Image Generator Segmentation for False Structure Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image resolution enhancement methods using machine learning models, such as SRGAN and network interpolation, often result in image quality degradation, including multiple edges and color changes, due to the presence of false structures and uneven perceived resolution.
Innovation Solution
A method involving a generator that produces two intermediate high-resolution images using different loss functions, one without a discriminator and one with, which are then combined to generate an estimated high-resolution image, thereby controlling the appearance of false structures and perceived resolution, while preventing image quality degradation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a generator learns resolution enhancement using a discriminator (SRGAN), then perceived resolution and texture quality are improved, but false structures appear causing subjective strangeness
Solution Approach 1:
The generator is divided into two separate models: a first generator trained without a discriminator to minimize false structures, and a second generator trained with a discriminator to maximize perceived resolution. Each generator specializes in one aspect, and their outputs are combined to achieve both goals simultaneously.
Solution Approach 2:
The invention changes the training parameters and objectives of the generator by creating two distinct training regimes. The first generator uses a loss function focused on structural accuracy without adversarial pressure, while the second generator uses adversarial training to enhance texture and perceived resolution. The final output blends these two different parameter-optimized results.
2Object-generated harmful factors
If network interpolation is used to balance false structure and perceived resolution, then a balance is achieved, but image quality degradation occurs such as multiple edges and color change
Solution Approach 1:
The invention extracts and separates the conflicting objectives into two independent generator models. Instead of trying to balance false structure and perceived resolution in a single generator through interpolation, the solution extracts these as separate specialized generators, then combines their outputs to avoid the quality degradation that occurs with traditional network interpolation.
3Device complexity
If a single generator is used for resolution enhancement, then device complexity is reduced, but it cannot simultaneously minimize false structures and maximize perceived resolution
Solution Approach 1:
The single generator is segmented into two specialized generators with different training objectives. The first generator focuses on structural accuracy without adversarial training, while the second focuses on texture quality with adversarial training. This segmentation allows each generator to excel at its specific task, improving overall reliability.
Solution Approach 2:
After the two generators are trained separately with different objectives, their outputs are merged through weighted combination. This merging allows the system to benefit from both generators' strengths while maintaining manageable complexity through modular architecture and automated weight adjustment.
Data Source
AI summary
A method for processing an image uses a generator which is a machine learning model. The generator converts an input low resolution image into a first feature map. From the first feature map, a first intermediate image and a second intermediate image each having resolution higher than resolution of the low resolution image are generated. Based on the first intermediate image and the second intermediate image, an estimated image having higher resolution than the resolution of the low resolution image is generated.