Abstract
Aiming at the mismatch between speed and accuracy of DSFD face detection algorithm, a lightweight network method based on improved DSFD is proposed. First, the backbone of the algorithm is replaced with EfficientNet B1, which greatly reduces parameters and computation. Then the feature pyramid network in feature enhance module is changed to BiFPN to further enhance the feature fusion. Finally, to improve the efficiency, dilated convolution layers in feature enhancement module are removed, and the convolution layers of BiFPN is replaced by ACBlock. The improved algorithm was evaluated on WIDER FACE and the effectiveness of the proposed methods are evaluated through ablation experiments. The detection speed of the improved algorithm is increased by twice. Compared with two popular algorithms, the improved algorithm achieves higher accuracy.
Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
