Research On Clothing Image Generation From Sketch Based On Deep Learning

Sketch based image generation is a research hotspot at present.Many image tasks about sketch generation have achieved good results,but the specific objectives are still some relatively simple objects,such as face,shoes and animal posture.These objects often have similar structures and are relatively simple in online and local depiction.This paper mainly studies how to generate clothing image by generating confrontation network.Clothing has complex texture structure and diverse styles.Image generation based on clothing sketch has certain difficulty.Firstly,the description of lines of clothing is much more complex than that of human face and cat face,and the structure is more various than animal posture,The attributes of clothing,such as color,texture,style and style,are unmatched by simple shoes.Therefore,this paper carries out experimental research on this problem.This paper mainly does the following three points:1.Generating countermeasure network to realize image conversion needs a large number of image data that can be used for learning as support,but due to the lack of clothing data set,it can't meet the research conditions.Therefore,a solution is proposed to construct the clothing data set needed in this experiment.The experimental data set of this paper is constructed by using the ready-made image target saliency edge detection algorithm to reverse generate clothing sketches from the collected real clothing images.2.Based on the existing diversified image conversion network,this paper focuses on the improvement of the generator of its network architecture.By adding the attention mechanism module to the generator of u-net structure,and adjusting the discriminator and encoder in the network architecture,the performance of the whole network is improved and the generated image is clearer.3.Based on the existing image conversion network,its generator and discriminator are improved.The attention module is added to the generator to improve the effect of image generation.Two scale discriminators are used on the discriminator to restrict the generator from generating higher quality images in global perception and local detail perception.In the loss function,the design of feature matching loss function and perceptual loss function is introduced to make the generated image more realistic.The experiments in this paper show that the method in this paper can generate more real clothing images.From the data comparison and data analysis of the experimental results,it can be proved that the diversified image conversion network improved in this paper has significantly improved compared with other relevant methods in the index of FID,and the image conversion network improved in this paper has significantly improved compared with other relevant methods in SSIM,PSNR and perceived loss.