Research On Unsupervised Image Style Conversion Algorithm Based On Cycle GAN

Image style conversion aims to convert the source domain image to the target domain image by designing an end-to-end model.Generally,the source domain provides the content of the image,and the target domain provides the "style" of the image(which can be image attributes or image style).Under the source domain content,the "style" of the target domain is realized,thus realizing the conversion of the source domain image to the target domain image.That is,image style conversion can be label map to scene map,line contour to color image,spring,summer,autumn and winter scene,or day and day.As long as it meets the task of end-to-end conversion,it can be achieved through image style conversion.At the same time,image style conversion can play an important practical role in style transfer,attribute transfer and image resolution improvement.Unsupervised image translation algorithms solve the problem of unmatched data sets by adding constraints,but they do not consider the particularity of such problems when processing object conversion tasks,making these algorithms still face the following challenges:(1)Due to the existence of network bottleneck layers,a lot of useful information has been lost here,resulting in significant changes in the background color of the generated image before and after style conversion.(2)In the object conversion task,because the network itself has no attention mechanism,the target that needs to be converted cannot be detected,resulting in no conversion at the place that needs to be converted,and conversion occurs at the place that does not need to be converted,such as generating target texture patterns in the background area.(3)How to select image quality evaluation criteria.The subjective evaluation result is the closest to the real quality of the image,but the operability is too poor.The evaluation result is affected by the evaluator’s educational background,evaluation motivation and other factors.Although objective evaluation methods have the advantages of simple use and easy integration,they often cannot truly reflect the quality of generated images,and there are fewer objective evaluation criteria for object conversion.This paper mainly studies the algorithm of image style conversion based on Cycle-GAN.Aiming at the problems of existing models in processing style conversion tasks,two improved schemes are proposed.Based on the method of adding jump,through the understanding of the basic idea of Cycle-GAN and the learning of network architecture,Cycle-GAN is used as the baseline to complete the basic framework of the model,and the jump connection is added to the generator network structure to realize the connection of output and input,and the low-level information is transferred to the generated image to solve the problem of low-level information loss and background distortion of the generated image.Based on the method of attention mechanism,understand the core idea of attention mechanism,obtain the areas that need most attention in the process of image style conversion by introducing attention mechanism,and use the key captured information to complete the establishment and connection of context information,so as to emphasize the areas of interest to us and suppress the irrelevant background areas,and prevent the target detection failure leading to the conversion of places that do not need conversion.After the training and validation of the model,this paper loads the corresponding weights for the final evaluation and comparison of the model.Master the physical meaning and evaluation methods of different evaluation indicators,and propose to use FID 、PSNR、and SSIM as evaluation indicators to evaluate the effect of image style conversion.Finally,the robustness of the model is analyzed and summarized,and the effect of style conversion is displayed on 2-3 data sets,and compared with the original Cycle-GAN model.