Research On Face Sketch Synthesis Algorithm Based On Generative Adversarial Networks

Face sketch synthesis is a branch of the field of image style migration and a hot research direction in the field of computer vision.The purpose of face sketch synthesis is to obtain the sketch images corresponding to other faces based on a given face and corresponding sketch images,and generate the sketch patterns after face style migration,which can be more truly output by the confrontation network.Face sketch synthesis is widely used in many industries such as life,criminal investigation,digital entertainment,cartoon production and film production.The traditional generation confrontation network is widely used in image style migration,but the sketch images obtained by these methods are not comparable to the real sketch,On the one hand,the mapping process of generating confrontation network from learning face to sketch mainly depends on the automatic learning and definition of potential loss function when training the network.On the other hand,it has a sense of abstraction when the face style is turned into sketch,and the style of sketch image is changeable and random,which increases the challenge to obtain fresh sketch map.Therefore,the generation of confrontation network itself has certain requirements for the training data set.Based on the above existing situation,this paper designs and improves the face sketch synthesis algorithm as follows.(1)Face sketch synthesis based on Cycle-Generative Adversarial NetworksThis paper proposed Cycle-Generative Adversarial Networks(Cycle GAN)has been used for Face sketch synthesis.By constructing multi-scale Cycle GAN;Combined with VGG16 module,it is designed into deep supervision U-Net + + structure;Design dense jump connection to extract multi-scale features from image information;In order to make the network reasonably share the weight of pixel low-frequency information and high-frequency information,a pixel attention module is added in the generator.(2)A multi-scale Self-attention mechanism Cycle GAN based face sketch synthesis is proposedCycle GAN with Self-attention mechanism is proposed.At the same time,reduce network redundancy;Improve down sampling and up sampling in U-Net structure,so as to improve feature resolution and obtain detailed information;The designed multi-scale feature aggregation module uses multiple parallel hole convolution with different sampling rates to integrate the spatial information from different scales;Capture image information in multiple proportions while maintaining a large receptive field of the image;Finally,in order to capture the feature dependencies in the spatial dimension and channel dimension,the pixel Self-attention module is designed to model the semantic dependencies in the spatial dimension and channel dimension,so as to enhance the performance of features.Compared with five classical algorithm models on the benchmark data set,the algorithm shows better performance.