Face Detection Algorithm Based On DCNN And Faster RCNN

This paper proposes a feature fusion single-stage CNN(convolutional neural networks)face detection algorithm and a cascade CNN face detection algorithm based on faster RCNN.First,this paper improves single-stage CNN face detection algorithm by feature fusion.To obtain features with better discrimination,this paper trains two different networks to extract features and fuses the features of the two networks.Then we adopt PCA to reduce dimension and get more compact feature,followed by SVM for binary classification.In the test stage,we first get image pyramid to detect multi-scale faces and then adopt sliding window method to obtain each window's feature.To reduce the computational complexity,we convert the network into a fully convolution network to share the convolution calculation and get feature for all sliding windows with only once forward computation.In the last pooling operation,offset pooling is used to reduce the step of the sliding window and thus improve the accuracy of bounding box.Experiments show that the fusion features are more discriminative and can better describe the face/non-face features.Second,this paper trains a cascade CNN face detector with faster RCNN and makes some improvements on faster RCNN.In the region proposal stage,we add the anchor scales and restrict the positive and negative anchor ratio to further balance ratio of positive and negative samples.In addition,we consider the special structure of face to achieve face classification and regression combined with online hard example mining(OHEM)to further improve the performance.In the fast RCNN stage,we replace the original fully connected layer with the response maps of the key parts of the human face.This greatly reduces the parameters and improves the speed.To further improves the performance of the detector,we adopt online hard example mining in our face detector.We achieve recall rate of 95.9%on FDDB with 500 false positives,and 95.71%with 351 false positives,which is slightly lower than Xiaomi INC,which achieves the best result of 95.89%recall rate with 351 false positives.And the average test speed is 100fps when using GPU.