Face Detection Based On Convolutional Neural Network

Face detection is the first step of face analysis,which is to judge whether there is a face in a given image,and mark the location of the face if there is.Although the traditional algorithm can detect the positive face image well,it cannot achieve the desired effect due to the influence of skin color,expression,shielding,illumination and other factors or in other complex environments.Since convolutional network can automatically extract features and reduce the influence of external factors,more and more scholars have applied convolutional neural network to face detection.However,most of the algorithms with high accuracy have more layers or use complex cascade structure,which requires a lot of training time and is difficult to realize real-time detection in practical applications.Therefore,based on convolutional neural network,this paper studies face detection and proposes two fast and accurate algorithms.The main work is as follows:(1)A multi-scale face detection method based on full convolutional neural network is proposed,which can quickly and accurately complete a single detection task for specific application scenarios such as online payment and traffic.Firstly,the full connection layer of the convolutional neural network model AlexNet,VGGNet and GoogleNet is changed to the full convolution layer to realize the input of any scale image and reduce the training parameters.Then,the classification layer was changed to face and non-face dichotomy,and the training accuracy reached 99.08%,98.99% and 98.98%,respectively.At last,the classification model with the best training results is used for face detection.After the detected image is input into the full convolution network through multi-scale transformation,the probability matrix of the feature image can be obtained,and then the non-maximum suppression is used to obtain the most accurate face frame.The detection results show that this method has high accuracy,short detection time and good performance in face detection.(2)An end-to-end multi-scale face detection method is proposed to improve the single detection.The image is divided into several grids and the face is detected on each grid with the prediction box.The specific work is as follows: first,in the training of convolutional neural network,an "anti-blocking layer" is added to preprocess the original image to enhance the robustness of the network against occlusion.Second,the ordinary convolution structure is changed to DL-CNN structure to extract more comprehensive features.Third,an improved inception structure is added between thefirst and second convolution to enable the network to extract features at different levels.Fourth,the improved spatial pyramid pool layer is added to realize the input of any scale image.Fifth,the loss function commonly used in detection algorithm is improved to increase the accuracy of the model.The test results of open face data set show that this method can extract meaningful facial features and has better performance than many competitive face detection algorithms.