Virtual Dressing Image Generation via Posture-Aligned Feature Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current virtual dressing technologies face challenges in generating detailed and authentic images of a user wearing target clothes, due to the rough image features extracted by existing image fusion models, leading to distortion and poor virtual dressing effects.
Innovation Solution
The proposed method involves obtaining a first body image and a first clothes image, transforming the clothes image to match the posture of the target body, and then performing feature extraction on the transformed clothes image, a bare area image, and the body image to obtain fine-grained features. These features are then used to generate a second body image with the target clothes, preserving detailed information and achieving a high-authenticity virtual dressing effect.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If feature extraction is performed by using an image fusion model, then the virtual dressing process can be completed, but the extracted image features are rough and lack detailed information, leading to distortion of image generation effect
Solution Approach 1:
The patent segments the feature extraction process into three independent streams: clothes feature extraction from the clothes image, skin feature extraction from the bare area image, and body feature extraction from the body image. This segmentation allows each stream to focus on extracting specific detailed features without interference, thereby improving the overall precision of feature extraction and subsequent image generation.
Solution Approach 2:
The patent introduces a new dimension by separately processing the bare area image to extract skin features. This additional feature dimension complements the traditional clothes and body features, enabling more precise reconstruction of skin areas in the virtual dressing result and reducing distortion.
2Ease of operation
If the clothes image is not transformed to match the body posture, then the processing is simpler, but the posture mismatch leads to poor virtual dressing effect
Solution Approach 1:
The patent performs clothes image transformation as a preliminary action before feature extraction and image generation. By pre-aligning the clothes image posture with the body posture through transformation operations, the system ensures that subsequent feature extraction and synthesis operations produce accurate virtual dressing effects without requiring complex real-time adjustments.
3Device complexity
If only the first body image and first clothes image are used, then the data processing is simpler, but the detailed information is insufficient for high-quality image generation
Solution Approach 1:
The patent extracts and isolates the bare area from the body image to create a separate bare area image for dedicated skin feature extraction. This extraction ensures that skin-related detailed information is not lost during the virtual dressing process, as skin features are captured from a specialized input rather than being mixed with body features.
Solution Approach 2:
By introducing the bare area image as an additional input dimension, the system enriches the feature space available for skin reconstruction. This additional dimension provides dedicated skin information that complements the body image features, reducing information loss in skin areas without significantly increasing overall system complexity.
Data Source
AI summary
This disclosure is related to an image generation method and apparatus. The method includes: obtaining a first body image including a target body and a first clothes image including target clothes; transforming the first clothes image based on a posture of the target body in the first body image to obtain a second clothes image, the second clothes image including the target clothes, and a posture of the target clothes matching the posture of the target body; performing feature extraction on the second clothes image, an image of a bare area in the first body image, and the first body image to obtain a clothes feature, a skin feature, and a body feature respectively; and generating a second body image based on the clothes feature, the skin feature, and the body feature, the target body in the second body image wearing the target clothes.