Perceptual Image Similarity via Set Mapping Induced Distance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for determining perceptual similarity between images lack effectiveness in measuring similarity at a pixel level, disconnecting the similarity metric from human perception and being computationally expensive, especially in image retrieval and management systems.
Innovation Solution
The set mapping induced similarity distance (SMID) method expands each image into a set of related images and defines similarity as the smallest average distortion per pixel between these sets, using a pixel-level distortion measure, linking it to the weighted bipartite matching problem and developing two-stage algorithms for approximation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If conventional signature-based similarity metrics are used, then computation is simplified, but the similarity measure becomes disconnected from pixel-level distortion and human perception
Solution Approach 1:
The patent segments the image into multiple blocks and performs pixel-level distortion calculation at the block level rather than requiring full-image pixel-by-pixel comparison. This segmentation approach maintains perceptual accuracy by preserving local structure information while reducing computational complexity through block-based processing.
Solution Approach 2:
The patent uses partial action by computing similarity based on representative pixel groups within blocks rather than all pixels in the image. This partial sampling approach achieves sufficient perceptual accuracy without the excessive computational cost of complete pixel-level comparison across entire images.
2Measurement precision
If pixel-level distortion measures are used, then perceptual similarity accuracy is improved, but computational complexity increases significantly
Solution Approach 1:
The patent divides images into blocks and selects representative pixel groups within each block, transforming the computational problem from O(N×M) pixel comparisons to O((N/B)×(M/B)) block comparisons where B is the block size. This segmentation maintains perceptual accuracy while dramatically reducing computational complexity.
Solution Approach 2:
The patent creates simplified representations (copies) of image blocks through representative pixel groups that capture essential perceptual information. These copies serve as proxies for full block comparisons, reducing computational load while preserving the ability to measure perceptual similarity accurately.
3Productivity
If traditional similarity metrics (histogram-based, SIFT) are used, then computational cost is reduced, but discriminating power and alignment with human perception deteriorates
Solution Approach 1:
The patent segments images into blocks and uses representative pixel groups to maintain local structural information, providing better discriminating power than global histogram-based methods. This segmentation enables the system to distinguish between images with similar global statistics but different local structures, improving reliability while maintaining computational efficiency.
Solution Approach 2:
The patent applies local quality by focusing computational resources on representative pixel groups within blocks that capture essential perceptual features. This local-focused approach provides better discriminating power than global metrics by preserving local structural relationships while reducing overall computational cost through selective processing.
Data Source
AI summary
Systems, methods and computer program products for determining perceptual similarity between independent images are described. The described systems, methods and computer program products use a similarity distance to determine the perceptual similarity between a first image having a first arrangement of pixels and a second image having a second arrangement of pixels. A processor can determine a first image array corresponding to the first arrangement of pixels and a second image array corresponding of the second arrangement of pixels. The processor can generate a first pixel group set based on the first image array and a second pixel group set based on the second image array. A plurality of pixel group pairs can be generated, and each pixel group pair can be associated with a corresponding pixel group pair distortion. The similarity distance between the first image and the second image can be determined based on the plurality of pixel group pair distortions.