Face Detection And Face Gender Recognition Based On Convolutional Neural Networks In The Natural Scenes

Facial features are crucial in the areas of public safety,intelligent services and human-computer interaction,because it has a large amount of information such as gender,expression,identity,and age.Face detection and face gender recognition relying on facial features are the focus of research in the field of computer vision,However,due to the natural scenes,the face images will be affected by such factors as light,posture,expression,occlusion,and size,which affects face detection and face gender recognition.Based on the theory of convolutional neural networks,this thesis deeply studies face detection and face gender recognition in the natural scenes.The main contents are as follows:(1)Face detection based on four cascade fully convolutional neural networksTraditional face detection has the problems of poor manual extraction feature characterization and low matching of features and classifiers.In view of the above situation,this thesis designs a face detection algorithm based on four cascade full convolutional neural networks.The basic idea is to construct the four-level cascade network,cascade training is used instead of end-to-end training to avoid the limitation of sharing only one weight.A deep network with differentiated functions can be obtained,which could improve the detection accuracy.In addition,each level of network is designed as a full convolutional structure to increase the detection rate.Then using the bootstrap method to optimize the network model.Finally the first three levels of network gradually filter and the final level of network correct output to get the face detection results.Experimental results show that the algorithm has good robustness to multi-pose,occlusion,and different skin color face types.The detection speed reaches 86 ms on a single picture,the true positive rate reaches 90.62% on the FDDB dataset.(2)Multi-scale face detection based on depth residual networksAt present,many convolutional neural networks have a simple structure and a single detection model,and there are still many missed-detection situations for face images with small size,large change in pose and more serious occlusion under the natural scenes.In view of the above problems,this thesis presents a multi-scale face detection algorithm based on deep residual network.The basic idea is to use a 18-layer depth residual network structure as a basic network,and design face detection modules of different sizes for different positions at the end of the basic network,which can increase the ability of the model to detect the different scale face,while CReLU is also introduced as the activation function to ensure the accuracy while improving the detection rate.We do some Experiments from multi-groups of datasets,which show that the true positive rateof this algorithm reaches 95.52% on the FDDB dataset,and the average detection accuracy can reach 93.30% on the Wider Face test set,which can effectively detect scale conversion and occlusion and other complex scenes under the face of the image,with high detection accuracy and robustness.(3)Face gender recognition based on Multi-layer feature fusion convolutional neural networksAt present,for face gender recognition,the recognition accuracy of some traditional gender recognition algorithms and convolutional neural network algorithms is not very high.There will also be misidentification.In order to further improve the face gender recognition accuracy;this thesis presents a face gender recognition algorithm based on multi-layer feature fusion convolutional neural network.The basic idea is based on the AlexNet network structure,while the feature map of the shallow Conv2 convolutional layer output is merged with the feature map of the deep Conv4 and Conv5 convolutional layer output after 2 times upsampling,fuse the characteristics of the multi-layer convolution layer to improve the characterization ability of the network model;In addition,Large-Margin Softmax Loss with adjustable target supervision mechanism is introduced as the loss function of the model to enhance the guidance of deep convolutional network learning,make the intra-class spacing between the same sex smaller and the inter-class spacing between different sex larger,Get better gender recognition.Finally,the algorithm of this thesis is tested and verified on six standard face datasets.Experimental results show that the recognition accuracy of this algorithm is higher than that of other traditional convolutional networks.