To address the issue of missed detection for small-scale occluded pedestrians in dense scenes, this paper proposes an improved YOLOv8 detection algorithm named Dense-YOLOv8. Firstly, to resolve the difficulty of extracting features from small-scale pedestrians in dense environments, a backbone network enhanced with deformable convolution and dynamic convolution is adopted to improve feature extraction capabilities. Simultaneously, a multi-scale linear spatial attention module is designed to amplify features of visible parts of occluded pedestrians while suppressing interference from complex backgrounds. Secondly, a small-scale pedestrian detection head is introduced in the neck of the YOLOv8 network to enhance detection performance for diminutive pedestrians. Finally, to improve training efficiency, a novel weighted loss function named DFL-SIoU is developed to accelerate model convergence. Experimental results demonstrate that the proposed algorithm achieves superior performance on two challenging dense pedestrian datasets, CrowdHuman and WiderPerson, significantly enhancing detection capabilities in dense scenarios. Comparative evaluations with other state-of-the-art pedestrian detection models further confirm the strong competitiveness of the proposed model.

In pedestrian detection, small-scale pedestrians often face challenges such as limited pixel values and insufficient features, often leading to wrong or missed detection. Therefore, this paper proposes a small-scale pedestrian detection method based on multi-scale structure perception and global context information. Firstly, to address the issue of decreasing features caused by the network deepens, we design a feature fusion strategy to overcome the constraints of the feature pyramid hierarchy. This strategy combines deep and shallow feature maps and leverages the advantages of Transformer to capture long-distance dependent features, incorporating a global context information module to retain a substantial amount of small-scale pedestrian features. Secondly, considering the confusion between small-scale pedestrian features and background information, we employ a combination of self-attention modules and channel attention modules to jointly model the spatial and channel correlations of feature maps. This utilization of small-scale pedestrian context and channel information enhances small-scale pedestrian features while suppressing background information. Finally, to address the issue of gradient explosion during model training, we introduce a novel weighted loss function named ES-IoU, significantly improving the convergence speed. Extensive experimental results on the CityPersons and CrowdHuman datasets demonstrate that the proposed method achieves a substantial improvement upon state-of-the-art methods.

To address pedestrian occlusion and multi-scale issues in crowded scenes, a crowded pedestrian detection algorithm that integrates global context information, spatial information, and pose information is proposed. Firstly, the feature pyramid network (FPN) is enhanced by integrating a dynamic sampling module and a weighted fusion branch, enabling more effective multi-scale feature aggregation and better handling of pedestrian scale variations. Secondly, a global context information module is designed to capture additional potential pedestrian features by fusing global context information and spatial information, which mitigates feature loss due to occlusion. Additionally, an enhanced Transformer-based pose module is constructed to extract pose information as supplementary resources, tackling the challenges of occlusion in crowded pedestrian detection. Since pose estimation is limited by pose estimation algorithms, an improved CA attention is incorporated into the visual information to complement the pose information. Finally, a novel data augmentation method is proposed to simulate occlusion, thereby improving the model’s generalization ability. Extensive experimental evaluations on the CrowdHuman and CityPersons benchmarks corroborate the robustness and superior detection capabilities of the proposed algorithm, which consistently surpasses current state-of-the-art approaches in crowded pedestrian detection scenarios.

Micro-expression recognition (MER) remains a formidable challenge in affective computing due to the subtle, localized, and fleeting nature of facial muscle actuations. Conventional spatial-temporal networks are easily overwhelmed by static facial topologies, leading to feature representations that are heavily biased toward identity-specific noise. To address this, we propose the Motion-Anchored Cross-Modal Fusion Network (MACFN), a novel dual-stream ViT architecture that explicitly decouples and synergizes spatial appearance and optical flow dynamics. Specifically, we introduce a motion-anchored spatial attention module, which translates latent motion features into a sparse spatial probability mask. It acts as an enhancement gate, forcing the texture stream to bypass static backgrounds and attend to genuine ME-related regions. Furthermore, we design a cross-modal bilinear fusion module to capture the second-order interactions across modalities, mapping the coupled features into a discriminative semantic manifold. Extensive experiments conducted on the CASME II, SAMM, and SMIC databases under the rigorous leave-one-subject-out composite database evaluation protocol demonstrate that MACFN is effective and achieves competitive performance compared to several recent methods.

Class-Incremental Learning (CIL) aims to enable models to continuously learn new categories while preserving existing knowledge and avoiding catastrophic forgetting. Although parameter-expansion architectures can alleviate task interference to some extent, the representations of previously learned classes often drift or degrade as the feature subspaces continuously evolve and expand, resulting in decreased recognition performance for old classes. To address this issue, we propose an efficient CIL method—Dynamic Gated Adapter for Subspace Alignment (DGASA). Based on a frozen pre-trained backbone, DGASA introduces lightweight adapters with attention-based gating for each task to construct task-specific subspaces, while dynamically fusing cross-task information via attention mechanisms. In addition, DGASA learns a linear mapping between the old and new subspaces to achieve consistent alignment of old class prototypes in the current subspace without accessing past data. Extensive experiments demonstrate that DGASA significantly improves classification accuracy and resistance to forgetting on multiple benchmark datasets, offering strong generalization and computational efficiency.