Attention-Map Model Training for Defect Detection Without Position Labels
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models for object detection require large amounts of training data with position information, which is labor-intensive, and there is a need to enhance the accuracy of detecting defective areas by emphasizing them over normal areas.
Innovation Solution
A training method that computes feature and attention maps from normal and defective images, calculates losses based on these maps, and updates model parameters to minimize total loss, focusing on suppressing normal areas and enhancing defective area detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If training data with position information is used for supervised learning, then detection accuracy is improved, but the amount of work for generating training data increases
Solution Approach 1:
The patent extracts only the essential label information from images, excluding position information and other detailed annotations. This allows training the machine learning model with simplified data that requires minimal manual work while still achieving the goal of detecting defective areas through attention maps.
Solution Approach 2:
Instead of directly training the model to detect defective areas with position information (supervised learning), the patent inverts the approach by training the model to generate attention maps that highlight important regions (weak supervised learning). This indirect approach reduces the need for detailed position information in training data.
2Loss of time
If weak supervised learning is used to reduce training data work, then detection accuracy may be improved, but the ability to emphasize defective areas and suppress normal areas deteriorates
Solution Approach 1:
The patent applies local quality by computing separate loss functions for different regions: a first loss for normal areas (suppressing them in attention maps) and a second loss for defective areas (emphasizing them). This region-specific loss computation ensures that the model learns to differentiate between normal and defective regions even with weak supervised learning.
Solution Approach 2:
The patent uses counterweight by introducing a suppression loss for normal areas that counterbalances the detection loss. This suppression loss actively penalizes attention in normal regions, creating a counteracting force that ensures defective areas are emphasized while normal areas are suppressed, improving defective area detection accuracy.
Data Source
AI summary
According to one embodiment, a training method computes a first feature map and a first attention map from a first image, and computes a second feature map and a second attention map from a second image, by inputting the first image, which does not include a defective area of a target, and the second image, which includes the defective area of the target, to a machine learning model. The training method computes a first loss, based on the first attention map. The training method computes a class classification of the target, based on the second feature map and the second attention map. The training method computes a second loss, based on the class classification. The training method computes a total loss, based on the first loss and the second loss. The training method updates a parameter of the machine learning model so as to minimize the total loss.