Semantic feature refinement of YOLO for human mask detection in dense crowded

Due to varying scenes, changes in lighting, crowd density, and the ambiguity or small size of targets, issues often arise in mask detection regarding reduced accuracy and recall rates. To address these challenges, we developed a dataset covering diverse mask categories (CM-D) and designed the YOLO-SFR convolutional network (Semantic Feature Refinement of YOLO). To mitigate the impact of lighting and scene variability on network performance, we introduced the Direct Input Head (DIH). This method enhances the backbone's ability to filter out light noise by directly incorporating backbone features into the objective function. To address distortion in detecting small and blurry targets during forward propagation, we devised the Progressive Multi-Scale Fusion Module (PMFM). This module integrates multi-scale features from the backbone to minimize feature loss associated with small or blurry targets. We proposed the Shunt Transit Feature Extraction Structure (STFES) to enhance the network's discriminative capability for dense targets. Extensive experiments on CM-D, which requires less emphasis on high-level features, and MD-3, which demands more sophisticated feature handling, demonstrate that our approach outperforms existing state-of-the-art methods in mask detection. On CM-D, the Ap50 reaches as high as 0.934, and the Ap reaches 0.668. On MD-3, the Ap50 reaches as high as 0.915, and the Ap reaches 0.635.