Weighted Feature Point Matching for Image Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Image recognition systems face challenges in accurately identifying objects when dealing with similar model images, as existing methods struggle with processing speed and memory usage due to high similarity scores and increased recognition depth, especially in mixed flow production lines where various types of work pieces with similar features are present.
Innovation Solution
The system assigns weights to feature points based on their saliency, calculating a weighted similarity score that increases with fewer matching model images, and combines this with a simple similarity score to accurately identify objects, reducing processing load and improving recognition accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If template matching is used to recognize images of similar model images, then recognition can be performed using simple similarity calculation, but recognition accuracy deteriorates because multiple model images receive perfect scores when they are subsets of each other or have large intersections
Solution Approach 1:
The patent applies local quality by assigning different weights to different feature points based on their discrimination power. Feature points that appear in fewer model images (higher discrimination power) are given higher weights, while feature points appearing in many model images (lower discrimination power) are given lower weights. This resolves the contradiction by maintaining the simplicity of template matching while improving accuracy through localized weighting of feature points.
2Measurement precision
If correlation computation is performed for all combinations of similar model images to achieve accurate recognition, then recognition accuracy improves, but processing load increases significantly
Solution Approach 1:
The patent changes the parameter of similarity calculation from unweighted feature point matching to weighted feature point matching. By introducing weights that reflect the discrimination power of each feature point, the system achieves accurate recognition without needing to compute correlations for all model image combinations. This resolves the contradiction by improving accuracy through parameter optimization rather than exhaustive computation.
3Productivity
If all feature points are used to calculate degree of similarity, then processing load is low, but recognition accuracy deteriorates because feature points with low discrimination power reduce the distinction between similar models
Solution Approach 1:
The patent applies local quality by assigning different weights to different feature points based on their discrimination power. Feature points that appear in fewer model images (higher discrimination power) are given higher weights, while feature points appearing in many model images (lower discrimination power) are given lower weights. This resolves the contradiction by maintaining the simplicity of template matching while improving accuracy through localized weighting of feature points.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
An image recognition device includes an identification unit configured to compare a feature point in an input image and a feature point in every model image to compute a first degree of similarity between the input image and the model image, and to identify the input image on the basis of said first degree of similarity. To compute the first degree of similarity for a model image, the identification unit adds to a score based on a matching feature point for each feature point in the model image matching a feature point in the input image. The score based on the matching feature point is a value that increases as the number of model images including the matching feature point decreases.