Improved DSFD Algorithm with Enhanced Feature Fusion and Information Utilization

Face detection on small-scale faces and faces in extreme poses is challenging but significant to many applications like intelligent monitoring and access control system. An improved DSFD face detection algorithm was proposed to address insufficient feature fusion and insufficient information utilization. Firstly, the original backbone of DSFD, ResNet, is replaced by the improved residual network, IResNet, to further enhance the extraction of facial features. Secondly, ConvNeXt V2 Block is adopted to combine with the feature enhance module to further enhance feature fusion and the detection effect for small-scale faces. Finally, hybrid dilated convolution is used to protect the continuity of information while expanding receptive field of the enhanced feature maps from feature enhance module, which enhances the utilization rate of face information and the detection effect for faces in extreme poses. The improved algorithm was evaluated on WIDER FACE and the feasibility of methods are evaluated through ablation experiments. The accuracy of the improved algorithm surpasses the original algorithm. Compared with other face detection algorithms, the improved algorithm has achieved better performance.