Super-Resolution Image Processing Apparatus Distinguishing Blur Types
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing super-resolution image generation techniques, such as Real-ESRGAN, fail to distinguish between blur caused by degradation processing and blur due to depth of field, leading to the sharpening of unwanted background areas and the generation of images different from the original.
Innovation Solution
An image processing apparatus that includes a learning unit to differentiate between blur caused by degradation processing and inherent blur in the image, using a discriminator to determine the degree of blur and a generator to perform super-resolution while preserving the original image's background blur, utilizing a machine learning model configured with a convolutional neural network to process and infer image data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If super-resolution is performed without distinguishing blur causes, then resolution is increased, but unwanted background areas are incorrectly sharpened
Solution Approach 1:
The patent segments the blur detection task by training the discriminator to specifically identify degradation-induced blur while ignoring depth-of-field blur. This is achieved by using training images where only certain regions contain degradation blur, enabling the model to segment different blur types and apply super-resolution selectively to the appropriate regions.
Solution Approach 2:
The patent applies local quality by making the blur detection and super-resolution application region-specific. The discriminator outputs blur likelihood values for different regions of the image, and super-resolution is applied only to regions where degradation blur is detected, preserving the original appearance of regions where depth-of-field blur is present.
2Manufacturing precision
If all blur is sharpened to increase resolution, then image detail is improved, but the original image characteristics are lost
Solution Approach 1:
The patent applies partial action by performing super-resolution only on portions of the image where degradation blur is detected, rather than applying it uniformly to the entire image. The discriminator identifies specific regions requiring enhancement, and super-resolution is applied selectively to those regions, avoiding unnecessary processing of already-sharp or intentionally blurred areas.
Solution Approach 2:
Different regions of the image receive different treatments based on their blur characteristics. Regions with degradation blur undergo super-resolution to restore detail, while regions with depth-of-field blur maintain their original appearance, preserving the artistic or functional intent of the original image.
3Device complexity
If blur detection is performed without considering depth of field, then processing simplicity is maintained, but incorrect regions are sharpened
Solution Approach 1:
The patent performs preliminary action by pre-training the discriminator on training images that contain degradation blur in specific regions. This pre-training enables the model to learn the characteristics of degradation-induced blur and distinguish it from depth-of-field blur, improving detection accuracy without adding complex post-processing steps.
Solution Approach 2:
The discriminator provides feedback in the form of blur likelihood values that indicate the probability of degradation blur presence in different regions. This feedback mechanism enables the system to make informed decisions about where to apply super-resolution, improving detection accuracy while maintaining processing efficiency through a unified neural network architecture.
Data Source
AI summary
To improve the quality of super-resolution performed on an image including a blurred portion, a method generates processed image data obtained by degradation processing performed on training image data based on a predetermined degradation processing parameter, causes a first machine learning model that discriminates a tag value in accordance with input image data to perform learning based on the processed image data and a tag value in accordance with the degradation processing parameter, causes a second machine learning model that generates output image data in accordance with input image data to perform learning based on training image data, the processed image data, and the tag value in accordance with the degradation processing parameter, performs inference using the first machine learning model on target image data that is a target of image processing as input, to output a tag value based on the target image data, and performs inference using the second machine learning model on this output tag value and the target image data to generate output image data based on the target image data.