The Research Of Face Detection Algorithm Based On Deep Learning

Face detection is one of the most classic applications in pattern recognition.Due to the complexity of natural scenes,the uncertainty of illumination and the complex and changeable face of human faces,face detection algorithm faces a huge challenge.Benefit from the development of deep learning,making the machine through learning to get better characterization algorithm of the data,at present,deep learning in academia and industry have achieved better than the traditional algorithm,has gradually been highly valued.This topic starts with the face candidate region,and combines the simple threshold segmentation of CMYK and HSV color space,apply Gaussian modeling in YCbCr to reduce the missing detection rate of skin color detection.Considering the convolution neural network as the depth increases,the expression of the model will be better,So the candidate regions obtained by the skin color detection are sent to the convolutional neural network for classification,and finally determine the face.However,due to the false detection of skin color detection,the number of candidate regions is large,so this topic adopts the method of full convolution neural network to generate the face heat map.According to the local hottest location on the heat map,the face candidate area is obtained,improving the accuracy and speed to a certain extent.Finally,the method of acquiring candidate regions in the heat map is integrated with the multi-scale network which based on the body information.The network takes into account the spatial relationship between the body and the human face,and uses the body information to carry out the assistant detection of the face.The face candidate regions and body regions are sent together into a multi-scale convolutional network,the low-level detail texture information and the high-level semantic information are merged to obtain the features of the face candidate region and the body region,and then the two features are merged and then subjected to regression classification.In the concrete realization,this topic adopts the open source deep learning framework Tensorflow&Caffe,uses OpenCV to carry on the image processing,uses Python to carry on the programming in the PC side,has obtained very good effect on the test data set.