Face detection algorithm based on improved Retinaface

In view of the insufficient use of the semantic information of high-level feature maps under the existing face detection algorithms, resulting in unsatisfactory face detection results in complex field conditions, an improved face detection algorithm based on Retinaface is proposed. Use the backbone feature extraction network ResNeSt-50 to extract features, combine the feature pyramid CE-FPN to perform multi-scale fusion of feature maps of different scales, and then transfer the information processed by the feature pyramid to the context module, and finally pass the training and loss function Soft non-maximum suppression to realize face detection and feature point positioning. Experimental results show that the average accuracy of this model on the WiderFace dataset is 94.8%, which is better than the original Retinaface model.