Research And Application Of Face Detection Based On Convolution Neural Network

The Convolutional Neural Network(CNN)is a distinct kind of deep neural network model,which is a new type of artificial neural network produced by the combination of artificial neural networks and deep learning networks.At present,CNN has obtained better results in the field of computer vision.The face recognition in practical applications is confronted with the influence of factors such as background,lighting,posture and facial expression in an unconstrained environment.Therefore,in recent years,the face detection algorithm with CNN for feature extraction has been continuously proposed,which can solve the problem that the precision rate in wild data sets is not high,and the method has achieved good results.Based on the CNN construction model,this thesis designs a multi-scale fusion convolutional neural network(MSF-CNN)structure for training face detector.As a feature extraction method,CNN has the ability of active learning.Therefore,when designing the MSF-CNN network structure,this thesis takes advantage of the ability to learn features autonomously.It divides the middle layer output into three channels for convolution operations,and sets convolution kernels of different number and size for each channel to obtain three-scale convolution features of the samples;then the features of different scales are normalized and then are scaled;finally,the classifier is connected to train the classifier.The main contributions of this method are as follows:(1)The multiscale feature extracted by the model contains more useful information,so that the classification accuracy is higher than the single-scale.(2)Compared with the cascaded CNN model,the scale becomes smaller and the complexity decreases.(3)To achieve endto-end training,there is no need to train multiple CNN models separately as in a cascade structure.Experiments show that the performance of the MSF-CNN model designed in this thesis is better than some existing detection algorithms on the public data set.Face detection and gender recognition are both two-category tasks,but there are certain differences.In the classification tasks of the former,faces and backgrounds are quite different.However,for the latter one's,the difference between male and famale is relatively smaller than the former one.Therefore,it is necessary to combine local features and overall features to train a classifier with good performance when solving gender identification problems.In this thesis,the MSF-CNN model is fine-tuning on the gender dataset,and the model is used to extract the deep global features,and then a simple shallow S-CNN model is selected to extract the local features of the shallow layer of gender samples.The combined local features and global features are classified by SVM algorithm.The experiments show that the improved network based on the MSF-CNN model has achieved good results in gender classification tasks,and is superior to some gender identification methods.