The detection of Focal Liver Lesions (FLLs) is crucial for disease diagnosis and treatment. Traditional detection methods face many challenges, and the application of deep-learning technology brings new opportunities. In view of this, this paper systematically reviewed the deep-learning-based FLLs detection methods, and provided specific research directions for the development of FLLs detection technology by analyzing the advantages and disadvantages of related technologies. First, the public datasets of liver radiological images were organized and summarized, and the key role of data preprocessing in improving model performance was expounded. Secondly, the 2D and 3D detection algorithms based on convolutional neural networks, Transformer, knowledge distillation, and other technologies were compared and analyzed, revealing the technical evolution path from local feature modeling to global spatio-temporal correlation. In addition, the temporal feature fusion methods for multi-phase images were examined in depth, providing new ideas for dynamic lesion characterization. The review showed that existing methods had achieved breakthroughs in detection accuracy and efficiency, but still faced challenges such as insufficient sensitivity to small lesions, weak cross-device generalization, and lack of clinical verification. Future research was recommended to accelerate the clinical transformation and application of deep learning in auxiliary diagnosis of liver lesions through multi-center data collaboration, lightweight algorithm design, and enhanced interpretability.

To address the challenges of morphological similarity and the resulting difficulty in classifying vehicle damage images uploaded by clients for auto insurance claims, a model named ResAWDNet was proposed for vehicle damage classification. Firstly, to effectively augment the model’s capacity for extracting damage features, the traditional down sampling operation was replaced with weighted dynamic convolution. This approach dynamically adjusted the weights of convolutional kernels based on the input features, thereby enhancing the model’s adaptability to features of varying scales and orientations. As a result, it enabled more precise capture of the subtle differences in vehicle damage. Secondly, to ensure that the model could concentrate on the salient discriminative regions and feature channels within the images, a dual attention mechanism was embedded after the convolutional layers of the backbone network. This mechanism concurrently learned the important weights in both spatial and channel dimensions, significantly enhancing the model’s ability to capture crucial information. Consequently, it further enhanced the decision-making accuracy of the model in the task of vehicle damage classification. Finally, experimental validation was conducted based on a dataset of vehicle damage images sourced from real accident cases. The experimental results demonstrated that the ResAWDNet model was feasible and offered significant advantages for vehicle damage classification tasks, achieving an accuracy rate of 73.79%. Compared with baseline models, ResAWDNet achieved higher accuracy in classifying multiple types of damages, robustly validating the effectiveness of the proposed model.

Existing research on multi-view stereo scheme utilizes depth-estimation algorithms to achieve stereo representation by establishing a mapping relationship between the physical and digital worlds. Supervised learning-based neural networks have achieved accurate and high-fidelity 3D reconstruction results through training. However, in-the-wild visual reconstruction remains challenging due to the lack of rendered depth priors and wide-baseline characteristics of images. A novel system was proposed to obtain optimized depth for naturally collected multi-view images without prior information by applying an unsupervised learning network and semantically optimized Neural Radiation Field (NeRF) rendering. First, preliminary depth information for wild multi-view images were produced without ground truth based on unsupervised deep learning. Subsequently, in a separate NeRF module, a diffusion model was used to construct a surface semantic rendering loss, enabling a fine-grained volumetric representation. Experimental results on the benchmark dataset validated the performance of the proposed system by improving an average of 24.6% of the overall metrics, compared with other state-of-the-art schemes. A novel wild wide-baseline dataset was also applied to verify the generalization performance, and the proposed system reduced the reconstruction error by up to 40.8% compared with all methods.

As large models continue to grow in scale, the memory footprint and computational overhead of model inference have become critical challenges. Mixed-precision quantization is an effective approach to reduce resource consumption, but existing methods suffer from insufficient outlier handling, significant quantization accuracy loss, and inefficient memory access. To address these issues, a memory-aligned mixed-precision quantization method for large models was proposed. First, weights were divided into SIMD-aligned groups, and outlier groups were identified via group-wise significance analysis, with high-significance groups quantized to 8 bit and others to 2 bit. A block-wise compensation strategy was introduced to mitigate accuracy degradation caused by 2 bit quantization. Furthermore, an efficient packing and storage scheme was designed for mixed-precision weights, where a bitmap was used to record the bit width of each data block, enabling random access. Experimental results demonstrated that the proposed method significantly reduced memory usage and improved computational efficiency while maintaining model accuracy. Specifically, on Llama2-7 B/13 B/70 B, the approach achieved perplexity reductions of 8.13/2.84/1.37 on WikiText-2 and 5.80 on C4 relative to state-of-the-art baselines. The quantized 70 B model reduced weight storage by approximately 87% compared with BF16. Across seven QA benchmarks, an average accuracy gain of 6.24% was achieved. Last, these results indicated that a mixed-precision quantization method for large language models via memory alignment could simultaneously improve compression ratio, memory-access efficiency, and overall model performance.

With the rapid development of deep learning, remarkable achievements have been made in image classification and related tasks. However, the success of these models heavily relies on large amounts of high-quality labeled data. In real-world applications, labeled data is often scarce, and manual annotation is time-consuming, labor-intensive, and costly, which limits the scalability and deployment of deep learning models. In recent years, active learning has gained significant attention due to its ability to improve model performance under limited annotation budgets. The core idea of active learning is to select the most valuable data for labeling based on certain criteria such as uncertainty, diversity, or representativeness. To address the limitations of traditional active learning methods, which often rely on manually designed heuristic sampling strategies that struggle to adapt to different task scenarios and are difficult to dynamically optimize, a Smart Reinforcement Active Learning (SRAL) approach for image classification is proposed. The sample selection process is modeled as a MARKOV DECISION PRocess (MDP), leveraging reinforcement learning’s adaptive strategy optimization ability to guide the model in dynamically selecting the most valuable samples from the unlabeled data for labeling. In this framework, the state is represented by features extracted from the unlabeled samples, the action indicates whether a sample should be selected for labeling, and the reward function is defined as the change in model accuracy after incorporating the selected sample into the training set. The Actor-Critic algorithm is adopted to optimize the sampling policy, and uncertainty-based heuristic ranking is incorporated as auxiliary information to improve the learning efficiency. Experimental results demonstrate that the proposed SRAL method significantly improves classification accuracy under the same labeling budget compared to other active learning approaches on datasets such as CIFAR-10, SVHN, and FASHION-MNIST. Furthermore, SRAL exhibits robust stability and strong generalization ability across these datasets. This confirms the effectiveness and advantages of SRAL in enhancing the performance of image classification models.

In night-vision environments, image quality was significantly degraded by glare from intense light sources, impairing the performance of night-vision assisted driving systems. Existing flare-removal algorithms suffer from limited robustness, high computational complexity, and loss of light-source information. To address these challenges, a lightweight image flare-removal method, Night Flare Removal Network+ (NFR-Net+), was proposed to enhance image clarity while meeting the real-time computational demands of mobile devices. The approach first incorporated a feature-filtering mechanism combined with residual connection strategies to strengthen feature extraction capabilities, effectively mitigating overfitting and ensuring robust flare removal across diverse lighting conditions and flare types. Additionally, a nonlinear, activation-free feature attention module was introduced. Via a lightweight design, an efficient attention mechanism was constructed that significantly improved image-detail reconstruction while reducing model parameters by approximately 8.28% and runtime memory by about 11.1%, thereby optimizing computational efficiency. To tackle the issue of diminished image naturalness due to excessive light-source removal in traditional methods, an enhanced light-source extraction module was developed within the segmentation network. This module employed an improved light-source separation strategy to accurately preserve brightness and texture details in light-source regions, ensuring the authenticity and naturalness of output images. Experimental results demonstrated that NFR-Net+ surpassed state-of-the-art methods on image quality metrics such as Structural Similarity Index Measure (SSIM), Peak Signal-to-Noise Ratio (PSNR), and Learned Perceptual Image Patch Similarity (LPIPS), exhibiting superior flare-removal performance and detail preservation. The method also demonstrated strong adaptability across various night-vision scenarios and hardware devices, fulfilling the efficiency requirements for real-time processing. Ablation studies further validated the effectiveness of individual components, highlighting the critical role of feature filtering and attention mechanisms in balancing performance and resource consumption. This approach provided an efficient, lightweight solution for applications such as nighttime autonomous driving and intelligent surveillance.

The organizational management of construction sites is a critical aspect in engineering management; however, traditional human supervision method is constrained by many environment limitations and low efficiency. In recent years, multiple government departments have issued relevant policies advocating deep integration of artificial intelligence with the real economy to promote high-quality and efficient economic development. The accuracy, efficiency, and automation advantages of Computer Vision (CV) technology have gradually led to its widespread application in the field of construction supervision. Meanwhile, the drones, which can efficiently obtain complex and varied visual data of construction scene, demonstrate their application potential in CV-based construction supervision tasks. However, the current researches on drone-based construction scene detection are limited, and the lack of overhead-perspective construction-scene image datasets restricts further development in the field. Therefore, the DJI Mavic 3T drone was utilized to obtain construction-site images to establish an open-source overhead image dataset for construction scene UB-CSD. Several advanced object-detection algorithms were selected for comparative experiments on the UB-CSD dataset, and the reasons for performance differences were analyzed from multiple dimensions such as model workflow design, computation principle, and task characteristics. The mAPs of every algorithm’s detection result were YOLOv8 and YOLOv10 (96.1%), YOLOv9 (96.0%), YOLO11 (95.7%), DETR (95.3%), Faster-RCNN (76.3%) and RetinaNet (72.1%). The analysis results indicated that the YOLO series algorithm constituted the most optical algorithm for drone-based object detection tasks in construction scenes. By establishing a new open-source special dataset and conducting comparative experiments, the conclusion drawn provided effective data and experimental cases to support future safety production management and object-detection algorithm research in the construction industry.

Traditional 3D point-cloud classification methods tend to suffer from insufficient generalization and catastrophic forgetting in Few-Shot Class-incremental Learning (FSCIL) scenarios. The pretrained vision-language model CLIP (Contrastive Language-Image Pre-training), which contains rich 2D shape priors, has been shown to effectively enhance 3D FSCIL performance. However, existing CLIP-based frameworks still lack flexibility and adaptability in multimodal feature extraction and fusion, which limits classification accuracy during incremental stages. To address these shortcomings, a 3D FSCIL approach with deeply fused multimodal features was proposed. An adaptive adapter based on gated units and residual blocks was introduced to achieve multi-scale feature alignment and redundancy suppression, and a multimodal global feature dynamic fusion module with self-attention was designed to adaptively adjust the weight allocation of different feature streams according to sample characteristics, thereby obtaining more consistent and complementary fused representations. Specifically, point clouds were rendered into multi-view depth maps, and features were extracted using both the original CLIP visual encoder and a CLIP encoder pretrained on depth maps, combined with point-cloud geometric features. After processing through the adaptive adapter, these features were fed into the attention-based fusion module and aligned with semantic features extracted by the CLIP text encoder for classification. In addition, contrastive learning loss, multi-view and geometric perturbation-based data augmentation strategies, and a memory-replay mechanism were incorporated to effectively mitigate overfitting and forgetting under few-shot conditions. Experiments on ShapeNet, ModelNet, and CO3D demonstrated that the proposed method consistently achieved higher accuracy across incremental stages compared with existing 3D FSCIL approaches, while significantly reducing both relative accuracy drop rates and maximum stage fluctuations.

To address the significant decline in matching accuracy and the number of correspondences exhibited by existing image-matching methods under large viewpoint variations, an improved image-matching approach based on E-LoFTR was proposed. Firstly, based on a strategy of viewpoint rectification followed by fine-grained matching, a novel two-stage SIFT-based viewpoint-rectification module was proposed, which leveraged the viewpoint invariance of the Scale-Invariant Feature Transform (SIFT) algorithm and the geometric alignment capability of homography to enhance matching accuracy under large viewpoint variations. Then, a directional-gated attention mechanism was designed that employed a cascaded structure of multi-directional convolutions and dynamic gating to extract queries (Q), keys (K), and values (V). The injected geometric priors significantly enhanced the model’s robustness. Lastly, to mitigate information loss during the upsampling of fused features, the Fusion-DySample module was incorporated to further improve performance. Experimental results on the public MegaDepth dataset showed that our method achieved relative pose estimation AUCs of 57.1%, 72.7%, and 83.9% under rotation error thresholds of 5°, 10°, and 20°, respectively, outperforming E-LoFTR by 0.7%, 0.5%, and 0.4%. On the newly constructed NewMega dataset based on MegaDepth and on a private industrial dataset, our method also demonstrated substantial improvements in both the number of matches and matching accuracy.

In intelligent traffic monitoring systems, small target vehicle detection in complex scenes faces challenges such as low feature resolution, severe occlusion interference, computational redundancy, and insufficient bounding-box regression accuracy. To balance detection accuracy with deployment efficiency on edge devices, an improved YOLOv8 framework based on dynamic sparse attention and a lightweight dual-branch structure was proposed. The method first introduced a bidirectional routing sparse attention mechanism (ReBiAttention) that enhanced the retention of shallow features for small targets by dynamically filtering key features through a two-level routing strategy. Subsequently, GSConv and VoV-GSCSP modules were integrated to reduce computational cost while dynamically adjusting multi-scale feature weights. An improved DynamicHead was applied for multi-task adaptive optimization, and a modified ShapeIoU loss function with shape- and scale-aware weighting was employed to improve localization accuracy. Experiments on the UA-DETRAC dataset showed that, relative to baseline YOLOv8n, Precision, Recall, and mAP@0.5 increased by 8.739%, 1.685%, and 7.225%, respectively, while the parameter count decreased by 4.3%. This method provided an efficient solution for accurate detection of small-target vehicles in complex traffic scenarios.