Research On Face Detection Method Under Complex Conditions Based On CNN

With the continuous breakthrough of hardware performance and the rapid development of computer vision,face recognition,identity verification,expression recognition and other face applications are favored by a wide range of users,which promotes the continuous development of face detection technology,and has become one of the most studied topics in the past decade.However,there are many uncertain factors in the real environment,such as illumination change,face occlusion,different posture,various expressions and different scales of human face,which make it difficult for face detection technology to be widely used in real life.Among them,face occlusion and multi-scale human face have the most serious impact on the performance of face detection,which is the existence of most face detection algorithms Main issues.In this paper,the face detection algorithm based on deep learning is analyzed,and the relevant improvement methods are proposed for the problems of face occlusion and multi-scale face,and applied to the more mature Faster R-CNN model at present.After experimental verification,the improved algorithm model proposed in this paper has better detection performance in these two aspects than most similar algorithms.The specific work of this paper is as follows:1.For face occlusion.Firstly,an improved region cross entropy loss function is proposed to learn the local features of the samples.The feature map is directly related to the category by the way of global average pooling,and then the contribution of the local region features to the classification is increased,so that the detector is more sensitive to the local region features.Secondly,aiming at the problem of missing face detection in positive samples,a more reasonable soft non maximum suppression method is used to obtain the best candidate frame,reduce the number of overlapping candidate frames,improve the recall rate of occluded face,and reduce the situation of missing face detection.Finally,aiming at the problem that the number of occluded face samples in the existing database is small,through the data-driven way,the occluded face samples are made manually,so that the model can fully learn the occluded samples,and then improve the detection rate of the occluded face in the network.2.For multi-scale face problems.This paper discusses the difficulties of face detection algorithm on different scales.Firstly,a multi-scale fusion strategy is proposed.The feature pyramid is constructed by using the feature extraction network to output the feature map which is gradually reduced.Through the two fusion of top-down and bottom-up,the final output feature contains more semantic information and location information,and enhances the Lu of the feature The robustness improves the sensitivity of the algorithm to small-scale face.Secondly,we improve the representation of anchor frame,increase the recommendation scale of candidate frame,match the output of multi-scale fused feature pyramid with anchor frame of appropriate scale,and improve the recall rate of conventional face and small-scale face.Finally,in order to alleviate the imbalance of positive and negative samples in the training process,the online hard case mining method is used to mine the hard cases in the samples,and through the way of back propagation to learn many times to enhance the detection ability of the model for the hard cases.