| In recent years,with the continuous increase of information transmission rate and network storage and the rapid popularization of image acquisition equipment,hundreds of millions of face images have emerged on the Internet,which provides many data samples for related research topics,such as face recognition and face attribute recognition.At the same time,with the rapid development of GPU computing performance,it becomes feasible to train complex depth networks.Therefore,deep convolutional neural networks have made important progress in many fields of image recognition.In face related recognition tasks,deep convolution neural network has become the mainstream method with its superior performance.In the task of face attribute recognition,because the face image may contain noisy and dynamic background,to remove the background interference and directly analyze the face part,the existing methods mainly include two types,one is to cut the face image through the boundary box,and the other is to directly segment the face image through the traditional semantics.The former may make the face image contain unnecessary parts,while the latter may lead to the loss of face texture details.Therefore,to extract the face and retain the local details of the face,this thesis designs and implements a dual path face attribute recognition algorithm based on deep learning and develops a face attribute recognition system based on this.The main work includes:(1)Abstract face images are generated based on the conditional generation countermeasure network Pix2 pix HD.The abstract face image retains the face texture information based on the face semantic segmentation image and removes the background interference at the same time.It is used to assist the face attribute recognition algorithm to locate the local part of the face,which effectively improves the performance of the algorithm.(2)Based on the convolution neural network architecture in deep learning,a dual path face attribute recognition algorithm is designed and implemented.Dual path face attribute recognition network includes original human face subnet and abstract human face subnet,which makes full use of the characteristics of original face image and abstract face image.At the same time,to further utilize the information from the abstract subnet,the features of the abstract human face subnet will be mapped and transmitted to the original human face subnet,and all the feature information of the two subnets will be obtained through feature fusion,which effectively improves the accuracy of face attribute recognition.The algorithm is evaluated and ablated on the public data sets Celeb A(Celeb Faces Attribute Dataset)and LFWA(Labeled Faces in the Wild Attributes Dataset),and the effectiveness of the algorithm is verified in the laboratory.(3)Based on BRPC framework,a face attribute recognition system is designed and implemented.The system can complete the tasks from face detection,face alignment to face attribute recognition.It includes multi face and multi-attribute recognition functions,achieves high recognition accuracy and fast recognition rate,and realizes the landing of face attribute recognition algorithm based on deep learning. |
