智能科学与技术学报 (Jan 2024)
Small Face Detection Based on Improved YOLOv5s Network
Abstract
In the complex real-world applications, the task of detecting small faces encounters numerous challenges, which include issues related to scale variation, abrupt lighting changes and suboptimal precision. In order to address the concerns of overlooking small facial structures within existing models, this study introduces a novel approach termed SK-YOLOv5s which is based on convolutional kernel attention mechanism. Firstly, we proposal a small face enhancement module to fuse multi-layer features, which can enhance the resolution of small faces and improve its features.Subsequently, we incorporates SKNet attention mechanism into the model to adaptively adjust receptive field sizes across multiple scales so that the detection efficacy of small face can be enhanced. Finally, EIoU is utilized as the loss function, which facilitate a direct optimization of the width and height discrepancies between predicted and actual bounding boxes. This approach enhances the precision and stability of small face detection. The performance on the WIDER Face dataset demonstrate a mean detection accuracy improvement of 3.1% over prevailing methods. The experimental results demonstrate the model's viability for detecting small faces in real-life scenarios.
