Image Feature Extraction Using Co-occurrence Matrices for Global Context
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing feature extraction methods for images struggle to incorporate both local and global information effectively, particularly in recognizing objects with few repeatedly-appearing patterns, and fail to restore lost information during the recognition process.
Innovation Solution
A feature extracting apparatus and method that calculates pixel features, sets areas in the image, maps coordinates between areas using affine transformations, and computes co-occurrence matrices to capture frequency combinations of pixel features, enabling the extraction of robust features that incorporate both local and global information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If features are extracted for each divided area using grid pattern, then local information is captured, but global information expressing relationships with distant areas is lost
Solution Approach 1:
The patent extends the feature extraction from local 2D area grids to a 3D spatial relationship dimension by calculating co-occurrence matrices that capture relationships between pixels at different spatial positions and directions. This adds a global contextual dimension to the previously local feature extraction process.
Solution Approach 2:
The patent merges local pixel features with global spatial relationship information by combining the co-occurrence matrix (capturing global relationships) with the original pixel feature vectors. This integration allows both local and global information to be utilized simultaneously in object recognition.
2Measurement precision
If co-occurrence matrix is calculated for texture images with repeatedly-appearing patterns, then texture recognition is improved, but recognition of objects with few repeated patterns (e.g., persons) is not effective
Solution Approach 1:
The patent changes the parameters of the co-occurrence matrix calculation by using multiple spatial distances and multiple directions (orientations). This allows the feature extraction to adapt to different object types - objects with repeated patterns benefit from the multi-directional analysis, while objects with few patterns benefit from the multi-scale spatial relationships captured by varying distances.
Solution Approach 2:
The patent creates a universal feature extraction method that works for both texture images with repeated patterns and objects with few repeated patterns. The co-occurrence matrix calculation with multiple distances and directions serves multiple functions: capturing texture regularities when present, and capturing spatial relationships when patterns are sparse or absent.
3Reliability
If feature extraction process is performed previously before recognition, then recognition process can be performed properly, but information lost in feature extraction cannot be restored in subsequent recognition process
Solution Approach 1:
The patent performs preliminary action by extracting comprehensive features including both local pixel properties and global spatial relationships before the recognition process. By calculating co-occurrence matrices that capture relationships at multiple distances and directions, the method prepares more complete feature information in advance, reducing information loss during subsequent recognition.
Data Source
AI summary
A feature extracting apparatus includes a pixel feature calculator that calculates a pixel feature for each pixel of image data; an area setting unit configured to set a plurality of areas in the image data; a coordinate mapping unit configured to map a first coordinate in one of the plurality of areas onto a second coordinate in at least one of the other plurality of areas; and a co-occurrence matrix calculator configured to calculate a co-occurrence matrix for each of the plurality of areas, the co-occurrence matrix being frequency of combinations of the pixel features at the first coordinate and the pixel feature at the second coordinates.