Research On Face Detection Algorithm Based On Improved Faster-RCNN In Natural Scene

In recent years,with the development of artificial intelligence,face detection as an important branch of computer vision has also become one of the hot research topics.In reality,many face detection scenes are everywhere,such as mobile phone face unlocking,face payment,station security,etc.At present,many face detection methods proposed at home and abroad have reached a high level of detection under certain constraints.Generally,these methods require that the input face image background is simple and required front face.However,in real natural scenes,the face scale is too small,the face pose is distorted,the illumination intensity is different,foreign objects are occluded,and the face image is blurred,and other factors cause the above methods to reduce the final detection accuracy and lead to high missed detection rate.Aiming at the above problems,this paper analyzes the basic process of face detection,studies the classic object detection algorithm Faster-RCNN,and combines the specific problem of natural scene face detection to improve Faster-RCNN.The accuracy of face detection under natural scene conditions is improved,and the rate of missed detection is reduced.The main work of this article is as follows:(1)The original Faster-RCNN backbone network VGG16 is not sufficient for deep-level image feature extraction,especially for some important fine-grained features of the face,which may cause the loss of important features of the face part.In this regard,this paper uses a deeper residual network ResNet-50 to extract facial features.In addition,in order to take into account the correlation between low-level features and high-level features,and fuse the context information of the network so that the network can detect faces of different scales,this article uses a multi-scale feature map fusion strategy to fuse the features of different convolution layers feature maps.The experiments were performed on the public face dataset WIDERPACE and tested on the FDDB face detection dataset,the results show that when the deeper residual network ResNet-50 is used and the feature maps of different convolutional layers are fused,the performance of the model is significantly improved.(2)In order to fully detect small-scale faces,this article designs a more refined Anchor in Faster-RCNN's RPN network.In addition,when the poses of the faces in the images are different,the faces are partially occluded,multiple faces are overlapped,and the image is not sufficiently exposed,the features extracted by the network are insufficient to detect these hard samples.In this regard,this article adds the Online Hard Example Mining Algorithm(OHEM)to concentrate the hard case samples encountered during training again into the network for training,making the network training more adequate.Meanwhile,the Soft Non-Maximum Suppression Algorithm(Soft-NMS)is added in order to solve the problem of missing detection caused by multiple faces overlapping to a certain extent.In addition,the Residual Attention Mechanism is introduced to make the model pay more attention to the facial features in the image and ignore the noise.The experiments were performed on the public face dataset WIDERFACE and tested on its test set.The experiments show that the improved network model can better adapt to face detection under natural scene conditions,the average accuracy of the detection and the recall rates have all improved.