GAN Image Sample Generation With Local Feature Discrimination
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge of obtaining sufficient training data for image recognition models, particularly when the specific feature occurs at a low probability, is addressed by generating image samples with a specific feature using a generative adversarial network (GAN) that incorporates location information as priori knowledge, reducing computation and improving discrimination accuracy and generation fidelity.
Innovation Solution
A method and system utilizing a SPADE-based generator structure and additional local discriminator in a generative adversarial network to generate images with specific features at specified locations, employing semantic segmentation images to guide the generation process and reduce computational burden.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a conventional generative adversarial network is used to generate image samples with specific features, then the generation process can be performed, but the computation amount is large and the discrimination accuracy and generation fidelity are insufficient
Solution Approach 1:
The discriminator is divided into a global discriminator that processes the entire image and a local discriminator that focuses on the specific feature region. This segmentation allows the local discriminator to specialize in evaluating the specific feature with higher accuracy while the global discriminator maintains overall image coherence, thereby improving discrimination accuracy without requiring a uniform increase in computational resources across the entire system.
Solution Approach 2:
The patent applies local quality by introducing a local discriminator specifically dedicated to evaluating the specific feature region, rather than using a single discriminator for the entire image. This allows the system to allocate discriminative capacity preferentially to the region of interest, improving generation fidelity of the specific feature while keeping the overall computational burden manageable through focused processing.
2Quantity of substance
If a large number of real image samples with specific features are obtained to improve training data sufficiency, then the training accuracy can be improved, but it is difficult to acquire sufficient real images when the natural occurrence probability of the specific feature is low
Solution Approach 1:
The patent employs preliminary action by using a pre-trained generative model to synthesize additional training samples before the actual training process. The model is first trained on available real samples, then used to generate synthetic samples that augment the training dataset. This preliminary generation of samples解决了 the problem of insufficient training data when specific features have low natural occurrence probability.
Solution Approach 2:
The patent uses copying by generating synthetic image samples that replicate the characteristics of real images containing specific features. The generative adversarial network creates copies of the specific features in various contexts and positions, producing realistic synthetic training data that expands the available dataset without requiring proportional increases in real sample collection.
3Productivity
If the generator processes the entire image to generate specific features, then the generation can be performed, but the computation process is inefficient when only a small defect region needs to be generated
Solution Approach 1:
The patent applies taking out by extracting and isolating the specific feature generation process from the overall image processing. The generator focuses computational resources only on generating the specific feature in its designated region, while the rest of the image remains unchanged. This extraction of the generation task to only the necessary region significantly improves generation efficiency and reduces computation time.
Solution Approach 2:
The patent implements partial action by having the generator process only the specific feature region rather than the entire image. The local discriminator similarly focuses evaluation only on the relevant feature area. This partial processing approach avoids the excessive computation that would result from processing complete images, thereby improving productivity while maintaining generation quality.
Data Source
Figure 1
Figure 2~3
Figure 4~5
AI summary
The present application provides a method and a system for generating an image sample having a specific feature. The method includes: training a generative adversarial network-based sample generation model, where the generative adversarial network includes a generator and two discriminators: a global discriminator configured to perform global discrimination on an image, and a local discriminator configured to perform local discrimination on a specific feature; and inputting, to a trained generator that serves as a sample generation model, a semantic segmentation image that indicates a location of the specific feature and a corresponding real image not having the specific feature, to obtain a generated image sample having the specific feature.