Image Anomaly Detection Using Encoder-Decoder Regression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing anomaly detection methods in images are limited by restricted receptive fields and prone to blurry reconstructions, leading to false positives and negatives, especially in applications like manufacturing where diverse defects are common.
Innovation Solution
A method using an encoder, decoder, and regression program to learn a representation of the whole input image, comparing reconstructed features to detect anomalies without relying on target representations, thereby avoiding false positives and negatives.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If autoencoder-based methods are used to reconstruct input images through a low-dimensional bottleneck, then anomaly detection is performed by comparing input image to reconstruction, but the reconstructions become blurry and inaccurate leading to false positives
Solution Approach 1:
The patent introduces a target representation as an intermediary between the input image and the reconstruction. Instead of directly reconstructing the input image, the system first transforms the input image into a target representation (which captures essential features while filtering out noise and irrelevant details), then reconstructs from this intermediate form. This intermediary step resolves the contradiction by providing a stable, feature-based representation that avoids the blurriness of direct low-dimensional bottleneck reconstructions while still enabling effective anomaly detection through comparison.
2Device complexity
If the receptive field is restricted in anomaly detection methods, then computational complexity is reduced, but detection accuracy decreases due to inability to capture global context
Solution Approach 1:
The patent resolves this contradiction by changing the dimensionality of the problem space. Instead of expanding the spatial receptive field in the original image domain (which would increase computational complexity), the system transforms the problem into a different dimensional space through target representation. This transformation allows the model to capture global context and long-range dependencies without increasing the spatial receptive field size, as the target representation inherently encodes global information in a compressed form that can be processed efficiently.
3Measurement precision
If decoder program produces inaccurate reconstructions, then false positives increase, but using regression program to compensate adds computational overhead
Solution Approach 1:
The patent implements a feedback mechanism where the regression program receives the target representation and generates a regression output that is compared with the decoder output. This feedback loop allows the system to identify and correct inaccuracies in the decoder's reconstruction by leveraging the regression program's predictions. The regression output serves as a reference that compensates for decoder errors, and this feedback-based correction mechanism reduces false positives while maintaining computational efficiency through the structured comparison and error correction process.
Data Source
AI summary
This invention relates generally to machine vision systems, and more particularly, to the detection of anomalies in scenes observed by imaging sensors.