Research On Face Detection Algorithm Based On Deep Learning

Face detection is an important research topic in the field of computer vision.In recent years,face detection technology has been gradually applied to people's daily lives,such as system monitoring,access control,network platform,supermarket platform and so on.At present,there are many challenges in face detection,such as low resolution,facial occlusion,face deflection angle,strong illumination,and multi-pose.Although early face detectors can be used to detect faces well,it is because of these factors that these detectors do not perform well on existing publicly challenging face data sets.This thesis studies the face detection under the framework of deep learning,and learns and analyzes the existing target detection algorithms Faster R-CNN and R-FCN.The improvements were made on the existing models,and the two data sets of WIDER FACE and FDDB were tested and analyzed.The contents are as follows:(1)Introduce the framework structure,principle,alternating training and end-to-end training methods of Faster R-CNN algorithm.Experiments were carried out on the dataset WIDER FACE.The experimental results were used to analyze the comparison of the two training methods in the image detection effect.The end-to-end training method of Faster R-CNN improved the detection performance of the network by sharing the convolutional layer.(2)According to the test results of the experimental Faster R-CNN,it is known that Faster R-CNN is not effective in the more challenging data set WIDER FACE.To this end,a context-based convolutional neural network is employed.It mainly uses the information of multi-layer convolution layer and RoI pooling to process tiny facial regions,introduces body information,and uses L2 normalization and RoI pooling to realize the fusion of face information and body information.Using the online difficult sample mining strategy,the difficult samples are added to the network for training,so that the context-based convolutional neural network has a good detection effect on the challenging data set WIDER FACE.(3)Analyze the principle of the existing target detection algorithm R-FCN.Since the existing target detection algorithm R-FCN is not a special face detection algorithm,the existing target detection algorithm R-FCN is improved.The R-FCN-FACE algorithm isproposed,and the third and fourth convolution feature maps are used for multi-scale training techniques,and online difficult sample mining and other strategies have achieved good results on the data set.