The Research Of Face Detection Algorithm Based On Convolutional Neural Network

Under unconstrained conditions,mainly referring to complex natural scenes,illumination,angles,and large-scale scales of faces and occlusions,many existing algorithms can no longer meet the real needs simply by changing the network model structure.Based on face detection process and convolutional neural network,several algorithms are proposed to improve the detection accuracy and speed of face detection.Firstly,a face detection method based on multi-resolution fully convolution neural network is proposed.Firstly,a multi-resolution sliding window is used to generate a face heatmap with multilevel resolution.The face candidate region is obtained according to the local hottest region on the heatmap,and finally the face candidate regions are sent to the existing classification network for classification and face location is obtained.Considering that the multi-resolution sliding window generates a lot of unnecessary calculations when generating the face candidate region,this paper proposes a scale and spatial forecast network.SSFN is a lightweight fully convolution network,in view of the large scale range of human face on image,the network can forecast the range of face scales that may exist on the image,which can effectively reduce the amount of redundancy calculation caused by the sliding window.Finally,considering the SSFN-based face candidate region detection method,the actual receptive field is much smaller than its theoretical receptive field,which is not enough to capture the global semantic information,for a certain degree of false detection problem by the face candidates to screen the human face.This paper proposes a multi-context face detection network,which uses the depth separable convolution method to add local context information around the target pixel to the original feature map without adding additional computational burden.The global average pooling method is used to add global context information,and finally multiple context information features are merged and input into Softmax classification regression.In the algorithm implementation,this paper uses Caffe to realize the deep learning algorithm framework,and achieved good results on FDDB,LFW and MALF data sets.