Existing low-light image enhancement methods generally suffer from high model complexity, considerable deployment cost, and semantic coupling between luminance and chrominance information, which often leads to fusion artifacts and restricts their application on resource-constrained devices. To address these issues, this paper proposes a lightweight low-light image enhancement network that combines multi-channel parallel attention and cross-fusion mechanisms. In the YCbCr color space, luminance and chrominance information are modeled separately. A lightweight denoising module and a multi-head self-attention mechanism are introduced to extract illumination structure features, while a multi-channel parallel attention module is introduced to enhance color-context modeling. In addition, a luminance-guided cross-fusion module is introduced to achieve collaborative enhancement of structure and color under joint channel-spatial attention optimization.

The input image is first converted into the YCbCr color space and decomposed into the Y, Cb, and Cr channels. The Y channel is processed by the DenoiseY module to suppress noise, and then passed through a multi-head self-attention module (MSA) to extract structural features. To improve contextual modeling efficiency, a pooling operation is inserted before MSA to reduce the number of tokens. Meanwhile, the Cb and Cr channels are enhanced by the MCPA module, which performs multi-scale attention modeling to extract richer color-context information.

Subsequently, in the LCA module, the denoised Y channel is used as a guidance feature to fuse with the Cb and Cr channels. This module incorporates channel attention based on the squeeze-and-excitation (SE) mechanism together with element-wise interaction, enabling luminance-guided structure-color fusion.

Finally, the fused features are restored to the RGB space through convolution layers to generate the enhanced image. By separately modeling luminance and chrominance information and introducing cross-guided fusion between the two branches, the proposed method effectively improves the overall quality of low-light images in terms of detail preservation, illumination enhancement, and color restoration.

The proposed network achieves 25.66 dB PSNR / 0.84 SSIM on LOLv1, 24.61 dB PSNR / 0.85 SSIM on LOLv2, and 20.26 dB PSNR / 0.57 SSIM on LSRW, outperforming recent methods such as Retinexformer. Meanwhile, the model contains only 0.059 million parameters and requires 10.06 GFLOPs, demonstrating its lightweight nature and computational efficiency. In addition, low-light face detection experiments conducted on the DARK FACE dataset show that all three detection metrics exceed 50%, indicating good generalization capability under low-light conditions.

Experimental results further show that MCPA is embedded into both the luminance and chrominance branches to extract contextual information at different scales and perform feature reweighting, thereby improving the network’s adaptive modeling ability for multi-scale structural patterns and color variations. The LCA module enables joint modeling of structure and color, where structural information provided by the Y branch dynamically guides the adjustment of chrominance features, effectively improving detail fidelity and color consistency in the enhanced images.

The proposed method achieves high-quality low-light image enhancement under resource-constrained conditions. Experimental results on multiple public datasets verify its superior performance, especially in achieving a favorable balance among structural fidelity, color naturalness, and computational efficiency. This study demonstrates that separate modeling of luminance and chrominance information, together with a luminance-guided cross-fusion strategy, can effectively alleviate the fusion artifacts caused by luminance-color coupling in conventional methods.

Light field super-resolution (LFSR) aims to reconstruct high-resolution light field images from low-resolution observations while preserving both fine spatial details and angular consistency among multiple views. Because spatial and angular information are tightly coupled in light field data, reconstruction is challenged by both feature complexity and view dependency. Existing approaches often suffer from two major limitations. First, high-frequency textures, edge details, and subtle structural patterns are easily degraded during feature extraction and upsampling, resulting in blurred outputs. Second, insufficient modeling of spatial-angular correlations may introduce inconsistencies across viewpoints, impairing geometric fidelity and visual coherence. To address these issues, this paper proposes a light field super-resolution framework that integrates Mamba-based subspace scanning with diffusion-based generative reconstruction. The framework is designed to enhance high-frequency detail recovery, strengthen long-range spatial-angular dependency modeling, and improve reconstruction accuracy and cross-view consistency under different upscaling settings.

The proposed framework adopts a dual-branch subspace scanning strategy based on the Mamba architecture. Considering that light field images exhibit complementary characteristics in different subspaces, two specialized branches are constructed for efficient modeling. The first branch, termed EPI-Mamba, focuses on Epipolar Plane Image (EPI) structures, which explicitly characterize geometric relationships across viewpoints. This branch is therefore used to capture directional continuity and structural variation in epipolar dimensions. The second branch, termed Spatial-Angular Mamba (SA-Mamba), is designed to model correlations between spatial content and angular variation, enabling the network to learn dependencies that are difficult to represent using conventional convolution alone. Both branches perform bidirectional scanning in their respective subspaces, allowing efficient long-range dependency modeling while maintaining relatively low computational complexity.

The serialized features extracted from the two branches are then fed into a Multi-scale Cross Interaction (MCI) module. This module promotes deep information exchange between EPI-aware features and spatial-angular features at multiple scales, thereby enhancing complementary fusion of geometric and texture information. To further refine the fused representations, a Spatial-Angular Modulation (SAM) module is introduced. This module jointly calibrates features from the spatial and angular perspectives, adaptively emphasizing informative responses and suppressing inconsistent activations. As a result, cross-view feature alignment is improved and the coherence of reconstructed light field content is strengthened.

To mitigate the loss of high-frequency details, a frequency-domain enhancement mechanism is incorporated into the framework. Specifically, Fast Fourier Transform (FFT) is applied to the fused features to obtain frequency-domain representations, from which informative high-frequency components are selectively enhanced. This process compensates for detail attenuation, reinforces discriminative structural responses, and suppresses irrelevant or noisy signals. The enhanced features are then input into a diffusion-based denoising network. Benefiting from the strong generative capability of diffusion models, the network progressively restores fine details through iterative denoising and reconstructs high-resolution light field images after upsampling. The cooperation between subspace-aware feature extraction and diffusion-based refinement enables the framework to balance reconstruction fidelity, structural accuracy, and perceptual quality.

Extensive experiments are conducted on multiple benchmark datasets to evaluate the proposed method. Quantitative comparisons show that the proposed framework consistently outperforms representative state-of-the-art methods under different magnification settings. In the 2× light field super-resolution task, the proposed method achieves a peak signal-to-noise ratio (PSNR) of 39.43 dB and a structural similarity index (SSIM) of 0.987, representing the best performance among the compared methods. In the more challenging 4× task, the proposed approach still attains the highest results, reaching 33.70 dB in PSNR and 0.945 in SSIM. In particular, the PSNR is improved by up to 1.44 dB over existing methods, demonstrating a clear quantitative advantage.

Qualitative results further confirm the superiority of the proposed framework. Compared with competing approaches, the reconstructed images exhibit sharper boundaries, clearer local textures, and more faithful structural recovery, especially in regions containing dense lines, repetitive patterns, or complex high-frequency details. In addition, the restored views show stronger cross-view consistency, with fewer artifacts such as blurring or misalignment. These observations indicate that the method not only improves distortion-based metrics but also enhances perceptual quality and geometric coherence, both of which are essential for light field imaging applications.

The performance gains can be explained from several aspects. First, the dual-branch Mamba subspace scanning strategy makes full use of the intrinsic properties of light field data by separately modeling EPI structures and spatial-angular dependencies. Second, the MCI and SAM modules strengthen feature interaction, adaptive fusion, and cross-view calibration, thereby improving both discriminative ability and reconstruction stability. Third, the frequency-domain enhancement mechanism directly compensates for high-frequency information loss, which is especially beneficial for recovering textures and edge details. Finally, the diffusion-based denoising network further refines the reconstructed results by exploiting generative priors, leading to more realistic and visually pleasing outputs. Together, these modules form a unified framework in which each component contributes to performance from a complementary perspective.

This paper presents a light field super-resolution framework that combines Mamba-based subspace scanning with diffusion-based generative reconstruction. By jointly capturing EPI structures and spatial-angular correlations, and by integrating multi-scale cross interaction, spatial-angular modulation, and frequency-domain enhancement, the proposed method effectively addresses two core challenges in light field super-resolution: high-frequency detail loss and cross-view inconsistency. Experimental results demonstrate that the method achieves state-of-the-art performance in both objective metrics and subjective visual quality, with notable advantages in texture recovery, edge preservation, and structural consistency. These findings indicate that the proposed framework provides an effective solution for high-quality light field super-resolution. Future work may explore more lightweight architectures and more efficient diffusion strategies to reduce computational cost and extend the framework to related applications such as depth estimation, view synthesis, and light field restoration.

Cardiovascular and cerebrovascular diseases remain leading causes of global mortality and morbidity, imposing a substantial clinical and socioeconomic burden. These diseases require accurate early diagnosis and minimally invasive image-guided intervention. Optical coherence tomography (OCT) is a high-resolution optical imaging modality based on low-coherence interferometry. It provides real-time cross-sectional imaging with micrometer-scale resolution, enabling detailed visualization of vascular microstructures. OCT operates without ionizing radiation and demonstrates strong resistance to metallic artifacts, making it particularly suitable for intravascular applications where stents and calcifications are frequently present. Owing to these advantages, OCT has become an essential tool for lesion characterization and interventional guidance. It enables precise evaluation of atherosclerotic plaques, arterial dissection, aneurysm morphology, stent deployment quality, and post-procedural vascular healing, thereby providing critical microstructural information for precision diagnosis and treatment planning.

From a technical perspective, OCT is categorized into time-domain OCT (TD-OCT) and Fourier-domain OCT (FD-OCT) according to signal acquisition and reconstruction strategies. TD-OCT relies on mechanical reference-arm scanning to obtain depth-resolved signals, which limits imaging speed and sensitivity. In contrast, FD-OCT reconstructs depth information through Fourier transformation of spectral interference signals without mechanical delay scanning. This approach significantly improves signal-to-noise ratio, imaging speed, and sensitivity. As a result, FD-OCT has largely replaced TD-OCT in clinical practice and has enabled high-throughput intravascular imaging during percutaneous interventions. Fiber-based endoscopic OCT probes are the core components for intravascular imaging systems. These probes are generally classified into side-view and forward-view configurations based on imaging geometry. Side-view probes are widely used in coronary and cerebrovascular imaging, providing 360-degree circumferential visualization of vessel walls and enabling accurate assessment of lumen morphology and plaque distribution. Continuous advances in microfabrication and optical design have reduced probe diameters to below 0.5 mm, allowing safe navigation in small and tortuous vessels while maintaining imaging stability. Forward-view probes, on the other hand, are more suitable for anatomical navigation and lesion targeting, particularly in complex vascular geometries and preclinical cerebrovascular or gastrointestinal applications where directional imaging is required. In recent years, multimodal OCT probe systems have emerged as an important research direction. These systems integrate OCT with complementary imaging modalities such as ultrasound, photoacoustic imaging, fluorescence lifetime imaging, and near-infrared spectroscopy. Such integration enables simultaneous acquisition of structural, functional, and molecular information, overcoming the intrinsic limitation of OCT in penetration depth and biochemical specificity. Multimodal imaging significantly improves the comprehensive assessment of vulnerable plaques by combining morphological features with compositional and functional biomarkers.

Clinically, OCT has been widely adopted in cardiovascular and cerebrovascular diseases. It provides high-resolution visualization of plaque microstructures, including lipid-rich necrotic cores, fibrous caps, calcification patterns, microvessels, and intraluminal thrombi. These features are essential for distinguishing stable plaques from vulnerable plaques that are prone to rupture. In coronary artery intervention, OCT plays a critical role in stent optimization. It allows quantitative assessment of stent expansion, malapposition, under-expansion, edge dissection, tissue prolapse, and neointimal hyperplasia. These measurements directly inform procedural decisions such as balloon sizing, post-dilation strategy, and implantation optimization, ultimately improving procedural safety and long-term outcomes. In cerebrovascular applications, OCT has demonstrated increasing clinical value in aneurysm evaluation, intracranial atherosclerosis assessment, and flow-diverter treatment monitoring. It enables detailed visualization of aneurysm wall microstructure, evaluation of stent apposition in tortuous intracranial vessels, and assessment of endothelial healing after endovascular treatment. These capabilities provide critical information for rupture risk stratification and postoperative outcome prediction, which are difficult to achieve with conventional imaging modalities.

Recent OCT systems show a clear evolution toward platform integration, multimodal fusion, and artificial intelligence (AI)-assisted analysis. Commercial systems developed by major manufacturers such as Abbott, Terumo, and Zeiss support rapid pullback imaging, automated lumen and stent analysis, and real-time image enhancement. These improvements significantly increase procedural efficiency and reduce operator dependence. Meanwhile, AI-driven OCT analysis has developed rapidly. Deep learning models based on convolutional neural networks and transformer architectures have been applied to vessel segmentation, plaque classification, calcification quantification, stent detection, and lesion identification. These models enable automated pixel-level segmentation and frame-level classification, improving both diagnostic accuracy and inter-observer consistency. Furthermore, AI systems facilitate large-scale quantitative analysis, enabling new opportunities for imaging biomarkers and outcome prediction models.

Despite these advances, several challenges remain in OCT technology and clinical translation. Limited imaging penetration restricts visualization of deep vessel wall structures. Metal-induced shadowing still affects quantitative evaluation in heavily calcified lesions. In addition, variability in imaging protocols and lack of standardized datasets hinder large-scale clinical validation of AI models. Interpretability and generalization of AI algorithms across institutions also remain important issues for clinical adoption. OCT has become an indispensable high-resolution intravascular imaging technology in cardiovascular and cerebrovascular precision medicine. Future development will focus on ultra-miniaturized probes with diameters below 0.3 mm, faster imaging systems with higher frame rates, deeper penetration imaging through optimized light sources, and improved multimodal real-time fusion platforms. Advances in artifact suppression techniques will further enhance image quality in complex vascular environments. In parallel, interpretable and robust AI models with strong cross-center generalization capability will promote the transformation of OCT from an image interpretation tool into an integrated clinical decision-support system. With continuous technological innovation and clinical translation, OCT is expected to significantly improve early detection of vulnerable plaques, refine interventional strategies, and reduce adverse cardiovascular and cerebrovascular events. These developments will ultimately support the advancement of precision and personalized medicine in vascular diseases.

Structured light 3D reconstruction is widely used in high-precision metrology and industrial inspection because of its non-contact and high-resolution characteristics. In practical industrial environments, however, captured fringe patterns are often corrupted by composite noise originating from sensor electronics, surface reflectance variations, and ambient interference. Such noise degrades phase quality, introduces unwrapping errors, and limits the achievable reconstruction accuracy. To address these issues, this paper proposes a multi-modal deep denoising method that integrates a dual-channel input, a squeeze-and-excitation channel attention mechanism, and a hybrid loss function combining MSE, SSIM, and Laplacian edge constraints. The objective is to effectively suppress multi-source noise while preserving fringe continuity and edge structures, thereby enhancing phase retrieval stability, reducing unwrapping failures, and ultimately improving both image-domain quality and 3D reconstruction accuracy under complex industrial conditions.

The proposed method employs a dual-channel residual convolutional neural network based on the DnCNN architecture. The input is constructed by concatenating the original noisy fringe pattern with a serially filtered image along the channel dimension. The serial filtering module first applies extreme value detection and selective median filtering to remove impulse noise, followed by two-dimensional Gaussian smoothing to suppress high-frequency random noise. Multi-scale convolutional layers with batch normalization and ReLU activation are then introduced to extract hierarchical fringe features from the dual-channel input, capturing both fine textures and coarse structures while preserving edge information.

Subsequently, squeeze-and-excitation (SE) channel attention modules are embedded after convolutional layers in each residual block to adaptively emphasize informative features. Each SE module performs global average pooling to compress spatial information, followed by a bottleneck mapping with ReLU and Sigmoid activation to generate channel-wise weights, which are then used to recalibrate the feature maps. A hybrid loss function is designed to balance global noise suppression and local structure preservation, combining mean squared error for luminance fidelity, structural similarity index for contrast and texture consistency, and a Laplacian-based edge loss to maintain high-frequency details and boundary sharpness.

Experiments are conducted on a binocular structured light system using three-frequency four-step phase shifting and heterodyne phase unwrapping. The left camera coordinate system is defined as the world coordinate system, and 3D coordinates are obtained via linear triangulation with calibrated camera and projector parameters. Training and validation data are collected from three representative objects: standard step blocks (rule geometries with sharp edges), plaster busts (smooth curved surfaces), and printed circuit boards (high-reflectivity components and fine textures). Reference images are obtained through multi-frame averaging.

Quantitative evaluations in the image domain show that the proposed method achieves a PSNR of 43.11 dB / SSIM of 0.9654 on the standard step block, 42.90 dB / 0.9627 on the plaster bust, and 42.69 dB / 0.9643 on the PCB. Compared with noisy fringe patterns, the average PSNR increases by approximately 7 dB, SSIM improves to about 0.96, and RMSE decreases by around 61.6%.Ablation studies confirm the contribution of each component: dual-channel input improves PSNR by approximately 1.05 dB over single-channel input; the SE attention module provides consistent gains in both PSNR and SSIM, especially in edge detail recovery; and the hybrid loss function achieves the best overall performance. Comparative experiments against traditional methods (NLM, BM3D) and learning-based methods (BM3D-Net, DIVA) demonstrate that the proposed method achieves the highest PSNR (42.90 dB), highest SSIM (0.9627), and lowest RMSE (0.0063) on the plaster bust dataset. Noise robustness analysis under varying Gaussian noise levels (σ_g = 5 to 20) shows that even at the strongest noise level (σ_g = 20), the proposed method maintains a PSNR of 27.6 dB and an SSIM of 0.8032, substantially outperforming the original DnCNN (25.6 dB, 0.7514) and noisy inputs (15.7 dB, 0.4057), thereby demonstrating a high noise tolerance.

In the phase domain, denoised fringe patterns produce wrapped and absolute phase maps with significantly fewer phase jumps and fringe discontinuities, particularly in edge regions and high-reflectivity areas. Phase gradient analysis shows that the denoised histogram is more concentrated near zero, with a marked reduction in high-gradient outliers. In the 3D reconstruction domain, the point-to-plane RMSE decreases by 47.1%, the 95% quantile error decreases by 42.4%, the maximum error decreases by 16.7%, and point cloud density increases by 43.6%. Connectivity analysis reveals that high-error regions transition from large clustered patches to sparse isolated spots, indicating effective error tail suppression and spatial homogenization without introducing new artifacts. Region-wise error analysis further shows that RMSE and 95% quantile error decrease in both flat and edge regions, while edge metrics such as boundary accuracy, boundary completeness, and normal angle errors all improve. Consequently, the proposed method preserves key structural and edge details while suppressing noise, confirming that no edge rounding or structural weakening occurs.

This paper proposes a deep denoising method for structured light fringe patterns that integrates multi-modal input and structure-preserving mechanisms. Based on an improved DnCNN framework, the method employs a dual-channel input combining the original noisy fringe pattern with a serially filtered image, introduces a squeeze-and-excitation channel attention mechanism, and adopts a hybrid loss function. Experiments on standard step blocks, plaster busts, and printed circuit boards demonstrate significant improvements in PSNR, SSIM, and RMSE, while maintaining texture preservation and fringe continuity. In the 3D reconstruction domain, point cloud density, accuracy, and structural fidelity are substantially enhanced. The proposed method provides a reliable preprocessing strategy for high-precision industrial measurement and defect detection.

As various emerging new pollutants continue to emerge and multiply in the environment, their potential ecological toxicity, bioaccumulation, and environmental persistence have become increasingly prominent, posing severe, multidimensional, and complex threats to natural ecosystems and human health. Notably, emerging new pollutants typically lack historical reference data, making it difficult to generate the labels required for supervised learning. Given that conventional detection approaches rely heavily on large-scale labeled datasets for model training, such a dearth of prior data severely constrains their applicability. Consequently, traditional methods often exhibit unsatisfactory performance in the identification of unknown new pollutants.

In response, an innovative method for constructing a dynamic feature code library by coupling three-dimensional fluorescence spectroscopy with long short-term memory (LSTM) networks and incremental learning was proposed, with the specific purpose of enabling accurate identification of uncharacterized new pollutants. This method first reconstructs three-dimensional fluorescence spectroscopy data into quasi-time series ordered by excitation wavelengths, thereby obtaining a dynamic sequence with intrinsic order and contextual correlations. A wealth of chemical information, including the species and concentration of target analytes, is concealed within this sequence. It then leverages LSTM networks to sequentially read the emission spectra corresponding to each excitation wavelength, enabling the capture of dependencies between different wavelengths. Finally, after iterating through all emission spectra, the final output of the last time step is taken as the feature representation. This output serves as the feature code for the fluorescence spectrum of the pollutant sample and can be used for identification. Through this process, discriminative feature codes that can highly represent the overall spectral information are extracted; these codes encapsulate the core characteristics of the spectra, laying a solid foundation for subsequent pollutant identification. Furthermore, an incremental learning mechanism is introduced. This mechanism constructs a dynamically expandable feature code library based on the previously derived feature codes, achieves automatic identification of unknown pollutants by leveraging the calculated similarity threshold, and further integrates the feature codes of these newly identified pollutants into the library to support subsequent detection tasks. This design ensures that while the model continuously learns the features of new pollutants, it still maintains robust recognition performance for existing contaminants, resolving the issue that traditional recognition methods are unable to identify pollutants not included in their training data.

To verify whether this method possesses the capability of continuously learning knowledge from untrained pollutant samples, three rounds of tests were designed. To simulate the variability present in real-world environments, test samples under different pH conditions were incorporated. The test results demonstrated that this method achieved an identification accuracy of 93.3% for known pollutant categories and 91.7% for unknown new pollutant categories. These results indicate that the method exhibits excellent performance in both mitigating catastrophic forgetting and adapting to knowledge pertaining to new categories, fully validating its capacity for continuous learning and incremental identification. Furthermore, to evaluate the effectiveness of the proposed method, simulated pollution experiments were conducted by adding five typical contaminants to river water samples. The results demonstrate the proposed method achieved an identification accuracy of 93.3% for single pollutants, outperforming principal component analysis, parallel factor analysis, residual neural networks, and the incremental learning benchmark iCaRL. This is attributed to the fact that the proposed method can effectively capture the complex dependencies inherent in pollutant sample data along the excitation wavelength dimension, thereby extracting feature representations with stronger discriminative power and significantly enhancing the robustness against background fluorescence interference in river water matrices. In addition, the proposed method achieves an accuracy of 70.8% in the complete identification of all components in mixed pollutants. Although this performance is slightly lower than that in single-pollutant identification tasks, it still surpasses other comparative methods. The slight performance limitation stems from the fact that the fluorescence peaks of two specific pollutants are relatively close in position. When mixed, these two components tend to cause feature-level confusion, leading the model to occasionally misclassify them as a single substance, which somewhat constrains the overall identification performance for mixed samples. Despite such challenges, the relative superiority of the proposed method across key metrics demonstrates its strong potential and promising application prospects in the field of mixed pollutant identification.

In summary, by capturing the dependencies across different wavelengths of three-dimensional fluorescence spectroscopy data to extract feature codes and integrating an incremental learning mechanism, the proposed method provides a robust solution for the identification of unknown new pollutants. Not only can this method effectively identify known pollutants, but it also possesses the capability to learn from unknown new pollutants and transfer the acquired feature information to subsequent detection tasks, thus demonstrating favorable knowledge transfer and continuous learning characteristics. This demonstrates its excellent scalability and adaptive capability. With the continuous incorporation of additional contaminant feature codes, the method can gradually expand its recognition scope, thereby offering a reliable technical solution for the long-term monitoring and identification of emerging contaminants.

Power transformer safe operation depends critically on oil-paper insulation condition. Traditional insulation aging detection approaches possess significant drawbacks, including long testing cycles, destructive procedures, and insufficient precision. Raman spectroscopy offers a rapid, non-destructive alternative by capturing molecular vibration characteristics associated with aging byproducts. However, conventional machine learning algorithms exhibit low efficiency, high computational cost, and weak generalization capabilities when processing high-dimensional Raman spectral data. Advanced two-dimensional convolutional neural networks demand excessive computational resources through artificial dimensionality expansion. To address these limitations, an intelligent insulation aging assessment approach utilizing a one-dimensional convolutional neural network (1D-CNN) integrated with a Dilated Inception-ResNet module is developed. The goal is to achieve accurate, rapid, and robust aging stage classification by automatically extracting multi-scale spectral features while mitigating gradient vanishing issues commonly found in deep networks, balancing computational cost and diagnostic performance effectively.

Accelerated thermal aging experiments generated three hundred oil-paper insulation samples. Mineral oil and kraft paper mixtures underwent continuous heating at 120 degrees Celsius for up to 480 hours. Samples were collected every 24 hours. Gas chromatography measured furfural content to establish ground truth labels, categorizing the samples into four distinct aging stages: initial, mid-term, late, and final. A portable Raman spectrometer collected molecular vibration spectra from the prepared samples. The laser power was set at 300 milliwatts with an excitation wavelength of 784.711 nanometers and an integration time of 500 milliseconds. Raw Raman spectral data underwent a rigorous serial preprocessing pipeline. Savitzky-Golay smoothing eliminated high-frequency noise interference. Subsequently, the adaptive iteratively reweighted Penalized Least Squares algorithm corrected baseline drift caused by fluorescence background interference. This preprocessing generated high-quality spectral data inputs. An enhanced 1D-CNN architecture was constructed. The core innovation involved designing a Dilated Inception-ResNet module. The network initially utilized a standard one-dimensional convolutional layer and max-pooling layer for preliminary feature mapping and dimensionality reduction. Two cascaded Dilated Inception-ResNet modules followed. Each module incorporated four parallel feature processing branches. The first branch utilized point convolutions for channel dimension linear transformations. The second and third branches applied a compress-expand strategy, using initial 1×1 convolutions followed by 1×3 and 1×5 one-dimensional dilated convolutions, respectively, to capture medium and long-range temporal dependencies without increasing parameter count. The fourth branch utilized max-pooling for significant feature retention. To prevent network degradation and gradient vanishing, adaptive residual connections linked the inputs and outputs of these modules, utilizing a 1×1 convolution for dimension matching when necessary. The network concluded with global flattening and fully connected layers for classification. Training utilized the AdamW optimizer and cross-entropy loss function.

Repeated random sampling validation experiments evaluated baseline model performance. The standard 1D-CNN achieved an average accuracy of 89.83% and a recall of 89.26%, outperforming traditional support vector machine and K-nearest neighbor classifiers. This demonstrated the inherent advantage of deep learning in automatically extracting representations from complex, high-dimensional spectral data without relying on manual feature engineering. Ablation studies verified the efficacy of the proposed Dilated Inception-ResNet architecture. The enhanced model achieved a maximum test set classification accuracy of 96.67%. This represented a significant absolute accuracy improvement of 6.67% over the original 1D-CNN and 3.34% over a standard Inception-1DCNN model without dilated convolutions or residual connections. The loss function curve demonstrated rapid and stable convergence within twenty epochs, confirming that the adaptive residual connections successfully facilitated smooth gradient backpropagation and eliminated gradient vanishing problems. Computational complexity analysis revealed that while parameters and floating-point operations increased moderately, the single-sample inference time remained exceptionally low at 0.1424 milliseconds, fully satisfying real-time monitoring requirements. Further extensive testing assessed model robustness and generalization capability under suboptimal data conditions. Three dataset configurations with varying total sample sizes and class distributions evaluated performance across different train-test split ratios. For a highly imbalanced dataset containing 230 samples, the proposed model maintained an average accuracy exceeding 92.5% across all split ratios. For a constrained small dataset containing only 170 samples, the average accuracy remained robust above 91.7%. These consistent performance metrics across varied data scenarios proved the multi-scale feature extraction mechanism successfully learned intrinsic physicochemical aging features rather than relying on statistical class distributions. The network architecture effectively prevented over-attention to majority classes and ensured reliable recognition of minority class samples representing critical severe aging stages.

The proposed Dilated Inception-ResNet 1D-CNN model provides a superior, non-destructive, and rapid diagnostic solution for oil-paper insulation aging assessment. Serial preprocessing techniques combining Savitzky-Golay smoothing and adaptive iteratively reweighted Penalized Least Squares algorithms significantly enhance Raman spectral data quality. The integration of multi-branch parallel dilated convolutions expands receptive fields for multi-scale feature extraction without escalating computational costs, while adaptive residual connections ensure stable deep network training. The model demonstrates exceptional classification accuracy, stability, and robustness, even when processing small or heavily imbalanced datasets. This intelligent diagnostic framework offers reliable technical support for transformer condition monitoring, predictive maintenance scheduling, and power system reliability assurance.

Infrared target detection plays an indispensable role in numerous critical domains, such as security surveillance, autonomous driving, and military reconnaissance, owing to its unique perceptual capability under complex environments (e.g., low-light conditions and severe weather). However, infrared images inherently suffer from low contrast, blurred details, and significant noise interference, which often lead to ambiguous target edges, missing texture features, and other challenges during the detection process. Existing deep learning-based infrared target detection algorithms (ITDA) exhibit inadequate performance in feature extraction and processing for infrared images, resulting in relatively high rates of missed detection and false detection. Moreover, our systematic analysis of infrared target detection tasks reveals that algorithms tailored for small infrared targets rely heavily on high-sensitivity feature extraction to capture subtle characteristics. Nevertheless, as the scale of detected targets increases, these algorithms tend to encounter overfitting to local textures and elevated false detection rates, thereby degrading overall performance. In practical applications, detection environments are dynamically changing with targets of varying scales; thus, multi-scale detection capability is critical to ensuring algorithms maintain high reliability and adaptability in complex real-world scenarios. Unfortunately, most state-of-the-art algorithms are optimized for single-scale targets, making it challenging to simultaneously satisfy the requirements of high-precision localization for small targets and effective semantic understanding for large targets.

To address the above issues, this paper proposes an adaptive multi-scale infrared target detection network based on YOLO (AFITDYOLO). This network is designed to receive infrared target images of different scales and employs a multi-layer feature extraction module and a multi-layer feature fusion module to enhance its multi-scale infrared target detection capability. Firstly, a multi-scale feature fusion module (MFFM) is proposed. This module enhances the correlation between features of different layers in the feature pyramid network (FPN), coordinates deep semantic features with shallow spatial detail features more effectively, and thereby improves the representational ability of multi-scale feature fusion. Secondly, a multi-kernel feature extraction convolution (MFEConv) is constructed. By utilizing heterogeneous convolution groups, MFEConv expands the receptive field and strengthens the model's feature extraction capability. Additionally, a cross-attention fusion module (CAFM) is designed. Through the comparative interaction of feature maps output by different layers in the detection network, CAFM leverages the complementary information among these feature maps to suppress infrared noise in images and further enhance feature representation capability.

To validate the effectiveness of the proposed method in improving detection performance, extensive training and evaluation are conducted on the CTIR dataset, which comprises road pedestrians and vehicles with multi-scale infrared targets. To further verify the adaptability of the method, additional experiments are performed on the SIRST-UAVB dataset—a single-frame UAV bird dataset characterized by more complex backgrounds and smaller target scales. Experimental results on these two datasets demonstrate that AFITDYOLO achieves mean average precision at 50% intersection over union (mAP50) of 88.9% and 90.7%, respectively, representing significant improvements of 5.6% and 6.5% compared with YOLOv10n. In terms of lightweight optimization, the proposed method achieves higher inference speed (measured in frames per second, FPS) while utilizing fewer model parameters (params) and floating-point operations (FLOPs). When compared with current mainstream methods, AFITDYOLO exhibits the highest detection accuracy, the lowest parameter count and FLOPs, and the fastest inference speed, demonstrating distinct advantages. Additionally, to evaluate the generalization ability of the proposed method, cross-dataset experiments are carried out on the HIT-UAV dataset (a high-altitude UAV infrared thermal imaging dataset) and the IRSTD-1k dataset (a classic infrared small target dataset). Experimental results indicate that while the precision (P) value of AFITDYOLO is slightly inferior to that of DEIM-N, it outperforms all other mainstream methods in all remaining evaluation metrics. These findings confirm that the proposed method achieves improved detection accuracy on the infrared datasets used in the generalization experiments, validating its strong generalization capability and further demonstrating its feasibility for cross-scenario deployment. Overall, the proposed method simultaneously achieves enhanced detection accuracy and lightweight optimization of the detection model, fully meeting the requirements of real-time detection applications.

The AFITDYOLO network proposed in this paper, which is an adaptive multi-scale infrared target detection network based on YOLO, enhances the detection accuracy of infrared targets of different scales under various backgrounds with a relatively small number of parameters. The proposed MFFM enhances the model's representational ability in multi-scale feature fusion by improving the correlation between features of different layers in FPN. Additionally, the lightweight convolution module MFEConv is designed to achieve an efficient and larger receptive field with minimal parameters by leveraging the target distribution characteristics of infrared images. Furthermore, the CAFM is introduced to highlight important feature information, filter out irrelevant background information, and suppress noise through the comparative interaction of feature maps output by different layers, thereby further boosting the model's feature representation capability. Experimental results demonstrate that the proposed method outperforms current mainstream algorithms, exhibiting excellent detection accuracy, lightweight performance, and generalization ability, along with cross-scenario deployment capabilities.

To address the limited spectral dimensionality of single-wavelength active detection and the redundancy of multi-band data in hyperspectral LiDAR (HSL), this paper proposes a recognition method for cat-eye systems based on spectral segmentation and maximum information coefficient (MIC) feature selection.

The experiment is divided into two groups. The first group uses a Hikvision DS-2CD2325E-I day-night hemispheric network camera, and the second uses a Tianzhiyan K811 dashcam as the cat-eye system. Experimental data are collected in an indoor corridor. The target distances to the HSL system range from 2.98 to 15.02 meters, and measurements are repeated at irregular intervals at progressively increasing distances. HSL first captures spectral-spatial point cloud data of cat-eye systems and interference targets. Then, the echo intensity curves are divided into three spectral regions, and representative wavelengths are selected from each based on their characteristics. These selected wavelengths serve as inputs to the support vector machine (SVM) classifier. To enhance classification accuracy and robustness, a feature selection framework based on maximum information coefficient (MIC) is applied, which combines a bootstrap strategy for assessing selection frequency with a redundancy-penalized greedy search to identify key wavelengths. The selected wavelengths are further tested, and the results are compared with those obtained using principal component analysis (PCA), linear discriminant analysis (LDA), random forest (RF), and multilayer perceptron (MLP) methods.

The results show that the wavelength in the fluctuation zone achieved the highest peak classification accuracy in the single-wavelength selection experiment. However, due to the saturation truncation effect of the echo signal, the quantized intensity values of some bands in this region are abnormal, thereby causing the classification accuracy to drop to its lowest level and exhibit severe fluctuations. Despite the near-infrared region exhibiting the highest average accuracy across bands, the actual peak response wavelength of the cat-eye system falls outside this zone. This indicates that the spectral characteristics of this region cannot adequately explain the underlying physical mechanism of the cat-eye effect. Despite having the lowest average accuracy, the visible light region demonstrates the smallest fluctuation range. The results indicate that relying on a single wavelength is difficult to fully characterize the spectral characteristics of cat-eye systems. In the optimization of MIC-based wavelength screening results, the high correlation between wavelengths within the same partition may stem from the small energy difference between waveforms emitted at adjacent characteristic wavelengths. Consequently, the echoes recorded across these channels show high similarity. The constraint of screening based on only a single sample leads to higher echo intensities from the cat-eye system at similar wavelengths, which further increases the spectral similarity of bands within the region. Therefore, each region ultimately retains only one representative wavelength. At the same time, selecting different feature bands for classification, the data shows that cross regional wavelength selection can better reflect the spectral differences between targets, further verifying the rationality of classification based on spectral features. In addition, different feature bands are used for classification. The results indicate that this method can more effectively capture spectral differences between targets, further verifying the theoretical basis of partitioning based on spectral features. The wavelength combination selected from all bands exhibits the poorest classification performance. This is primarily attributed to the absence of partition constraints, which can lead to the selected bands being concentrated in spectrally similar regions. Consequently, these bands exhibit high correlation and provide limited discriminative power, thereby adversely affecting classification accuracy. Therefore, the partitioning strategy avoids spectral redundancy by extracting wavelengths from different regions. In the end, this method only requires 3 wavelengths to achieve effective recognition of cat-eye systems, and the average classification accuracy with SVM reaches 0.858. To verify the applicability of the proposed method in different environments, comparative experiments were conducted in three different scenarios. In dark and low-light scenes, the influence of external light is weak. The signal-to-noise ratio of HSL laser echo signals is higher, and the extraction of spectral features is more stable. Therefore, the method performs best, with an average accuracy of 0.84 and an average F1 value of 0.85. In experimental scenarios with external light interference, the method proposed in this paper can still achieve good target classification, with an average accuracy of 0.77 and an average F1 value of 0.80. Compared with normal lighting scenarios (average accuracy of 0.81 and average F1 value of 0.84), it has decreased by 4.94% and 4.76%, respectively.

The cat-eye system recognition method proposed in this study addresses the problem of poor classification performance in traditional MIC feature selection, which is easily affected by band correlation and sampling instability, by introducing stability screening and redundancy constraints. It achieves robust extraction and combination optimization of key wavelengths, thereby more effectively mining the differential information between hyperspectral bands and improving classification performance. The band selection method in this article outperforms the other four types of feature selection methods in terms of average accuracy and F1 value, and also demonstrates better stability at the lowest accuracy, providing new ideas and solutions for quickly and efficiently identifying high reflection targets.

Image classification serves as a fundamental component in computer vision and supports a wide range of applications including object recognition, scene understanding, and intelligent perception systems. Deep convolutional neural networks have made substantial progress through hierarchical feature learning and residual architectures. Despite these advances, conventional convolution operations rely on fixed sampling grids and rigid kernel geometry, which limits their ability to adapt to geometric transformations such as rotation, scale variation, and orientation changes. When object orientation varies, features extracted by fixed convolution kernels become spatially misaligned, leading to inconsistent representations within the same category and reduced classification robustness. Deformable convolution partially alleviates this limitation by learning adaptive sampling offsets, yet fully data-driven offsets lack explicit geometric constraints and often result in offset divergence, unstable sampling patterns, and weak sensitivity to directional structures in deeper layers. In addition, existing convolution-based networks tend to assign similar importance to foreground and background features, which amplifies the influence of noise and reduces discriminative capability. The objective of this work was to construct an image classification network that simultaneously enhances rotation robustness and improves feature selection stability under complex geometric variations.

A rotational deformable convolution-based image classification network, named RoDeNet, was developed on the basis of the ResNet-34 backbone. The core component of RoDeNet was the rotational deformable convolution module (RDCM), which integrated explicit geometric priors with adaptive offset learning. Instead of relying solely on unconstrained learned offsets, RDCM introduced predefined rotational offsets corresponding to four base orientations, namely 0°, 45°, 90°, and 135°. These offsets provided stable directional references for convolutional sampling. A compact offset prediction network was employed to generate dynamic offsets from feature responses. The predefined offsets and learned offsets were combined through element-wise addition, which constrained the deformation process within a geometrically meaningful range while preserving data-driven adaptability. RDCM adopted a dual-path structure. One path employed standard convolution to preserve stable base features and ensure consistent low-level representations. The other path applied rotational deformable convolution to capture orientation-adaptive features through direction-aware sampling. Feature outputs from both paths were concatenated along the channel dimension to form enriched representations. To prevent information loss during feature enhancement, a feature preserve module (FPM) was integrated after feature fusion. FPM consisted of an attention-guided enhancement branch and a feature retention branch. The attention branch amplified discriminative spatial responses, while the retention branch maintained the original structural information, ensuring balanced feature optimization. To further improve feature discrimination and noise suppression, a hierarchical gated attention (HGA) mechanism was incorporated into residual blocks. HGA divided feature channels into multiple groups and processed them through two cooperative branches. The spatial pooling branch extracted horizontal and vertical contextual information to model long-range spatial dependencies. The depthwise separable convolution branch focused on local spatial patterns with reduced computational cost. The outputs of the two branches were fused to generate adaptive attention weights, which were used to reweight feature responses across spatial and channel dimensions. In the overall architecture, RDCM replaced the initial convolution layer and the first residual block of each stage, while HGA was inserted after the second convolution within residual blocks to avoid interference with early geometric correction.

Extensive experiments were conducted on CIFAR-10, CIFAR-100, Imagenette, and Imagewoof datasets to evaluate classification accuracy, robustness, and efficiency. RoDeNet achieved classification accuracies of 96.87% on CIFAR-10, 81.81% on CIFAR-100, 94.72% on Imagenette, and 89.69% on Imagewoof. Compared with the baseline ResNet-34, accuracy improvements of 0.58%, 3.01%, 4.74%, and 5.09% were achieved, respectively. The performance gains were more pronounced on Imagenette and Imagewoof, which contain significant orientation variation and background complexity, indicating improved geometric robustness. Parameter sensitivity experiments demonstrated that moderate predefined rotation offset strength produced the most favorable performance balance. Excessively strong geometric priors constrained adaptive learning, while purely data-driven offsets failed to provide stable rotational modeling. Ablation studies confirmed that RDCM and HGA each contributed independently to performance improvement. RDCM enhanced direction-sensitive feature extraction, while HGA improved discriminative feature selection and suppressed background noise. Their combination produced greater performance gains than either module alone, indicating complementary functionality. Computational analysis showed that RoDeNet introduced a moderate increase in parameter count and computational complexity compared with ResNet-34, while remaining significantly more efficient than many Transformer-based classification models. Visualization analysis further revealed that RoDeNet focused more accurately on object regions and structural contours, whereas baseline models exhibited dispersed or background-dominated activation patterns. These observations confirmed the effectiveness of the proposed geometric adaptation and hierarchical attention mechanisms.

A rotational deformable convolution-based image classification network was presented to address the limitations of fixed convolutional structures under geometric transformations. By integrating predefined rotational priors with dynamically learned offsets, RDCM enabled stable and direction-aware adaptive sampling. The feature preserve module balanced feature enhancement and information retention. The hierarchical gated attention mechanism improved multi-level feature discrimination through grouped attention modeling. Experimental results demonstrate that the proposed network achieves superior classification accuracy and robustness under rotation and orientation variations while maintaining acceptable computational efficiency. The method provides an effective solution for rotation-sensitive image classification tasks. Future work will focus on reducing computational overhead and extending the framework to continuous rotation modeling and to scenarios involving non-rigid geometric deformations.

Immersive experience is a core determinant of user engagement and perceptual quality in metaverse-oriented virtual reality systems. Display characteristics play a central role in shaping immersion, as they directly regulate visual information delivery in near-eye environments and mediate the perceptual coupling between users and virtual scenes. In immersive perception research, multiple display dimensions have been incorporated into analytical frameworks, with field of view, resolution, and luminance commonly treated as fundamental parameters due to their direct influence on spatial coverage, visual clarity, and contrast perception. Color saturation, although closely associated with visual vividness, realism, and affective response, is frequently configured as a scene attribute rather than examined as an independent perceptual control variable. As a result, the quantitative relationship between saturation modulation and immersive experience remains insufficiently characterized, particularly under controlled display conditions. This gap limits the perceptually informed optimization of immersive display systems. The objective of this study was therefore to systematically investigate the effect of color saturation on immersive perception in virtual reality display systems and to develop an objective immersion quantification framework that integrates display parameters with neurophysiological responses.

A controlled experimental paradigm was designed based on a virtual reality near-eye display system to isolate the perceptual effects of color saturation while minimizing confounding visual factors. Six virtual scenes were constructed with identical spatial structure, interaction logic, object arrangement, environmental layout, and motion cues, thereby ensuring that any observed perceptual differences could be attributed primarily to saturation manipulation rather than scene content variation. Color saturation was manipulated at six predefined levels spanning low to high saturation conditions, while all other visual parameters, including luminance, contrast, resolution, and viewing geometry, were held constant throughout the experiment. Participants performed standardized interaction–perception tasks within each scene to maintain consistent cognitive demand, attentional allocation, and sensorimotor engagement across experimental conditions. During task execution, electroencephalogram (EEG) signals were continuously recorded using wearable devices, enabling real-time capture of neural activity associated with immersive perception under naturalistic interaction states. Following each task, participants reported their perceived immersion using structured subjective questionnaires administered immediately after exposure to reduce memory bias. This experimental protocol enabled synchronized acquisition of behavioral context, subjective evaluation, and physiological signals under systematically varied saturation conditions. Statistical analyses were conducted to assess the effects of color saturation on subjective immersion scores and EEG-based objective indicators. On this basis, representative EEG features were extracted and combined with saturation parameters to construct an objective immersion quantification model capable of producing a continuous immersion index across conditions, thereby linking display-level manipulation with perceptual and neurophysiological responses.

Data analysis indicated that color saturation exerted an extremely significant influence on subjective immersion scores (p < 0.001) and a highly significant influence on objective immersion evaluation results based on EEG measurements (p < 0.01). Further analysis revealed a clear enhancement effect within the medium-to-high saturation range. Immersion increased significantly as saturation rose from approximately 75% to full saturation, whereas lower saturation conditions were associated with reduced immersion stability and weaker neural engagement. These results suggest that color saturation functions as an effective immersion-enhancing factor when maintained within an appropriate perceptual range, likely by strengthening visual salience, scene vividness, and perceptual presence without inducing sensory overload or discomfort. Importantly, the observed trends were consistent across participants and scene configurations, indicating that the saturation–immersion relationship was not driven by individual preference alone but reflected a more general perceptual mechanism.To bridge subjective experience and physiological measurement, an objective immersion quantification model was constructed by integrating color saturation parameters with EEG-derived features. The predicted immersion scores generated by the model showed high consistency with subjective questionnaire ratings, yielding a correlation coefficient of r = 0.885. This strong correspondence indicates that the proposed model effectively captured essential variations in immersive experience across different saturation conditions. In addition to predictive accuracy, the model offers an interpretable mapping between display-level color modulation and neural response patterns, supporting its use as an analysis tool rather than a black-box predictor. The results further demonstrate the feasibility of employing lightweight EEG signals as a complementary tool for immersion assessment, reducing exclusive reliance on post-experience questionnaires and enabling more continuous and objective evaluation of user experience.

Color saturation is confirmed as a significant and quantifiable display parameter influencing immersive experience in virtual reality systems. Increasing saturation within an appropriate range enhances immersion at both subjective and neurophysiological levels, with particularly strong effects observed in medium-to-high saturation conditions. The integration of saturation parameters with EEG signals enables objective and continuous immersion evaluation, offering a practical pathway toward real-time immersion monitoring in interactive virtual environments. The proposed framework provides methodological support for isolating display-parameter effects and offers actionable guidance for immersive display design and content color tuning in metaverse applications. More broadly, the modeling approach establishes a foundation for adaptive display systems that dynamically adjust visual parameters in response to user perceptual states, thereby contributing to the development of perceptually optimized, user-centered virtual reality experiences.

Multimodal three-dimensional object detection combines camera semantics with LiDAR geometry and plays a central role in autonomous driving perception. Yet performance often degrades in complex traffic due to unstable semantics, sparse geometry, and cross-modal misalignment. Camera features become unreliable under glare, low illumination, motion blur, and cluttered backgrounds, introducing uncertainty into bird’s-eye-view (BEV) projection. LiDAR point clouds are sparse at long range or in adverse weather, weakening local geometry and causing boundary blur and missed small objects. In BEV fusion, misalignment between modalities introduces conflicting cues and redundant responses, reducing confidence and accuracy. This study proposes an efficient BEV detector that enhances camera semantics, stabilizes sparse geometry, and achieves adaptive cross-modal fusion for improved accuracy and robustness without excessive computational cost.

The proposed framework, MEPFusion (multi-layer perception enhancement and gated alignment fusion), improves camera semantics, stabilizes sparse-geometry modeling, and enables gated alignment fusion in BEV space. For camera encoding, SimPVT enhances global context and salient-region representation with low overhead. A pyramid vision transformer backbone extracts multi-scale features to capture global and local dependencies. A parameter-free saliency reweighting unit amplifies target-related activations and suppresses background distraction, improving semantic separability. To reduce computation, the highest-resolution stage is removed while preserving three-scale outputs for projection. Features are lifted to BEV space via a lift–splat–shoot transformation supervised by depth estimation, stabilizing the depth distribution and reducing projection noise to improve spatial consistency. This camera branch aims to provide more reliable semantics under appearance variations, reducing texture-induced false positives and improving recall for visually small instances.

For LiDAR encoding, raw point clouds are voxelized and processed through a SECOND-style sparse convolution backbone. Standard sparse convolutions are replaced with spatial density-modulated convolution (SDMConv) to address equal-weight aggregation in sparse regions. SDMConv injects a position-dependent density prior into convolution responses via multiplicative modulation, producing a center-enhanced and periphery-decayed pattern within the receptive field. The modulation mitigates feature dilution and preserves boundary cues and shape continuity, especially in long-range or partially occluded areas, without increasing parameter count. The resulting sparse 3D features are compressed along height to produce LiDAR BEV features. In practice, SDMConv acts as a plug-in operator that strengthens local geometry when point evidence is weak, benefiting thin structures and small obstacles that are easily under-sampled.

Cross-modal fusion occurs in BEV space through a gated alignment fusion module (GAFM) designed to suppress misalignment while maintaining complementarity. Each modality undergoes channel recalibration to emphasize discriminative channels. A residual mutual-guidance path performs bidirectional correction: LiDAR features refine camera BEV features in geometry-sensitive regions, while camera semantics complement LiDAR features where geometric evidence is weak. Spatial gates are predicted for each BEV cell and normalized via softmax to assign adaptive fusion weights, allowing the more reliable modality to dominate locally while reducing inconsistent responses. The fused BEV features are decoded by a multi-scale neck and a transformer-based detection head to produce 3D bounding boxes and attributes. This strategy improves optimization stability and encourages reliability-aware weighting rather than uniform mixing.

Extensive experiments on the nuScenes benchmark demonstrate the effectiveness of MEPFusion. With six cameras and one LiDAR input, MEPFusion achieves 67.1% mAP and 70.8% NDS, surpassing the BEVFusion baseline by 2.8% and 1.7%, respectively. Regression errors in translation, scale, orientation, velocity, and attribute are reduced, confirming improved spatial alignment and motion estimation. Category-level analysis shows notable gains for small or geometry-sensitive objects (e.g., traffic cones, barriers, motorcycles, and pedestrians) while maintaining improvements for major vehicle classes. These trends align with module contributions: SimPVT strengthens semantics in cluttered scenes, SDMConv alleviates sparsity-induced degradation, and GAFM reduces cross-modal interference through adaptive gating and complementary fusion. Qualitative inspection suggests fewer duplicated boxes and cleaner spatial responses in crowded regions, consistent with suppressing redundant activations caused by misalignment.

Distance-stratified evaluation further verifies consistent improvements across near, mid, and far ranges. The far range, where point sparsity and scale ambiguity are most severe, still achieves clear gains, highlighting the benefits of density-aware encoding and local gating. Robustness tests under adverse conditions also validate generalization. On the night subset, MEPFusion improves mAP and NDS while reducing error terms, indicating stable semantics and selective fusion under low illumination and glare. On the rain subset, performance again improves, confirming resilience to reduced contrast and weaker LiDAR returns. Ablation studies reveal both independent and synergistic effects: SimPVT alone enhances accuracy with fewer parameters; SDMConv and GAFM each bring additional gains; combining all three yields the best overall performance. Efficiency analysis shows SDMConv introduces minimal overhead, and the full model maintains comparable throughput with reduced total parameters, indicating a favorable accuracy–efficiency trade-off for on-vehicle inference.

In summary, MEPFusion integrates semantic enhancement, sparse-geometry reinforcement, and gated alignment fusion within a unified BEV framework. The method achieves higher accuracy, improved regression stability, and consistent robustness across distance and environmental variations while retaining deployment-friendly efficiency. Stable semantics before view transformation, density-modulated LiDAR encoding, and adaptive BEV fusion jointly mitigate misalignment and suppress redundant noise. MEPFusion thus offers an effective and reliable multimodal perception solution for urban autonomous driving.