Multi-Valued Teaching Signal for Object Detection Overlap
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image recognition methods require effective setting of teaching signals for machine learning, which is challenging in selecting and detecting objects within large images, particularly in determining the relevance between detection and target regions.
Innovation Solution
An image processing device and method that acquires images for learning, sets a target region including a detection target, and sets a detection region with a teaching signal that can take three or more values based on the relevance between the detection and target regions, utilizing units for target region setting, detection region setting, and teaching signal calculation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a teaching signal with only two values (presence/absence) is used, then the learning process is simple, but the detection accuracy decreases when the detection region only partially overlaps with the target region
Solution Approach 1:
The teaching signal is changed from a binary parameter (0 or 1) to a multi-valued parameter (0, 0.5, 1) that reflects the degree of overlap between detection region and target region. This allows the learning model to distinguish between complete overlap, partial overlap, and no overlap, thereby improving detection accuracy while maintaining reasonable signal complexity
Solution Approach 2:
Instead of requiring complete overlap between detection region and target region for a positive teaching signal, the patent applies partial action by assigning intermediate values (0.5) to partially overlapping cases. This enables the model to learn from partial matches and improve overall detection performance
2Productivity
If the detection region is set to fully include the target region, then the teaching signal setting is simple, but detection regions that only partially include the target are lost
Solution Approach 1:
The teaching signal is made local and specific to each detection region by calculating individual overlap ratios between the detection region and target region. Each region receives a customized teaching signal value (0, 0.5, or 1) based on its specific overlap characteristics, enabling precise learning for different spatial configurations
Solution Approach 2:
The overlap ratio is calculated in advance during the learning data preparation phase, and the appropriate teaching signal value is assigned before the actual detection task. This preliminary classification of overlap degrees enables the model to learn from diverse spatial configurations without requiring complex real-time calculations
Data Source
AI summary
Provided is an image processing device having an acquisition unit that acquires an image for learning used for machine learning; a target region setting unit that sets, to the image, a target region including a detection target; a detection region setting unit that sets, to the image, a detection region in which a teaching signal is required to be set; and a teaching signal setting unit that sets, to the detection region, a teaching signal that may take three or more values in accordance with a relevance between the detection region and the target region.