Multi-Level GAN Architecture for Stock Image Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing technologies face challenges in generating stock-quality digital images from user-provided images, especially when the image depicts fashion items in different poses, with partial self-occlusions, or under challenging lighting conditions, making it difficult to perform effective visual searches on e-commerce platforms.
Innovation Solution
The use of multiple levels of generative adversarial networks (GANs) to train digital image generators that can produce stock-quality images by pre-processing user-provided images through person segmentation and human pose estimation, and then processing them through a combination of GANs to generate images that can be used for visual searches.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If a single GAN is used to generate stock-quality images from user-provided images, then the generation process is simpler and faster, but the quality and accuracy of generated images are insufficient for effective visual search
Solution Approach 1:
The patent divides the image generation task into multiple specialized GANs, each responsible for specific aspects: a first GAN for generating initial stock-quality images, a second GAN for improving image quality, and a third GAN for enhancing searchability. This segmentation allows each GAN to specialize in particular functions, achieving high generation quality while maintaining manageable system complexity through modular architecture.
Solution Approach 2:
The patent implements a nested GAN structure where the output of one GAN serves as input to the next GAN in the hierarchy. The first GAN generates initial images, which are then processed by the second GAN for quality enhancement, and finally by the third GAN for search optimization. This nested arrangement allows progressive refinement of image quality while organizing complexity in a hierarchical manner.
2Measurement precision
If user-provided images with varying poses, occlusions, and lighting conditions are used directly for search, then the search process is faster, but the search accuracy decreases due to image quality issues
Solution Approach 1:
The patent applies preliminary actions by preprocessing user-provided images before they enter the GAN generation pipeline. Image preprocessing techniques are applied to normalize poses, handle occlusions, and correct lighting conditions in advance. Additionally, the multi-stage GAN structure performs preliminary generation and quality enhancement in sequence, ensuring that final search images are optimized without requiring excessive processing time during actual search operations.
3Adaptability or versatility
If traditional image search methods are used without GAN-generated stock images, then the system is simpler to implement, but the ability to match items across different poses and conditions is limited
Solution Approach 1:
The patent creates multiple synthesized copies of user-provided images through the GAN system, generating numerous variant images that represent the same item under different poses, lighting conditions, and occlusion scenarios. These synthesized copies expand the search capability by providing multiple representations of each item, enabling robust matching across diverse conditions while maintaining a unified system architecture.
Data Source
AI summary
Various embodiments described herein utilize multiple levels of generative adversarial networks (GANs) to facilitate generation of digital images based on user-provided images. Some embodiments comprise a first generative adversarial network (GAN) and a second GAN coupled to the first GAN, where the first GAN includes an image generator and at least two discriminators, and the second GAN includes an image generator and at least one discriminator. According to some embodiments, the (first) image generator of the first GAN is trained by processing a user-provided image using the first GAN. For some embodiments, the user-provided image and the first generated image, generated by processing the user-provided image using the first GAN, are combined to produce a combined image. For some embodiments, the (second) image generator of the second GAN is trained by processing the combined image using the second GAN.