Research On High Resolution Tiled Clothing Generation Method Based On Image Fusion

The number of online shoppers is increasing every day with the development of Internet technology.And the sales of clothing have taken a large proportion of revenues from online products available for browsing and purchasing.In this background,how to efficiently retrieve and recommend online clothing products has become an urgent problem to be solved.Therefore,a method that can effectively extract the features of specific clothing regions is required to ensure the accuracy of subsequent clothing retrieval and recommendation systems.However,traditional approaches based on object detection technologies trend to crop clothing regions firstly,and then use the extracted images as the input to the associated systems.Nevertheless,the extracted features’ quality of these methods is often harmed by complex background and clothing distortion.Then some improved generation-based methods can translate the model images into tiled clothing ones directly,so as to avoid the disturbance of various noises.But the generation-based methods are sensitive to the quality of input images and the generated clothing can’t restore the textures and patterns in high resolution.The performance of the subsequent system depends on the extracted feature will further enhanced if the improvement about these approaches can be made for overcoming these defects.In the existing image generation methods,the performance of the generative adversarial network can deliver a relatively better effect.The general image-toimage translation technology based on GAN has been applied in many fields.So,in this study,a method based on GAN and image fusion is constructed to translate the clothing region in wearer to a tiled clothing;and on this basis,a highresolution tiled clothing generation method is further proposed.Specifically,a GAN based on cycle-consistence and feature loss is used to translate the input clothing regions to tiled clothing templates.Further,a context-based TPS transformation is applied to align the clothing region with the generated template and a refinement network fuses the template and warped clothing to generate a characteristic preserved tiled clothing.Finally,a super-resolution operation is performed based on the generated tiled clothing to enhance the resolution,then a refinement network further fuses the generated tiled clothing and clothing region in high resolution for texture restoration.For verifying the superiority of the proposed model,a large-scale paired dataset between human body clothing and tiled clothing images is created by using a crawler system.Based on this dataset,various experiments were carried out.Experimental results demonstrate that the resultant images generated from the proposed framework can preserve the internal features of clothing region in wearer and deliver higher quality than other existing methods in terms of the reality and the retrieval performance.