Face Photo-Sketch Synthesis And Recognition Based On Unsupervised Generative Adversarial Network

Face photo-sketch synthesis and recognition have been widely used in many fields,especially in tracking suspects,which can facilitate the investigation of cases by generating sketch images of criminal suspects.At present,only a few forensic sketchers can reach the professional level,which is difficult to meet the needs of target reconnaissance and solving cases.Face photos and sketch images are heterogeneous images,and it is not efficient to identify face photos directly from sketch images.To synthesize high-quality sketch face images that are helpful for recognition,the main work content of this paper is as follows:(1)A face photo-sketch synthesis and recognition method based on unsupervised adaptive generative adversarial network is proposed.In terms of constructing the dataset and data pre-processing,different scenarios are reasonably divided into training and testing sets using machine learning methods,and the generalization ability of the model is enhanced through data pre-processing.In terms of network structure,the residual layer and adaptive normalized residual layer of the generative adversarial network generator are added,and a multi-scale model is constructed on the basis of different residual layers and adaptive normalized residual layers to extract the deep and shallow features of the image.Through ablation experiment and comparative experimental analysis and comparison,the effectiveness and advanced nature of the network are shown in terms of synthesis quality,recognition efficiency and complexity of the network model,respectively.(2)A method of face photo-sketch synthesis and recognition based on unsupervised attention generation adversarial network is proposed.The attention mechanism helps to pay better attention to the detailed features of images in the study of computer vision.Therefore,the self-attention module is added to the generator of the generative adversarial network,and the residual layer and adaptive normalization layer are reduced by half in the up-sampling-down-sampling modules before and after,and the residual layer and adaptive normalization layer are reduced on the multi-scale model,simplifying the network structure.After the introduction of the experimental environment and experimental details,the effectiveness of the network was verified under the ablation experiment,and the synthesis,recognition,and model complexity results were analyzed in the comparative experiment,and the synthesis quality and recognition accuracy of the proposed method were improved.(3)A facial photo-sketch synthesis and recognition method based on unsupervised lightweight generative adversarial network is proposed.To achieve better qualitative effect of network training,this method can use classification activation mapping image to qualitatively determine the key area of the network in the generated image,so as to improve the efficiency of model training optimization.The number of convolutional layers of the discriminator in the generative adversarial network is reduced,the number of residual layers in(1)and the number of adaptive normalized residual layers in the generator before and after up-sampling-down-sampling modules are retained,and the number of residual layers and adaptive normalization layers in(2)are retained in the multi-scale structure,while the loss function is improved.While realizing the network synthesis quality and recognition efficiency,the model lightweight is greatly improved,and the hardware resource cost is reduced.In terms of result analysis,the ablation experiment on the necessity of the network and the superiority of the loss function is carried out,which verifies the advanced nature of the method in this chapter,and compares other unsupervised and semi-supervised methods,and qualitatively and quantitatively analyzes the synthesis quality,recognition efficiency and model complexity to show the effectiveness of the proposed method.It can effectively improve the quality of image composition,improve recognition efficiency,and realize lightweight improvement of existing models.