Image Restoration Using Directional Regularization Strength
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image restoration methods fail to maintain the direction of contours in images with significant pixel value variations, as they do not consider the direction of pixel value variation during the restoration process.
Innovation Solution
A degradation restoration system that computes direction and amount of variation for each pixel, adjusting regularization strength based on these values to ensure the restored image maintains clear contours by using a variation computation unit, regularization strength computation unit, and image restoration unit.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If regularization is performed to reduce pixel value variation between neighboring pixels, then a unique solution can be determined, but the direction of contours is not maintained
Solution Approach 1:
The patent applies different regularization strengths to different spatial locations based on local image characteristics. Specifically, regions with large pixel value variations (edge regions) receive different treatment compared to regions with small variations (flat regions), allowing contour preservation while maintaining restoration accuracy
Solution Approach 2:
The patent dynamically adjusts the regularization strength parameter based on local image properties. By computing pixel value variations and using them to modulate the regularization term, the system adapts the restoration process to preserve contours while removing noise and artifacts
2Measurement precision
If adaptive regularization strength is assigned to each pixel based on pixel value difference, then sharpness is maintained in high-variation regions, but contour direction is still not preserved
Solution Approach 1:
The patent makes the regularization term spatially varying by incorporating local pixel value variation information. This allows the restoration process to maintain different qualities in different regions: sharp edges in high-variation regions while preserving smoothness in flat regions
Solution Approach 2:
The patent extends the regularization approach by considering not just the magnitude of pixel variations but also their spatial distribution patterns. This additional dimensional information helps preserve contour directions by identifying directional structures in the image
3Ease of manufacture
If uniform regularization is applied across the image, then computation is simplified, but regions with different characteristics cannot be restored optimally
Solution Approach 1:
The patent implements adaptive regularization where the regularization strength varies spatially according to local image characteristics. This is achieved by computing pixel value variations at each location and using them to determine local regularization parameters, optimizing restoration quality for different regions
Solution Approach 2:
The patent makes the regularization process dynamic by allowing the regularization strength to adapt based on local image content. The system dynamically adjusts regularization parameters during the restoration process based on computed pixel variations, rather than using fixed uniform parameters
Data Source
AI summary
For each pixel of an input image, a direction toward which variation in pixel value between the relevant pixel and its peripheral pixels is largest and an amount of the variation in pixel value are computed, where a direction unit vector that indicates the direction toward which the variation is largest and a variation vector that represents the amount of the largest variation in the pixel value are computed. For each pixel, a regularization strength is computed utilizing the direction unit vector, the variation vector, and a regularization strength computing formula so that the larger the amount of the variation in pixel value toward a direction indicated by the direction unit vector, the less the regularization strength. An optimization function is determined based on the input image and the regularization strength for each pixel thereof and assigns a value for each pixel, by which the determined optimization function has the minimum value, to the corresponding pixel of a restored image to be generated.