Learning Device for Face Identification Using Integral Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing face identification methods are weak against environmental changes, such as variations in illumination, due to their reliance on specific feature extraction filters and calculation processes.
Innovation Solution
A learning device and method that selects and combines multiple feature extraction filters to extract features from images, calculates correlations between these features, and learns a same-subject classifier using boosting techniques to improve identification accuracy and speed.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple Gabor filters are used for feature extraction at each feature point, then identification precision is improved, but processing time increases significantly
Solution Approach 1:
The patent pre-calculates and stores integral images before feature extraction. These integral images contain pre-computed cumulative sums that enable rapid calculation of rectangular region features during identification, eliminating the need for repeated pixel-by-pixel calculations when applying multiple Gabor filters.
Solution Approach 2:
The patent divides the feature extraction process into separate stages: first computing integral images independently, then using these pre-computed structures to efficiently extract features with multiple Gabor filters. This segmentation allows the computationally intensive integral image calculation to be performed once, rather than repeatedly during filter operations.
2Productivity
If a single feature extraction filter is used, then processing speed is improved, but robustness against environmental changes deteriorates
Solution Approach 1:
The patent combines multiple weak classifiers, each using different Gabor filters with varying characteristics (sizes, orientations, frequencies), into a single strong classifier through boosting. This ensemble approach merges the results of multiple filters to achieve robustness against environmental changes while maintaining efficient processing through the integrated classifier structure.
Solution Approach 2:
The patent creates a composite classifier system that integrates multiple Gabor filter responses into a unified feature representation. By combining features from filters with different characteristics into a composite feature vector, the system achieves environmental robustness similar to composite materials that combine different properties for enhanced performance.
3Ease of operation
If feature difference calculation is performed between correspondence points, then processing simplicity is improved, but adaptability to environmental changes deteriorates
Solution Approach 1:
The patent implements a dynamic feature extraction process where multiple Gabor filters with different characteristics are adaptively applied to feature points. The system dynamically selects and combines filter responses based on local image characteristics, allowing the feature extraction to adapt to varying environmental conditions such as illumination changes, rather than using a fixed simple difference calculation.
Data Source
AI summary
Provided is a learning device including: an acquisition section that acquires a plurality of image pairs in which the same subjects appear and a plurality of image pairs in which different subjects appear; a setting section that sets feature points on one image and the other image of each image pair; a selection section that selects a plurality of prescribed feature points, which are set at the same positions of the one image and the other image, so as to thereby select a feature extraction filter for each prescribed feature point; an extraction section that extracts the features of the prescribed feature points of each of the one image and the other image by using the plurality of feature extraction filters; a calculation section that calculates a correlation between the features; and a learning section that learns a same-subject classifier on the basis of the correlation and label information.