Learning Apparatus for Abnormality Detection Using Image Similarity
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing techniques for generating an estimation model for detecting abnormality face challenges in efficiently collecting training images, particularly requiring a large number of images indicating abnormal behavior.
Innovation Solution
A learning apparatus and method that acquire images, compute similarity with pre-defined abnormal state images, register images with low similarity as normal state images, and generate a model for discriminating between normal and abnormal using machine learning with both types of images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a large number of training images indicating abnormal behavior are collected, then the accuracy of the estimation model for detecting abnormality is improved, but the difficulty and time required for collecting training images increases significantly
Solution Approach 1:
The patent applies preliminary action by pre-defining abnormal state images before the actual training process. These pre-defined abnormal images serve as reference points that enable the system to automatically identify and collect corresponding normal images through similarity comparison, eliminating the need to manually collect large numbers of abnormal behavior images.
Solution Approach 2:
The patent uses copying by creating similarity-based copies of the pre-defined abnormal state images. By comparing newly acquired images against these abnormal templates and identifying those with low similarity, the system automatically generates a dataset of normal images without requiring direct examples of abnormal behavior, thus solving the data collection bottleneck.
2Reliability
If traditional methods are used to collect training images for abnormality detection, then the estimation model can be trained, but the user load and complexity of the process increases
Solution Approach 1:
The patent implements self-service by enabling the system to automatically perform the entire training image collection process without extensive user intervention. The system autonomously compares acquired images against pre-defined abnormal templates, identifies normal images through similarity thresholds, and prepares training datasets, thereby reducing user load and operational complexity.
Solution Approach 2:
The patent applies parameter changes by utilizing similarity threshold parameters to automatically differentiate between normal and abnormal images. By adjusting and comparing similarity values against predefined thresholds, the system transforms the complex qualitative judgment of image normality into a quantitative parameter-based classification process, simplifying the overall workflow.
Data Source
AI summary
The present invention provides a learning apparatus (10) including: an acquisition unit (11) that acquires an image; a similarity computation unit (12) that computes a similarity between the acquired image, and a first image being accumulated in advance and indicating an abnormal state; a registration unit (13) that registers, as a second image indicating a normal state, the acquired image whose similarity is equal to or less than a first reference value; and a learning unit (14) that generates an estimation model for discriminating between normal and abnormal by machine learning using the first image and the second image.