Improving the Lightweight Classroom Face Detection Algorithm of YOLOv5

This study addresses issues in images captured by classroom surveillance cameras, such as unclear facial contours and indistinct texture features. Particularly, the small size of students' faces in the back rows increases the likelihood of recognition errors or omissions. To improve recognition accuracy and accommodate deployment on mobile devices, we refined and lightweighted the YOLOv5s algorithm. Firstly, we replaced the original CIOU loss function with the EIOU loss function to optimize the training model and achieve faster target recognition. Secondly, we introduced the GhostConv module and CBAM attention mechanism from GhostNet to enhance feature extraction capability and optimize the network structure. Through ablation experiments, we confirmed that the improved algorithm maintains high recognition accuracy while reducing the parameter count. Experimental results demonstrate significant reductions in Params and FLOPs compared to the original version, with only a 0.6% decrease in mAP and a 0.7% decrease in Precision. This makes it more suitable for deployment on mobile devices, effectively addressing facial detection challenges in classroom environments.