YOLO-OSAM: Reassembly Spatial Attention Mechanisms for Facial Expression Recognition

Facial Expression Recognition (FER) is crucial for accurately interpreting human emotions in human-computer interactions.However, FER remains challenging due to many variations, such as facial expressions, head poses, and illumination.Spatial attention mechanisms in Convolutional Neural Networks (CNNs) help address these challenges by enhancing feature extraction, directing focus to crucial facial regions while suppressing irrelevant information.However, traditional spatial attention modules, which apply average and max pooling followed by a convolutional layer, may have limited capacity to capture complex spatial dependencies, leading to suboptimal feature representation in FER tasks.YOLOv5 was selected from among various YOLO series because of its ability to deliver high accuracy object detection and classification, its lightweight architecture, and overall efficiency to overcome these limitations, we propose YOLO-OSAM, an enhanced version of YOLOv5 designed to refine feature learning for FER.Our approach introduces (1) a fusion layer that integrates attention mechanisms to ensure robust feature extraction across varying facial expressions and (2) an enhance spatial attention mechanism incorporated into the YOLOv5 architecture to capture fine-grained facial details.In this paper, the proposed attention mechanism module separately applies max and average pooling to generate feature maps, which are refined through three convolutional layers with batch normalization and Leaky ReLU activation.These processed maps are then concatenated and further optimized using additional convolutional blocks and SoftMax activation, with residual connections enhancing feature representation.Finally, we integrate this enhanced attention mechanism into the YOLOv5 neck, improving feature extraction and refinement.Experimental results demonstrate that YOLO-OSAM achieves 79.7%, 41.7%, and 98.1% accuracy on the RAF-DB (basic), RAF-DB (compound), and CK+ datasets, respectively-outperforming the original YOLOv5 by 1.3%, 0.8%, and 1.6%.Additionally, YOLO-OSAM surpasses baseline models such as VGG16, YOLOv3, and YOLOv5, highlighting its effectiveness in enhancing FER through improved spatial attention and feature extraction.