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.

White-light interferometry is widely employed in precision manufacturing sectors, including semiconductor processing, thin-film thickness measurement, and metal defect detection, owing to its merits of high precision, high measurement speed, and non-contact nature. The frequency domain analysis (FDA) technique has been applied to surface profile measurement by utilizing the slope of the phase distribution of the white-light scanning interference signal in the wavenumber domain. Through simulations and experiments, this study demonstrates that the accuracy of the phase obtained in the wavenumber domain is related to the amplitude distribution, whether in the presence of additive noise or phase noise, and quantitatively provides the optimal phase selection range for the FDA method. When the sample has strong spectral absorption, parts of the amplitude distribution in the wavenumber domain decrease significantly, rendering the corresponding phase information invalid and unreliable. To address this limitation, an improved algorithm is proposed in this paper, which incorporates the wavenumber-domain amplitude as a weighting factor into the least-squares fitting procedure. This method avoids the selection of the phase range for the FDA method and ensures high measurement accuracy.

In the simulation, the zero-mean Gaussian noise with a standard deviation (STD) of 10 nm is used as the phase noise ε_n(z) to generate a white-light interference signal. The Fourier transform is performed on the signal to obtain the amplitude I_F($\sigma $) and phase ($\sigma $) in the wavenumber domain. For the FDA method, the selected phase range $R_\sigma $ is defined as the wavenumber range corresponding to amplitudes in I_F($\sigma $) that are greater than CImax, where the coefficient C∈(0, 1) and I_max is the maximum value of I_F($\sigma $). The STD of z_s represents the measurement accuracy of the FDA approach. The influence of phase noise on the measurement accuracy and its relationship with amplitude distribution I_F($\sigma $) are investigated by calculating the STD of z_s across different $R_\sigma $. Furthermore, additive noise is added to the white-light interference signal with phase noise, and the STD of z_s is evaluated under varying $R_\sigma $ to determine the optimal phase selection range. Since conventional phase selection is not suitable for interference signals with complicated spectral distributions, this paper proposes a special frequency domain algorithm based on amplitude optimization in the wavenumber domain (amplitude optimization for frequency domain analysis, AOFDA). In this improved strategy, the amplitude distribution I_F($\sigma $) is introduced as a weighting factor in the phase linear fitting process to obtain the optimized surface measurement parameter z_w.

The simulation results indicate that when only phase noise exists, the accuracy of phase extraction improves with the increase in the amplitude corresponding to the selected phase, thereby enhancing the overall measurement precision. However, the phase extraction accuracy has an upper limit. When the range $R_\sigma $ corresponding to C≥0.5 is selected in the FDA method, the measurement results exhibit higher accuracy. Additive noise with different signal-to-noise ratio (SNR) is introduced into the white-light interference signal containing phase noise. For interference signals with the same SNR, the STD of z_s obtained with C≥0.5 is stable and relatively low. For signals with stronger additive noise, the STD of z_s shows a slight increase when C>0.5. In summary, if only phase noise exists, selecting the range $R_\sigma $ corresponding to C≥0.5 yields high measurement precision. When both phase noise and additive noise exist, C=0.5 provides higher measurement precision. Therefore, the $R_\sigma $ corresponding to C=0.5 is the optimal phase selection range for the FDA method.

To verify the feasibility of the AOFDA method for interference signals with complex spectral distributions, a band-stop filter spectral distribution was used as the absorption spectrum of the measured object in the simulation. In the band-stop range, simulations with different reflectivities were conducted to investigate the measurement accuracy. When the reflectivity of the sample is ≥20%, it exerts little influence on the measurement accuracy of the FDA method. However, the reflectivity of ＜20% moves the measurement results away from their true values. When the reflectivity of the sample is ＜20%, the measurement results obtained by the AOFDA method always exhibit better performance. This indicates that the measurement accuracy of the AOFDA method is largely unaffected by the absorption spectrum of the sample.

The flat surface of wedged window and the step surface were measured in the experiment. When the different value of C is selected to obtain the surface profile of the wedged window, the C=0.5 has the better performance that agrees with simulation result. The S_z of the wedged window surface calculated by the AOFDA method and the FDA method (C=0.5) are 34.6 nm and 50.2 nm, and the S_q are 5.5 nm and 6.7 nm, respectively. The average step heights, whose value is regarded as 3.0 μm, of 10 measurements calculated by the AOFDA method and the FDA method (C=0.5) are 3.0054 μm and 2.9566 μm, and the repeatability values are 17.9 nm and 19.4 nm, respectively. The experimental results demonstrate that the AOFDA method can achieve the measurement accuracy comparable to that of the FDA method when selecting the optimal wavenumber-domain phase range. Since the selection of an appropriate wavenumber-domain phase range is no longer required for the AOFDA method, it enables the white-light interferometer to better adapt to complex measurement environments.

The measurement accuracy of the FDA method is degraded by environmental vibration and background noise. This work confirms that the measurement precision of the FDA method is correlated with the amplitude distribution in the wavenumber domain through the selection of diverse phase ranges, and proposes a feasible phase selection strategy to enhance its measurement performance. It is verified that the FDA method is difficult to adapt to the situation where the sample has a strong absorption spectrum. To address these limitations, the AOFDA method is developed in this paper, which incorporates the wavenumber-domain amplitude distribution as a weighting function into the phase fitting procedure. The AOFDA method can achieve a measurement accuracy comparable to that of the FDA method processed with the optimal phase range, while eliminating the need for determination of the optimal phase range.

Achieving both a large field of view and high resolution remains a persistent and critical bottleneck in the field of optical imaging. Conventional imaging systems are constrained by a fundamental trade-off: expanding the field of view typically leads to a noticeable reduction in spatial resolution or requires the adoption of optical components that are bulkier, heavier, and more complex in structure, which limits their applicability in scenarios requiring portability or miniaturization. Existing solutions aimed at addressing this dilemma, such as those relying on mechanical translation stages, oscillating mirrors, or multi-lens arrays, suffer from inherent limitations. Mechanical translation stages are restricted by mechanical inertia, resulting in scanning frequencies as low as 10 Hz that severely hinder imaging efficiency; oscillating mirrors are prone to mirror surface deformation due to centrifugal forces at high speeds, with their maximum scanning frequency capped at around 100 Hz; while multi-lens array systems significantly increase overall system complexity, require highly efficient processing algorithms, and rely excessively on precise calibration procedures, with an equivalent scanning frequency of merely 50 Hz. To overcome these challenges, this thesis conducts in-depth research on large field of view stitching imaging technology based on galvanometer fast scanning. The primary objective is to fully leverage the high-speed scanning advantage of galvanometers (capable of kHz-level scanning frequencies, far exceeding traditional mechanical systems) to construct an efficient and reliable image acquisition system with an overlap compensation mechanism. Simultaneously, the traditional Harris algorithm, a classic corner detection method in computer vision, is improved to enhance its performance in feature extraction and matching for stitching tasks. Ultimately, this research develops a practical method for acquiring overlapping sub-images and seamlessly stitching them into a single high-resolution panoramic image, thereby resolving the trade-off between field of view and resolution.

The research methodology is based on two core aspects: the construction of a scanning imaging system with an overlap compensation mechanism and the introduction of key improvements to the traditional Harris corner detection algorithm. The scanning imaging system was constructed using a high-precision galvanometer and a GigE industrial camera, enabling the rapid acquisition of a grid of sub-images. On the algorithmic front, the thesis proposes two main modifications. First, a priori non-extremely large value suppression shape is adopted. This method determines the optimal suppression shape (square, circle, or ellipse) by calculating gradient directions from downsampled images, allowing the algorithm to adapt to different image textures. Second, a center region shielding adaptive non-extremely large value suppression method based on the overlap rate is proposed. This technique detects the horizontal and vertical overlap ratios between images and strategically applies suppression only within the overlapping regions, effectively reducing interference from non-overlapping areas. These improvements enhance the accuracy of corner point screening and the computational efficiency of the algorithm. The overall stitching process follows a designed "group-of-four" iterative strategy and employs homography estimation and weighted linear blending for seamless image fusion.

The effectiveness of the proposed method was rigorously verified through simulation and experimental studies. Simulation verification involved partitioning ten high-resolution 9000 pixel×12000 pixel source images into sub-arrays with preset overlap rates of 30%, 40%, 50%, and 60%. The results demonstrated that, compared with the traditional adaptive non-extremely large value suppression algorithm, the proposed method achieved notably faster computational speed. It also generally yielded a higher number of valid corner matches and a higher correct matching rate, with the advantage being particularly pronounced at lower overlap rates. Computational efficiency was improved by more than 60% under the different overlapping rates tested. The stitched image exhibited high quality, with PSNR values exceeding 40 dB and SSIM values close to 1, indicating excellent fidelity and minimal distortion. For experimental validation, a 6×11 array of 2592 pixel×1944 pixel images was first captured using the constructed galvanometer scanning system. Applying the improved algorithm to this dataset successfully generated a seamless panoramic image with a final resolution of 9553 pixel×19149 pixel. Compared to the traditional method, the complete stitching process achieved a 39.69% improvement in computational efficiency and a 10.31% increase in the feature matching correct rate, which directly contributed to significantly enhanced visual consistency in the final large-field-of-view composite. These results substantiate that the integration of high-speed galvanometer scanning with the enhanced Harris algorithm effectively addresses key limitations such as the low scanning speed of conventional mechanical systems and the computational inefficiency of traditional global feature detection, establishing a practical and effective approach for high-resolution imaging over a large field of view.

This research successfully develops a large field of view stitching imaging technology based on galvanometer scanning. By constructing a rapid scanning system and improving the Harris algorithm through a priori suppression shape and overlap-rate-based center shielding adaptive non-extremely large value suppression, the study provides an effective solution to the classic trade-off between field of view and resolution. The technology demonstrates substantial improvements in processing speed and stitching accuracy, enabling the generation of ultra-high-resolution panoramic images. The results offer a new technical paradigm for applications ranging from astronomical observation and satellite remote sensing to pathology slice scanning and endoscopic imaging. Future work may involve integrating deep learning techniques to further optimize the algorithm's adaptability and efficiency in complex scenarios.

Industrial steel surface defect detection is a core step of industrial quality inspection. However, traditional detection algorithms suffer from low accuracy, high model complexity, and difficult deployment due to the various types and large scale variations of steel surface defects. Existing deep learning-based detection algorithms have improved detection accuracy to a certain extent, but most of them feature large model size and high computational complexity, which makes them hard to deploy on resource-constrained industrial edge devices and thus restricts their practical application in real-time industrial detection scenarios. This study aims to propose a lightweight steel surface defect detection algorithm that balances high detection accuracy and low computational complexity, realizes efficient feature extraction for multi-scale steel surface defects, and meets the requirements of real-time detection and edge device deployment in industrial scenarios.

A lightweight steel surface defect detection algorithm named RGCE-YOLO was proposed based on an improved YOLOv11n. The backbone network was reconstructed with a novel HGRCBlock structure combining lightweight HGNet and improved reparameterized RepLightConv, and multi-branch convolution and normalization steps were fused during inference to reduce computational load while maintaining feature extraction capability. For the neck network, the lightweight GSConv replaced the original downsampling structure to reduce redundant information via mixed convolution and address the accuracy bottleneck of depthwise separable convolution; the C3K2 structure was redesigned by integrating the Spatial and channel synergistic attention (SCSA) mechanism to alleviate sub-feature semantic differences and enhance multi-scale object feature fusion. The CMBC_Detect module, constructed based on EfficientNet's inverted bottleneck structure and combined with the coordinate attention (CA) mechanism, was introduced into the detection head’s regression branch; the SE attention was replaced with the CA mechanism to retain spatial detail information, enhance small object detection ability, and compress the structure simultaneously. Channel-wise knowledge distillation (CWD) was applied to the improved model: each channel’s feature map was normalized to a soft probability map, and the asymmetric KL divergence between the teacher and student models was minimized to make the student model focus on salient regions of defect features, reducing feature loss without increasing computational cost. A self-built steel surface defect dataset with four defect types (rust, scratch, pit, crack) was established and expanded via data augmentation strategies including random cropping, dynamic brightness adjustment, salt-and-pepper noise, and Gaussian noise, and then divided into training, validation, and test sets at an 8:1:1 ratio. Ablation experiments verified the effectiveness of each improved module, and comparative experiments on distillation methods and mainstream detection algorithms were conducted under the same conditions. Generalization experiments were carried out on the public NEU-DET and GC10-DET datasets, and the model was converted to ONNX format for deployment verification on an ARM industrial computer. Precision, mAP50, mAP50-95, parameters, FPS, and FLOPs were used as key evaluation indicators to comprehensively assess the algorithm's performance.

Ablation experiments confirmed the effective optimization role of each improved module: HGRCBlock reduced model parameters and FLOPs while enhancing multi-scale feature extraction; GSConv reduced the neck network’s redundant information and computational complexity; SCSA improved the model’s perception of multi-scale objects and detection robustness; CMBC_Detect strengthened the capture ability for small, easily missed defects and further compressed the model size; CWD improved model accuracy without changing model size and computational complexity. On the self-built dataset, RGCE-YOLO achieved a precision of 83.4% (3.9% higher than YOLOv11n), with mAP50 and mAP50-95 increased by 3.9% and 4.6% to 80.5% and 50.3%, respectively; parameters decreased by 34.5% to 1.69 M and FLOPs by 33.3% to 4.4 G, realizing dual optimization of accuracy and lightweight design. Distillation comparison experiments showed that CWD had a better optimization effect than other distillation methods, and YOLOv11m was selected as the optimal teacher model by balancing performance and complexity. Mainstream algorithm comparison experiments indicated that RGCE-YOLO outperformed traditional algorithms (Faster R-CNN, RetinaNet, SSD) and other lightweight algorithms (MobileNetV4, StarNet, NanoDet-Plus-m) in detection accuracy, and had obvious lightweight advantages over other YOLO series algorithms, with an FPS of 185.8 meeting industrial real-time detection needs. Generalization experiments showed that RGCE-YOLO achieved a precision of 81.2% (8.4% higher than YOLOv11n) on the NEU-DET dataset and 73.1% (4.2% higher than YOLOv11n) on the GC10-DET dataset, with stable mAP improvements, demonstrating excellent generalization ability. ONNX deployment verification on the ARM industrial computer showed that the model maintained a precision of 81.7% and an FPS of 35.1 with no significant performance decline, proving its adaptability to edge devices. Visual results showed that RGCE-YOLO had a lower missed detection rate, stronger small object feature extraction ability, and more accurate bounding box positioning than YOLOv11n. The excellent performance of RGCE-YOLO stemmed from the synergy of all improved modules: RepLightConv’s reparameterization balanced feature representation and computational efficiency; GSConv and SCSA solved feature redundancy and semantic differences in multi-scale defect detection; the CA mechanism in CMBC_Detect retained spatial details and solved small object detection information loss; CWD enhanced dense object feature capture and generalization without additional computational cost.

RGCE-YOLO effectively solves the problems of low accuracy and poor lightweight performance of existing steel surface defect detection algorithms, and the synergy of the improved modules achieves dual optimization of detection accuracy and model lightweight design. On the self-built dataset, the algorithm not only has significantly higher precision, mAP50, and mAP50-95 than the YOLOv11n benchmark model, but also greatly reduces parameters and computational complexity, with its inference speed meeting industrial real-time detection requirements. It exhibits excellent generalization performance on public steel surface defect datasets and can be stably deployed on edge devices with no obvious detection performance decline, providing a reliable technical solution for real-time industrial steel surface defect detection. The design idea of combining a lightweight network structure, an attention mechanism, and knowledge distillation in this study also provides a reference for the lightweight optimization of other industrial object detection algorithms. In the future, constructing a dataset with more steel surface defect types and conducting targeted improvements for tiny object defects will further enhance the model’s comprehensive detection ability, promoting its wider application in industrial quality inspection scenarios with more complex defect types and higher detection requirements.

Nighttime low-light imaging is often affected by underexposure, limited dynamic range, and amplified sensor noise. These factors weaken edge, texture, and contour cues, and they increase background interference and feature ambiguity. Such degradation reduces detector reliability and causes missed detections, false positives, and localization bias, with more severe impact on small objects, distant targets, and dim instances. Based on YOLOv11n, LMD-YOLO was proposed as a low-light object detection method that integrates multi-scale features and local detail enhancement. Without substantially increasing model complexity, it strengthens multi-scale feature fusion and local detail extraction for low-light images and reduces high noise sensitivity in low-illumination scenes, thereby improving low-light detection accuracy.

LMD-YOLO combined multi-scale feature fusion and local detail enhancement through four components. 1) A Multi-Pool Spatial Pyramid Fast Pooling module (MPSPPF) was inserted into the backbone. It contained parallel pooling branches and a gradient-enhanced concatenation part. The parallel pooling used MaxPool2d and AvgPool2d. Max pooling strengthened salient responses and benefited key-object extraction in low-contrast images. Average pooling produced smoother representations and reduced noise sensitivity. The pooled outputs were fused with weighted aggregation to preserve global structure while retaining local detail cues. MPSPPF also used repeated sequential pooling. The same pooling operation was applied stage by stage to the previous pooled output to form a deeper pyramid. This design enlarged the effective receptive field with limited cost and built a richer feature hierarchy. It captured local and global context that supported recognition of blurred, partially occluded, or weakly illuminated objects. The gradient-enhanced concatenation improved information flow across pooling stages and reduced detail attenuation in early feature extraction. 2) A Cross Stage Partial–Enhanced Dual Hybrid Attention Network module (CSP-EDHAN) was constructed for fine-grained low-light features. Depthwise separable convolution reduced computation while maintaining spatial sensitivity. Residual connections stabilized feature propagation under severe noise. A multi-path fusion structure increased cross-stage information flow and strengthened weak-target cues. A dual hybrid attention mechanism combined channel emphasis and spatial selection. Channel emphasis highlighted informative channels under illumination degradation. Spatial selection focused responses on object regions and suppressed random-noise activations. This module improved feature contrast for dim targets and reduced false activations caused by background clutter. 3) A Dynamic Detail–Semantic Fusion Pyramid Network (DDS-FPN) was designed for neck fusion. A channel–spatial attention unit from the SDFM module guided feature selection during cross-scale aggregation. Shallow detail cues and deep semantic cues were fused with dynamic weights, which raised responses in target regions and lowered responses in noisy background regions. An adaptive downsampling module (ADown) aligned deep semantics with shallow details during scale transitions. It reduced feature misalignment, improved feature consistency across pyramid levels, and compensated for detail loss caused by low illumination. A bidirectional pyramid path strengthened bottom-up detail propagation and top-down semantic guidance. Multi-scale features were aligned and enhanced for small, medium, and large objects under low-light conditions. 4) A dual-group detection head was proposed to balance accuracy and efficiency. Input multi-scale feature maps first passed through a shared feature-extraction stem composed of two grouped convolution layers. Grouped convolution increased feature diversity at low cost and improved localization and recognition. The classification and regression branches shared the stem parameters to reduce redundant computation. Two independent output branches were then applied. Each branch used a 1×1 convolution layer to predict class probabilities and bounding-box distributions. The decoupled branch design reduced task interference and improved training stability for noisy low-light inputs.

Comparative experiments were conducted on EXDark and PASCAL VOC 2012 with YOLOv11n as the baseline. On EXDark, LMD-YOLO improved precision, recall, mAP50, and mAP50-95 by 1.3%, 2.3%, 3.1%, and 2.2%. On PASCAL VOC 2012, the same metrics increased by 1.4%, 1.5%, 1.8%, and 1.3%. Additional evaluation on a complex low-light scene dataset showed gains of 2.2% in mAP50 and 1.3% in mAP50-95 over the baseline. The results indicated that MPSPPF reduced early noise sensitivity through multi-branch pooling and context aggregation. Weighted fusion preserved structure cues while limiting noise amplification. Sequential pooling improved long-range context modeling, which benefited targets with weak boundaries. CSP-EDHAN strengthened weak-target representation through cross-stage fusion and attention-guided enhancement. Depthwise separable convolution preserved efficiency and improved feature robustness. DDS-FPN improved multi-scale alignment through dynamic detail–semantic fusion and attention-based selection. ADown reduced information loss during downsampling and improved semantic–detail consistency across pyramid levels. The dual-group head maintained lightweight computation and improved optimization through decoupled branches. Performance gains on PASCAL VOC 2012 suggested improved general feature quality rather than low-light domain bias. The method also showed stable behavior under mixed illumination and cluttered backgrounds, which are common in real nighttime scenes.

LMD-YOLO improves low-light object detection through noise-robust feature extraction, weak-target enhancement, and dynamic multi-scale fusion while preserving a lightweight design. The method achieves consistent accuracy gains on low-light benchmarks and remains effective under normal illumination. Future work targets broader illumination distributions, stronger robustness to extreme noise, glare, and motion blur, and improved generalization in natural scenes.

Micro-vibrations generated by internal moving components during spacecraft on-orbit operations constitute a critical factor limiting the optical axis stability and imaging quality of space-based optical payloads. Isolating these disturbances and maintaining precise pointing are essential for high-precision tracking and imaging tasks. Incorporating a payload platform capable of both vibration isolation and precision pointing is considered a "multi-win" technical solution. While piezoelectric actuators are common, their limited stroke restricts millimeter-level applications. Voice coil motors (VCMs), conversely, offer large strokes and high linearity but often lack guidance in fast steering mirror applications, leading to lateral drift. Furthermore, traditional hexapod platforms based on the Gough-Stewart configuration, while accurate, involve six-degree-of-freedom fully coupled control. This introduces structural redundancy and control complexity for payloads that primarily require pitch and yaw stabilization. To address the dual requirements of vibration isolation and tracking pointing in pitch and yaw directions, a simplified quadruped platform integrated with voice coil actuation is proposed. The design targets a natural frequency of approximately 5 Hz to facilitate passive isolation of high-frequency disturbances and a pointing adjustment range of ±1° for low-frequency tracking. This study aims to balance structural simplification, large-stroke capability, and high-precision motion, providing a physical foundation for subsequent active vibration control and compound axis tracking systems.

A simplified multi-degree-of-freedom parallel mechanism configuration was adopted. The platform consists of a payload platform, a base, and four supporting legs arranged symmetrically at 90° intervals along a circumference. A push-pull actuation strategy was implemented where legs 1 and 3 control pitch motion, and legs 2 and 4 control yaw motion, enabling decoupled control. VCMs were selected as actuators to leverage their zero stiffness, large stroke, and high resolution. To ensure precise axial motion and eliminate friction, a flexible guidance structure and flexure hinges were developed. A double diaphragm spring system was designed to provide low axial stiffness for actuation compliance while maintaining high radial stiffness to prevent lateral drift and contact between the coil and magnet. The diaphragm springs were designed using 60Si2Mn steel, and Finite Element Analysis (FEA) was used to verify structural integrity. A theoretical model based on dynamic principles established the relationships between leg inclination, pointing range, and resonant frequencies. FEA was further conducted to validate the design, simulating the stiffness of the diaphragm springs and performing modal analysis on a complete 3D model with a 40 kg payload to identify vibration modes. A physical prototype was manufactured, and a specific suspension system utilizing a tripod, tension springs, and flexible cords was constructed. This system effectively offloaded the gravity of the 40 kg payload to simulate a micro-gravity environment, ensuring that the platform's dynamic characteristics were tested without the bias of static gravitational forces. An experimental test system comprising accelerometers, linear optical encoders, and a frequency response analyzer was established to perform sine sweep excitations (1 Hz to 300 Hz), stroke capability tests, and decoupling verification.

FEA results indicated that the double diaphragm spring configuration achieved an axial stiffness of 3.83 N/mm and a radial stiffness of 1703.87 N/mm. Stress analysis showed a maximum stress of 330.33 MPa at the spring's flexure points, which is well below the 1180 MPa yield strength of the 60Si2Mn material, resulting in a safety factor of 3.57. Simulation confirmed that a leg stroke of 2.7 mm was sufficient to achieve a platform tilt of 1°. Modal analysis predicted the first and second order modes (pitch and yaw) at 4.7413 Hz, aligning closely with the 5 Hz design target, while the third mode (Z-axis translation) appeared at 5.7381 Hz. Higher-order modes were pushed beyond 60 Hz, avoiding interference with the control bandwidth. Experimental testing on the prototype demonstrated the following: 1) Individual legs achieved a total stroke exceeding 6.8 mm. Under sinusoidal actuation, the platform successfully verified a pointing range of ±1°, meeting tracking requirements. 2) During single-axis excitation, parasitic motion in the non-driven axis was minimal. For a 2 mm pitch motion, yaw cross-coupling was approximately 2% (0.041 mm), confirming effective decoupling. 3) With a 40 kg payload, measured resonant frequencies were 5.2 Hz for pitch and 5.6 Hz for yaw. These values agree well with the FEA prediction (4.74 Hz). 4) Tests with varying loads validated the dynamic model: 20 kg load resulted in ~7 Hz resonances, and no-load conditions yielded ~16 Hz.

A voice coil-actuated integrated vibration isolation and pointing quadruped platform was successfully designed and verified. The simplified parallel mechanism, utilizing flexure hinges and double diaphragm springs, effectively resolves the conflict between large-stroke pointing and high-precision vibration isolation. Experimental results confirmed the platform meets critical specifications: a natural frequency near 5 Hz under a 40 kg load and a pointing range of ±1°. The study demonstrates that pitch and yaw motions are effectively decoupled, and the experimental data aligns well with theoretical models. This work establishes a verified physical platform capable of passive high-frequency isolation and active low-frequency tracking. Future developments may include integrating this platform with a fast steering mirror to form a two-stage composite control system, enabling coarse tracking via the platform and fine tracking via the mirror for enhanced precision.

Homodyne coherent wireless optical communication enables the direct recovery of baseband signals by interfering a local oscillator with the received signal light under identical frequency and stable phase conditions. This approach provides outstanding detection sensitivity, strong resilience against interference, and high spectral efficiency, making it a promising candidate for next generation high capacity and long reach optical links. However, its practical realization depends critically on maintaining precise and stable phase alignment between the signal and the local oscillator. This requirement imposes rigorous demands on laser source coherence, phase locking accuracy, and adaptive control algorithms. Addressing these challenges is fundamental to harnessing the full potential of homodyne detection in real world dynamic environments such as free space optical channels, which are often affected by atmospheric turbulence, polarization fluctuations, and other time varying impairments.

Significant progress has been made in the key technologies of the robust homodyne coherent receiver, including wavefront distortion correction technology, local oscillator noise suppression technology, phase locking technology, polarization control technology, mixing technology and dual balanced detection technology. Wavefront distortion correction, primarily implemented through adaptive optical systems, compensates for phase errors induced by atmospheric turbulence and optical imperfections. Early conceptual work proposed using deformable mirrors, and subsequent development of wavefront sensors enabled closed loop operation. Modern adaptive optical systems employ multi actuator deformable mirrors, liquid crystal spatial light modulators for static or quasi static correction, and hybrid schemes that combine fast steering mirrors with high order deformable mirrors. These improvements enhance correction bandwidth, spatial resolution, and suitability for real time operation in turbulent channels. Local oscillator noise suppression targets phase, frequency, intensity, and quantum noise sources. Optical injection locking and self injection locking techniques effectively reduce phase noise. Electronic feedback and servo control, often using PID algorithms, provide precise frequency stabilization. Intensity noise is mitigated through balanced detection architectures and by operating the local oscillator at an optimal power level that maximizes the signal to noise ratio. Quantum noise, a fundamental limit, is addressed through techniques such as squeezed state generation and nonlinear optical processing, though these remain areas of active research. Phase locking technology ensures the local oscillator accurately tracks the signal phase. Optical phase locked loops form the core of this subsystem, evolving from discrete component based designs to architectures incorporating digital signal processing. Key developments include decision driven optical phase locked loops, Costas loop based designs for carrier recovery, and hybrid optical phase locked loops that combine optical and electronic feedback. Phase recovery algorithms, adapted from digital coherent communications, play a complementary role. The Viterbi Viterbi algorithm, blind phase search, and Kalman filtering are widely used to estimate and compensate residual phase noise, especially for high order modulation formats like 16 QAM and 64 QAM. Polarization control ensures the polarization states of the signal and local oscillator are matched at the mixer. Early systems relied on mechanical polarization controllers with slow feedback. The field progressed towards integrated electro optic devices based on lithium niobate, offering faster response. Recent trends focus on system level integration, combining polarization tracking with phase recovery algorithms, and on novel devices like metasurface based polarization controllers that offer compact size and programmable functionality. Polarization diversity receivers and polarization insensitive mixer designs also help alleviate the control burden. Mixing techniques generate the in phase (I) and quadrature (Q) components required for complex signal recovery. Fiber based mixers using 3 dB couplers were common but suffer from polarization sensitivity. Integrated optical hybrids on platforms like silica or indium phosphide offer improved stability and compactness. Crystal based mixers, particularly those utilizing the electro optic effect in periodically poled lithium niobate, provide inherent phase accuracy and polarization insensitive operation. Recent advances in thin film lithium niobate enable broadband, low loss mixers suitable for high speed applications. Dual balanced detection converts the optical interference signals into differential electrical outputs, effectively rejecting common mode noise and amplifying the weak signal component. The core of this technology is a pair of matched photodiodes in a differential configuration. Developments have focused on improving photodetector bandwidth, responsivity matching, and integration with planar lightwave circuits. Systems achieving common mode rejection ratios exceeding 40 dB have been demonstrated, directly translating to higher receiver sensitivity and dynamic range.

Homodyne coherent systems exhibit superior sensitivity and spectral efficiency compared to conventional heterodyne and direct detection schemes. The continuous refinement of wavefront correction, noise suppression, phase locking, polarization control, mixing, and balanced detection has progressively lowered implementation barriers, enabling laboratory demonstrations and early field trials. Current performance benchmarks show clear advantages in receiver sensitivity, tolerance to optical signal to noise ratio, and resilience to certain channel impairments. Nevertheless, challenges remain for widespread deployment. Further reduction of phase noise is critical for higher order modulation formats. Managing the coupled dynamics of phase, polarization, and frequency under rapidly varying atmospheric conditions requires more intelligent, adaptive control systems. System complexity and cost also need reduction for practical applications. Future research will likely focus on several key directions: 1) Deep photonic integration, combining lasers, modulators, mixers, and detectors on a single chip using platforms like silicon photonics or thin film lithium niobate to reduce size, power consumption, and cost. 2) Advanced digital signal processing and machine learning algorithms for joint estimation and compensation of multiple impairments in real time. 3) Exploration of quantum inspired techniques, such as phase squeezed states, to surpass classical sensitivity limits. 4) Development of robust and adaptive adaptive optical systems tailored for specific turbulent channel characteristics. Progress along these paths will enhance the robustness, scalability, and energy efficiency of homodyne coherent receivers, solidifying their role in future high performance terrestrial and satellite based wireless optical communication networks.

In clinical practice, pathological examination is considered the "gold standard" for diagnosing various types of cancer. In the field of pathological image intelligent diagnosis, the detection performance of deep learning models typically relies on large-scale and high-quality manually annotated datasets. Due to the high cost and expertise required in annotating medical images, large-scale annotation is often infeasible, resulting in challenges such as limited training data and significant domain variation in clinical scenarios. Therefore, achieving accurate, efficient, and generalizable medical image recognition under limited annotation conditions has become a important challenge in the field of intelligent medical image processing. Mainstream methods adopt multiple instance learning (MIL), which enables whole slide image(WSI) classification using only slide-level labels. However, MIL-based approaches still suffer from issues such as overly concentrated attention allocation and model overfitting when handling complex pathological patterns.

To address these challenges, this study proposes a MIL framework based on a mixed distillation training strategy. First, Starting from publicly available datasets, the processes of image segmentation, image partitioning, and image feature extraction for digital pathology slides have been implemented. The OTSU segmentation method is used to remove the background images that do not contain any pathological information. Each WSI in the dataset is divided into non-overlapping image patchs to facilitate model training and feature extraction. These preprocessing steps help improve the efficiency and accuracy of the model, enabling it to focus more on learning the key pathological features. The self-supervised Vision Transformer-S/16(ViT-S/16) was used to train the model on unlabeled pathological image data. The ViT-S/16, which is a self-supervised learning method based on contrastive learning, can learn key feature representations from images. This process enables the model to learn transferable features from unlabeled data and provide better initialization for the downstream classification task. Then, a mixed distillation mechanism multiple instance learning (MDM-MIL) model integrating dual-path adaptive attention and random instance concealing is designed. The adaptive attention module dynamically adjusts the degree of focus on different texture features and captures semantic relationships between them to further enhance the classification accuracy. The first path adaptive gated attention submodule calculates the attention scores of all instances, and then aggregates the attention scores adjusted by random instance masking to integrate the remaining instance information. The second path uses an adaptive multi-head attention submodule to calculate the attention scores of all instances, and then aggregates the attention scores adjusted by random instance masking to integrate the remaining instance information. This adaptive method not only enhances the model's ability to recognize local features but also significantly improves its understanding of global context. Meanwhile, the random instance concealing (RIC) mechanism randomly masks instances with high attention scores during training, forcing the model to explore additional informative regions and mitigating overfitting to a few critical patches. Mixed distillation mechanism (MDM) integrates the feature information of the remaining instances from the two paths after being processed by the RIC method, thereby enhancing the learning efficiency and classification accuracy of the model. This method captures semantic relationships between different texture features of an image to further increase the classification accuracy. Finally, the classifier achieved accurate classification results on a dataset of breast cancer whole slide pathological images. This paper proposes a multi-instance learning model based on a mixed distillation strategy to address the issues of excessive attention allocation and model overfitting in existing multi instance learning. This model combines a collaborative training dual-path adaptive attention model with a random instance masking method, and ultimately uses a mixed distillation strategy to concatenate the aggregation results of the dual paths. The model can fully learn more key instance feature information while considering the correlation between various types of instances, ultimately improving the performance of the model and alleviating model overfitting.

This study has the characteristics of uneven distribution of categories and weakened sample labels. Therefore, AUC and F1-score are comprehensively adopted as the main evaluation indicators to ensure the scientificity and comprehensiveness of the performance assessment. Experimental results demonstrate the superiority of the proposed method. On the Camelyon16 dataset, the model achieves an accuracy of 98.445%, an AUC of 0.984, and an F1-score of 0.983. On the BRACS dataset, the corresponding values reach 77.852%, 0.905, and 0.764, respectively, showing significant improvements over existing MIL approaches. The experimental results on the Camelyon16 and BRACS datasets show that compared with other advanced multiple instance learning models, This model has significantly better classification metrics, indicating that our model has more advantages in pathological image classification.

This thesis proposes novel algorithm designs and conducts comparative experimental evaluations on corresponding public datasets, achieving superior results compared to the current state-of-the-art methods. This paper contributes to the advancement of automatic analysis techniques for pathological images based on deep learning, which holds significance in assisting cancer diagnosis, and improving the efficiency of clinical analysis.