YOLO-SOD: Color probability map-based fuzzy framework using deep models for face detection

This is the authors' abstract. We don't add key points for this paper.

Detecting small faces remains challenging due to scale variation and limited discriminative cues at low resolutions. We address this with a two-pronged approach. First, we modify YOLOv5 by adding extra convolutional layers and small-object–focused detection heads that operate on larger, higher-resolution backbone feature maps, improving sensitivity to tiny targets. Second, we introduce a dynamic, image adaptive patching strategy for scale selection. We compute Cr channel color probability maps and measure the number and density of pixel clusters; a Fuzzy Inference System (FIS) uses these cues to select the most suitable patch size for near and far faces. On the WIDER FACE benchmark, our method attains 93.07% mAP@50 (Easy), 91.25% mAP@50 (Medium), and 84.45% mAP@50 (Hard). On our proposed local dataset, the proposed model achieves 49.36% mAP@50 for different size faces. Beyond faces, the approach generalizes to small object detection on VisDrone-DET2019, reaching 46.76 mAP@50. These results demonstrate that coupling small-object-aware architecture design with color-guided, fuzzy patch selection yields competitive mAP@50 for small and distant targets.