GAN Training Data Generation for Visual Search Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models for visual search face challenges in training due to the scarcity and labeling complexity of 'in the wild' images, as most such images are unlabeled and labeling them is time-consuming.
Innovation Solution
A system that generates training data by clustering labeled images from catalogs, retrieving additional 'in the wild' images using category-based searches, and using generative adversarial networks (GANs) to generate images resembling 'in the wild' scenarios, followed by object transference to create feature model training data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If more 'in the wild' images are used for training, then model effectiveness is improved, but data labeling time and cost increase
Solution Approach 1:
The patent uses GANs to generate synthetic training images that copy the visual characteristics of real 'in the wild' images without requiring actual physical images or their labels. The generator creates realistic images while the discriminator ensures they match the desired distribution, effectively copying the essential features needed for training.
Solution Approach 2:
The system uses the model's own predictions and the catalog structure to automatically generate training data. The GAN training process is self-supervised, using the discriminator's feedback to improve the generator without external human intervention for labeling.
2Manufacturing precision
If manual labeling of 'in the wild' images is performed, then training data quality is improved, but productivity decreases
Solution Approach 1:
The patent replaces the manual mechanical process of human labeling with an automated computational system. The GAN framework automatically generates labeled training data by synthesizing images and using the catalog structure to provide ground truth labels, eliminating the need for manual human annotation while maintaining data quality.
Solution Approach 2:
The GAN acts as an intermediary between the catalog data and the training requirements. Instead of directly using real images that need labeling, or creating simple synthetic images that lack realism, the GAN generates intermediate synthetic images that have both the realism needed for quality training and the automatic labelability provided by the catalog structure.
3Reliability
If GANs are trained on clustered catalog images, then generated image realism is improved, but device complexity increases
Solution Approach 1:
The patent segments the catalog images into distinct clusters based on visual features and categories before training the GAN. This segmentation allows the GAN to learn from organized, homogeneous groups of images, improving the realism and consistency of generated images while making the training process more manageable despite the overall system complexity.
Data Source
AI summary
Systems, device and techniques are disclosed for training data generation for visual search model training. A catalog including catalog entries which may include images of an item and data about the item may be received. Labels may be applied to the images of the items based on the data about the items. The images of the items may be sorted into clusters using cluster analysis on the labels. Each cluster may include labels as categories of the cluster. Additional images may be received based on searching for the categories. Generative adversarial network (GAN) training data sets may be generated from the images of the items, the additional images, and the categories. GANs may be trained with the GAN training data sets. The GANs may generate images including images of generated items, which may be replaced with images of items from the catalog entries to create feature model training images.