Deep Learning Based Image Synthesis And Interactive Editing

It always been a challenge task in the field of artificial intelligence that how to synthesize photorealistic images.Image is a common media bearer in people's daily life.The demand for image synthesis and editing is increasing,and the requirements are also stricter.Whether it is popular retouching or professional designers,image synthesis algorithm is needed to support.Recent studies have shown that the deep learning based algorithm has made a significant breakthrough in the field of image editing and synthesis.Neural networks try to synthesize images by learning semantic information and potential features in images.Image synthesis is a pixel level task,and it is difficult to synthesize high-resolution images.In this work,we will based on the deep neural network,to generate high-quality and high-resolution garment images from the garment contours map.This algorithm can be used to assist the workflow of garment production.Our major contributions in this work can be summarized as follows:1.We propose the bi-colored edge representation which reaches a good balance between the quality of the synthesized images and the workload of users.Existing image synthesis algorithms try to provide a small texture patch to generator,and propagate the texture patch to the corresponding regions of the given contour map.Although this simple input setting requires less effort for user,it is difficult to generate a high-resolution garment image just from a small-size texture pattern.Therefore,based on the above discussion,we propose the bi-colored edge representation,our experiment demonstrate this representation has sufficient capability to generate diverse high-quality garment images.2.We propose to interactively drawing shading edges and separately infer the shading information.During the test phase of the image generator,we found that it is hard to draw the shading information,and lead to the lack of stereoscopic sense in the output garment images.Thus,to make the generated image contains a better stereoscopic shape,we introduce a shading generator to enhance the shading on garment image by drawing some sparse binary edges.In this way,the user achieves more control on the shading information of a garment.3.We propose an elaborate interactive system for fashion images design.Based on garment image generator and shading generator,the interactive system is completed in combination with the Easy Gui module of python graphics development library.Our interactive system consists of three modules,the drawing plate,the output plate and the color regulator,which can meet the demands of garment design.Thanks to the powerful generation ability of image generator,the flexibility of the bi-colored edge representation and the shading generator to increase the photo-realism of the results,our interactive system can design diverse and colorful garment.