Object Recognition via Edge Gradient Similarity Tables
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional object recognition and location methods face challenges with broken or hidden objects, complex backgrounds, and non-linear light intensity variations, leading to reduced accuracy and increased computational complexity.
Innovation Solution
A method that analyzes a template image to obtain sampled edge points and gradient angles, generating similarity score tables, which are then used to calculate predictive gravity positions in a live image, correlating the position with the highest similarity score to effectively recognize and locate objects, even with broken parts or complex backgrounds and non-linear light variations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If connected component analysis is employed to recognize objects, then the method is simple to implement, but the recognition rate is largely reduced when the object has broken parts or hidden parts
Solution Approach 1:
The patent segments the object recognition task into multiple feature extraction components (edge points, gradient angles, area, perimeter) rather than relying on a single connected component analysis. This segmentation allows the system to handle broken or hidden parts by combining multiple feature types for more robust recognition.
2Device complexity
If SAD algorithm is used to locate the object, then binary operations are not needed, but the algorithm fails to overcome the problem caused by linear light intensity variation
Solution Approach 1:
The patent changes the parameters used for location from simple pixel intensity comparison (SAD) to gradient angle and edge point features. By transforming the feature space to use gradient information instead of raw intensity, the system becomes insensitive to linear light intensity variations while maintaining algorithmic efficiency.
3Reliability
If NCC method is used to locate the object, then the problem caused by linear light intensity variation is overcome, but the similarity score is largely reduced when the object is hidden or non-linear light intensity variation exists
Solution Approach 1:
The patent creates a composite feature representation by combining multiple types of features (edge points, gradient angles, area, perimeter) rather than relying on a single NCC method. This composite approach leverages the strengths of different feature types to maintain accuracy under various lighting conditions while reducing the weaknesses of any single method.
4Reliability
If GHT method is used to extract edge points, then the problems of non-linear light intensity variation and hidden object are overcome, but the detection result is unstable when the background is complicated and the number of non-edge parts is huge
Solution Approach 1:
The patent extracts only the most discriminative features (edge points with gradient angles) from the image while discarding redundant non-edge information. By selectively extracting only the essential features needed for recognition and location, the system avoids the instability caused by complicated backgrounds and non-edge parts.
5Speed
If chamfer transform is used to accelerate location, then the location speed is improved, but the locating efficacy is impaired when the object has hidden parts and many non-edge parts are generated
Solution Approach 1:
The patent applies different processing qualities to different parts of the image: edge points receive full gradient angle analysis while non-edge parts are discarded. This local quality differentiation maintains high locating efficacy for important features while enabling faster processing by skipping unnecessary analysis of non-edge regions.
Data Source
AI summary
A method for recognizing and locating an object includes an offline mode process and an online mode process. In the offline mode process, plural sampled edge points of a template image of the object and respective gradient angles and a gravity position of the plural sampled edge points are obtained, and plural similarity score tables are obtained according to the plural sampled edge points, a predetermined detecting distance range and a predetermined gradient angle difference range. In the online mode process, plural edge points of a live image and respective gradient angles are obtained, plural predictive gravity positions are calculated, and plural similarity scores corresponding to the plural predictive gravity positions are summed up. The predictive gravity position with the local maximum of the similarity scores higher than a threshold value is correlated with the gravity position of the template image so as to recognize and locate the object.