Bag-Level Identifier Training for Subtle Defect Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning methods for identifying normal or abnormal instances in product images require costly instance-level teaching and struggle with overlooking flaws that are shallow or not clearly differentiated, especially when teaching is performed at a coarser bag level.
Innovation Solution
A training apparatus and method that uses a neural network to calculate bag likelihood and expected instance quantity, updating parameters based on teaching data at the bag level to improve detection accuracy by minimizing losses for overlooked or over-detected instances.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If instance-level teaching data is used to train the identifier, then detection accuracy is improved, but training cost and complexity increase significantly
Solution Approach 1:
The teaching data is segmented into two levels: bag-level (coarse) and instance-level (fine). The identifier processes instances within bags, and the training system segments the learning process into bag-level supervision and instance-level probability calculation, allowing efficient training without requiring expensive instance-level annotations for all data
Solution Approach 2:
The multiple instance learning framework acts as an intermediary between bag-level teaching data and instance-level detection. It introduces probability sequences as an intermediate representation that bridges the gap between coarse bag-level labels and fine-grained instance predictions, enabling accurate detection without direct instance-level teaching
2Ease of manufacture
If bag-level teaching data is used to train the identifier, then training cost is reduced, but detection accuracy deteriorates due to training termination after finding one abnormal instance
Solution Approach 1:
The system implements feedback through probability sequences that indicate the likelihood of each instance being abnormal. This feedback mechanism allows the training process to continue beyond finding the first abnormal instance, using the probability information to guide further training and improve detection of multiple or subtle abnormalities
Solution Approach 2:
The system performs excessive action by calculating probability sequences for all instances even when one abnormal instance is found. This ensures comprehensive detection of multiple abnormalities and prevents early termination, maintaining high detection accuracy while using efficient bag-level teaching data
3Productivity
If training terminates after finding one abnormal instance, then training efficiency is improved, but detection completeness deteriorates when abnormal portions are large
Solution Approach 1:
The system performs excessive action by continuing to evaluate all instances and generate probability sequences even after finding one abnormal instance. This ensures that large or multiple abnormal portions are detected, maintaining detection completeness while managing training efficiency through the bag-level teaching framework
4Measurement precision
If instance-level teaching maps are created, then detection precision is improved, but labor and cost increase significantly
Solution Approach 1:
The system extracts only the necessary information from teaching data at the bag level, avoiding the need to create expensive instance-level teaching maps. By using multiple instance learning, it extracts meaningful patterns from coarse bag-level labels and applies them to instance-level detection, achieving high precision with minimal labor and cost
Data Source
AI summary
A training apparatus includes a circuitry. The circuitry inputs a bag that is a set of an instance to an identifier and outputs a probability sequence including a probability that each instance belonging to the bag corresponds to a target. The circuitry calculates a bag likelihood that is a probability that specific instances corresponding to the target are included in the bag. The circuitry outputs an expected instance quantity regarding a quantity of the specific instances included in the bag. The circuitry calculates a parameter update amount based on teaching data, the bag likelihood, and the expected instance quantity, and updates a parameter of the identifier based on the parameter update amount.