Posture and texture guided fashion costume design synthesis method

Abstract

The invention discloses a posture and texture guided fashion costume design synthesis method. The method comprises the steps of 1, collecting task data by means of an existing fashion data set, preprocessing the data, and constructing a fashion image, posture information and semantic information data set; 2, by taking generation of a natural and accurate fashion image as a target, constructing a two-stage generation model; the generation model comprises a semantic layout generation network and a texture generation network, effective texture transfer is realized, and diversified fashion images are generated; 3, training a semantic layout generation network and a texture transfer network by using the collected data groups under the condition of minimizing adversarial loss, cross entropy loss, pixel-level loss, perception loss and style loss; and 4, training network parameters in the generated model through a back propagation algorithm until the whole model converges, and generating a corresponding fashion image. According to the method disclosed by the invention, an experiment is carried out on a Fast-Gen data set, and good results are obtained in both quantitative and qualitative aspects.

Description

technical field [0001] The present invention mentions a new method of Pose and Texture Guided Multi-View Fashion Design Synthesis (Pose and Texture Guided Multi-View Fashion Design Synthesis), which mainly involves a method of using a semantic layout generation network to convert input human body poses into a series of A method for the semantic layout of the human body, and a method for implementing texture transfer using a texture transfer network and generating realistic fashion images through a generative adversarial network. Background technique [0002] Due to the high demand for real-life applications and the breakthrough development of related theories and technologies such as deep learning, machine learning, computer vision, and multimedia technology, the task of combining artificial intelligence with fashion has received considerable attention in recent years, such as clothing recognition, In terms of clothing retrieval, fashion recommendation, and fashion trend pre...

AI Technical Summary

Problems solved by technology

[0006] Human body posture is usually represented by two-dimensional joint points, which only contain human body joint point information, not shape information. Existing methods are difficult to infer human body structure and clothing structure from rough posture information.

[0007] 2) Difficulty in implementing texture transfer

[0008] Due to the locality of feature processing in ordinary convolutional networks, there is no dedicated texture transfer mechanism in existing fashion image generation methods to achieve effective transfer of fashion image textures

Second, since clothing regions are usually irregular, how to use texture blocks of any size to accurately transfer textures to corresponding clothing regions and synthesize natural and realistic fashion images is also a challenge.

The existing fashion image generation methods only realize the generation of pure color textures, and cannot effectively transfer complex textures. What are generally realized is the generation of local textures or the generation of incorrect textures.

[0009] 3) Diversity limitation of fashion clothing generation

[0010] Existing fashion image generation methods generally use human body posture information or human body semantic information for guidance, and the clothing structure type is fixed, so it is impossible to generate fashion images corresponding to various clothing categories and fashion styles in real situations

Images