Image Anomaly Training Using Similarity-Based Reference Maps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image anomaly detection methods for infrastructure inspection, such as road surfaces, face challenges in accuracy due to individual differences in data labeling and the influence of background variations and image capturing angles, particularly in weakly supervised learning scenarios.
Innovation Solution
A training apparatus and method that utilizes a first machine learning model for feature extraction and a second model for likelihood mapping, selecting reference images based on similarity to the subject image, and calculating differential feature maps to improve anomaly detection accuracy by minimizing loss values.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If weakly supervised learning is used to train a machine learning model with only presence/absence teaching labels, then data labeling time is reduced, but anomaly position estimation accuracy becomes low
Solution Approach 1:
The patent introduces a reference image as an intermediary element between the subject image and the teaching label. The reference image, selected based on similarity to the subject image, serves as a mediator that provides contextual information about normal states. This intermediary enables the model to infer anomaly positions more accurately without requiring detailed labeling, thus resolving the contradiction between reduced labeling time and improved estimation accuracy.
2Ease of manufacture
If a randomly sampled normal image is used for training, then training data preparation is simplified, but detection performance lowers due to background variations and image capturing angle differences
Solution Approach 1:
The patent changes the selection parameter from random sampling to similarity-based sampling. Instead of randomly selecting normal images, the system selects reference images based on their similarity to the subject image in terms of background and capturing conditions. This parameter change maintains the simplicity of automated selection while significantly improving detection reliability by ensuring that reference images are contextually appropriate.
3Measurement precision
If supervised learning with circumscribed rectangle teaching labels is used, then anomaly position detection is improved, but data labeling work takes time and shows individual differences
Solution Approach 1:
The patent uses copying by selecting reference images that replicate the normal state conditions of the subject image. Instead of creating detailed circumscribed rectangle labels for every training image, the system copies the contextual information from reference images that match the subject's background and capturing conditions. This approach maintains detection accuracy while dramatically reducing the time-consuming labeling process.
Data Source
AI summary
According to one embodiment, a training apparatus includes processing circuitry. The processing circuitry calculates a similarity between a subject image and at least one normal image. The processing circuitry selects at least one reference image from the normal image based on the similarity. The processing circuitry calculates first feature maps of the subject image and second feature maps of the reference image using a first machine learning model. The processing circuitry calculates differential feature maps that are differences between the first and second feature maps. The processing circuitry calculates a likelihood map based on the first feature maps and the differential feature maps using a second machine learning model. The processing circuitry calculates, based on the likelihood map and a teaching label of the subject image, a loss based on a likelihood. The processing circuitry updates the first and second machine learning models based on the loss.