Image Signature Matching via Variable Extraction Regions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Image signatures with feature vectors having high correlation between dimensions result in low discrimination capability, especially for images with repeated patterns or textures, and fail to detect frequency components with cycles matching the rectangle's dimensions, leading to reduced robustness and accuracy in image matching.
Innovation Solution
The use of extraction regions with different shapes and relative positions for each dimension in the feature vector, allowing for varied patterns and reducing correlation between dimensions, thereby enhancing discrimination capability and robustness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If the same shape extraction regions are used for each dimension, then the device complexity is reduced, but the discrimination capability of the image signature is lowered
Solution Approach 1:
The patent applies local quality by assigning different shapes to extraction regions based on their specific dimension and location. Each dimension's extraction region is optimized for its particular function, with shapes selected from multiple candidates (e.g., vertical rectangles for vertical edge detection, horizontal rectangles for horizontal edge detection). This localized optimization improves discrimination capability without requiring complete system redesign.
Solution Approach 2:
The patent changes the shape parameter of extraction regions across different dimensions. By varying aspects ratios, orientations, and sizes of rectangular regions for different dimensions, the system captures diverse image features more effectively. This parameter variation reduces redundancy and enhances the discrimination capability of the image signature.
2Measurement precision
If extraction regions are optimized for specific frequency components, then the discrimination capability is improved, but the robustness against frequency variations is reduced
Solution Approach 1:
The patent segments the frequency detection task across multiple dimensions, each with extraction regions optimized for specific frequency components. Rather than using a single extraction region type for all frequencies, the system divides the work into specialized dimensions that collectively cover a broader frequency range, improving both discrimination capability and robustness.
Solution Approach 2:
The patent creates a composite approach by combining multiple extraction region shapes and orientations in different dimensions. This composite structure allows the image signature to capture diverse frequency components through the combination of specialized regions, achieving robustness against frequency variations while maintaining high discrimination capability.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
An image signature to be used for matching is generated by the following generation method, First, region features are extracted from respective sub-regions of a plurality of pairs of sub-regions in an image, and for each of the pairs of sub-regions, a difference value between the region features of two sub-regions forming a pair is quantized. When performing the quantization, the difference value is quantized to a particular quantization value if an absolute value of the difference value is smaller than a predetermined value. Then, a collection of elements which are quantization values calculated for the respective pairs of sub-regions is used as an image signature to be used for discriminating the image. An image signature matching device matches an image signature of a first image and an image signature of a second image, generated by the above-described generation method, in such a manner that a weight of an element having the particular quantization value is reduced.