Research And Optimization Of Face Selection Algorithm

As a popular technology in the moment,face recognition has achieved high recognition accuracy in academia.However,under unconstrained conditions,due to the limitations of the acquisition device and the complexity of the acquisition environment,the quality of the captured face images is uneven,making face recognition not getting better performance.This thesis studied the face select algorithm.It is expected to obtain high quality pictures of the same individual,filter low quality pictures,and improve the robustness of face recognition in practical applications.The paper divided the face screening algorithm into two parts,face quality assessment and face recognition.The qualities of the faces will be scored by quality assessment part,and the low-quality images that may affect the accuracy of the recognition algorithm will be removed.After this,face images will be recognized by face recognition part to achieve selecting purpose.The specific work is as follows:(1)In terms of face recognition,this thesis proposed a new loss function,improved the existing network,and achieved a higher recognition accuracy.This thesis eselected the CASIA-Webface data set as the training data set of the algorithm,and preprocessed the data set through the MTCNN algorithm and the expansion method of random horizontal mirror flip.Mobilefacenet was used as the base network to design Mobilefacenet-1 with faster computing speed.The computational complexity of this network is much smaller than that of the mainstream face recognition model,which is only 1/2 of Mobilefacenet,while the recognition accuracy is not much lower than the original Mobilefacenet.Several popular loss functions were evaluated and visualized by plotting the two-dim features on the Mnist dataset.Based on Focal Loss,the thesis proposed Focal-angle Loss with a much more specific semantic definition for difficult samples.The face recognition model trained by the loss function achieved higher recognition accuracy than Focal Loss.(2)In terms of face quality assessment,this thesis proposed a new face quality score calibration method and designed a new face quality assessment model using convolutional neural network.This thesis proved through experiments that low quality images can affect the performance of face recognition algorithms.A new face quality assessment method was proposed by using class weight vector as the standard face feature vector.By compared the EvR(Error versus Reject)curves with the existing method of manual selection,the new method was confirmed to be a better guideline on the face recognition algorithm.By constructing a network that combines low-level features and deep features,a good face image quality assessment model was trained.The guideline of the model was verified on the EvR curve.