Design And Implementation Of A Method For Unaligned Fashion Editing

With the continuous and rapid development of public aesthetic appreciation and demand for fashion,more and more consumers are experimenting with independent design and customization of fashion clothing.The advancement of digital and intelligent technology has provided opportunities to solve the challenge of human-computer interaction in the process of fashion design and customization,giving rise to fashion editing,a problem that has received extensive attention from researchers in the field of computer vision and multimedia.Fashion editing is a task where the user interactively edits a real photo of a fashion item.Given a real image of a fashion item,the user makes simple changes to the image,and the algorithm generates a modified real image with high realism.Previous works relied on userset attribute labels,drawn sketches and colors,etc.to make editing directly on the real fashion item image.However,in the real-world fashion design workflow,designers tend to use the design drafts to express their inspiration and show the fashion designs to clients so that they can modify the design drafts based on the feedback,and finally make them into real clothes.For most users without design experience,it is more feasible and natural to perform fashion editing through the design draft with clear sketches,simple colors and rich semantics,in other words,the user first makes changes to the design draft,and then the algorithm generates the modified real photo.Inspired by the fashion design workflow,this paper proposes a design-draft-driven fashion editing framework,in which a given real fashion item image is converted into a design draft by the network,allowing the ordinary users to edit the design draft as freely and conveniently as fashion designers,and then render the edited fashion item image by another network.The goal is that the edited real photo should accurately reflect the design draft modifications in the editing area with clear structure and rich textures,and at the same time,should be consistent with the original real fashion item image in the non-editing area.Based on the design-draft-driven fashion editing framework,this paper further proposes an Unaligned Fashion Editing Network(UFE-Net).To address the serious pixel misalignment between the fashion item image and the design draft,resulting in distorted shapes and textures in the "real photo-design draft-edited design draft-edited real photo" cycle translation,this paper proposes a progressive architecture to gradually reduce the domain gap through the stages of coarse-to-fine alignment,feature-based editing,structure and appearance optimization,and finally obtain high-quality edited results with realistic structures and details.To guide the network to better reason about the content of the editing region,this paper designs an alignment-driven fashion editing module,which facilitates each other by jointly learning alignment and editing tasks to obtain more accurate alignment results and more robust editing results.This paper also proposes a coarse-to-fine alignment strategy and designs a mask-guided local alignment sub-module,which achieves reasonable optical flow inference for editing regions and accurate optical flow estimation for non-editing regions.The sub-module improves the performance of optical flow alignment algorithms for image alignment problems with large structural differences,and provides a higher-quality reference for image editing.In addition,this paper designs a reference-guided refinement module that introduces gated convolution,spectral normalization,and dilated convolution to further optimize the structure and texture,as well as remove artifacts for the edited real image,and improve the overall quality of the edited result.This paper compares UFE-Net with baseline methods on the D2 R dataset.UFE-Net excellently performs design-draft-driven fashion editing task,and significantly outperforms other baseline methods in terms of visual effects and quantitative metrics,which is the stateof-the-art algorithm on unaligned fashion editing task.This paper also verifies the rationality of the design of overall architecture and each module through ablation studies,and confirms good generalization ability through the cases.