Fuzzy Color Maps and Modified YOLOv5 Sharpen Small Face Detection

Detecting small faces in images remains a stubborn problem in computer vision, because tiny targets offer few distinguishing features and vary enormously in scale. A new framework called YOLO-SOD tackles this challenge by combining an adjusted deep detection architecture with an adaptive, color-guided patching strategy that decides how an image should be divided before analysis.

The first prong of the approach modifies YOLOv5, a widely used object detection network. The researchers added extra convolutional layers and detection heads focused on small objects, and these heads operate on larger, higher-resolution feature maps from the backbone. Because fine details are often lost as feature maps shrink through a network, working from higher-resolution maps improves sensitivity to tiny faces that standard configurations tend to miss.

The second prong is a dynamic, image-adaptive patching strategy for scale selection. The method computes color probability maps from the Cr channel and measures the number and density of pixel clusters in these maps. A Fuzzy Inference System takes these cues as inputs and selects the patch size best suited to the image, choosing finer patching for near faces and coarser patching for far ones.

On the WIDER FACE benchmark, the method reached 93.07% mAP@50 on the Easy subset, 91.25% on Medium, and 84.45% on the Hard subset. On the team’s proposed local dataset covering different face sizes, the model achieved 49.36% mAP@50.

The framework also generalizes beyond faces. Tested on the VisDrone-DET2019 dataset, which features small objects captured from drones, it reached 46.76 mAP@50. The authors report that coupling small-object-aware architecture design with color-guided, fuzzy patch selection yields competitive mAP@50 for small and distant targets.

Subject of Research: A fuzzy inference framework using Cr channel color probability maps and a modified YOLOv5 for small face and small object detection

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

Article References: Saba, A., Shirazi, M. F., Rehman, Y., & Khan, M. M. (2026). YOLO-SOD: Color probability map-based fuzzy framework using deep models for face detection. PLOS One, 21(10), e0360794. https://doi.org/10.1371/journal.pone.0360794

Image Credits: AI Generated

DOI: 10.1371/journal.pone.0360794

Keywords: face detection, YOLOv5, small object detection, fuzzy inference system, Cr channel, color probability map, WIDER FACE, VisDrone-DET2019, computer vision, deep learning, mAP, YOLO-SOD