Small face detection based on improved YOLOv5s

At present, in the complex real-world application scenairos, the task of small face detection encounters numerous challenges, which include small face scale, abrupt lighting changes and low accuracy. In order to address the concerns of overlooking small face detection within existing models, this study introduced a novel small face detection model termed SK-YOLOv5s which was based on convolutional kernel attention mechanism. Firstly, we proposed a small face enhancement module to fuse and upsample multi-layer features, which enhances the resolution of small face feature maps and strengthens their distinctiveness. Subsequently, we incorporated the SKNet attention mechanism into the model, which can adaptively adjust receptive field sizes across multiple scales and enhance the detection efficacy of small face. Finally, EIoU was utilized as the loss function, which directly reduced the width and height discrepancies between predicted and actual bounding boxes, and FReLU was utilized as the activation function, which could enhance the nonlinear expressiveness of feature maps to improve the precision and stability of small face detection. The performance of the enhanced model on the WIDER FACE dataset demonstrated mean average precision improvement of 7.9% over YOLOv5s. The experimental results demonstrate the viability of the enhanced model for small face detection in real-world scenarios.