Zero-Shot Image Anomaly Detection With Two-Stage Reconstruction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image anomaly detection methods require a large number of balanced positive and negative samples, and autoencoders often fail to accurately reconstruct anomalous areas due to low-resolution or high-resolution issues, leading to misclassification or loss of pixel information.
Innovation Solution
A two-stage anomaly detection method using an autoencoder structure based on mutual information to reconstruct low-resolution and high-resolution images without anomalies, combined with an expert network for accurate anomaly measurement.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a large number of sample-balanced positive and negative samples are used to train the deep learning model, then the detection accuracy is improved, but the data collection cost and training time increase significantly
Solution Approach 1:
The patent performs preliminary action by pre-training the autoencoder exclusively on normal samples before the actual anomaly detection task. This preliminary training phase establishes a baseline model that learns normal patterns, eliminating the need to collect and train on rare anomaly samples during the main detection process, thus saving significant time while maintaining detection accuracy.
Solution Approach 2:
The patent segments the training process into two distinct phases: (1) autoencoder training on normal samples only, and (2) anomaly detection using the pre-trained model. This segmentation allows the model to focus on learning normal patterns first, then detect anomalies through reconstruction error, avoiding the time-consuming process of balancing anomaly samples.
2Productivity
If low-resolution reconstruction is used in autoencoder, then the training speed is improved, but the anomaly localization accuracy deteriorates due to extra image difference
Solution Approach 1:
The patent segments the reconstruction process into two resolution stages: low-resolution reconstruction for efficient processing and anomaly detection, followed by high-resolution reconstruction for precise anomaly localization. This segmentation allows the system to benefit from both fast low-resolution processing and accurate high-resolution detail recovery.
Solution Approach 2:
The patent transitions between resolution dimensions by first processing images at low resolution for speed, then upsampling to high resolution for detailed anomaly localization. This dimensional transition enables the system to optimize both training speed and localization accuracy by operating at different resolution levels at different stages.
3Measurement precision
If high-resolution reconstruction is used in autoencoder, then the anomaly localization precision is improved, but the anomalous area may be reconstructed together causing loss of pixel information
Solution Approach 1:
The patent performs preliminary low-resolution reconstruction to identify potential anomaly regions before performing high-resolution reconstruction. This preliminary action allows the system to focus computational resources on reconstructing only the normal areas at high resolution, while preserving the original high-resolution anomaly pixel information by not over-reconstructing those regions.
Solution Approach 2:
The patent applies different reconstruction qualities to different regions: high-resolution reconstruction is applied to normal areas to maintain detail, while anomaly areas are preserved from aggressive reconstruction to maintain their distinctive pixel characteristics. This local differentiation prevents anomaly information loss while maintaining overall localization precision.
4Device complexity
If single-stage autoencoder is used, then the device complexity is reduced, but the anomaly detection reliability deteriorates due to inability to handle both resolution requirements
Solution Approach 1:
The patent segments the anomaly detection system into two functional modules: an autoencoder for normal pattern learning and reconstruction, and an expert network for anomaly scoring and localization. This segmentation allows each module to specialize in its function, improving overall reliability while keeping individual module complexity manageable.
Solution Approach 2:
The patent introduces an expert network as an intermediary between the autoencoder reconstruction output and the final anomaly detection result. This intermediary processes the reconstruction error and original image to produce reliable anomaly scores, bridging the gap between simple reconstruction and complex anomaly interpretation, thereby enhancing overall system reliability.
Data Source
AI summary
The embodiments of this disclosure disclose an image anomaly detection method. A specific implementation of the method comprises: obtaining a test image; inputting the test image to an autoencoder to obtain a first reconstructed image; inputting the first reconstructed image to an expert network to obtain a second reconstructed image; based on the test image, the first reconstructed image, the second reconstructed image and a perceptual measurement method, generating an anomaly score matrix; based on the anomaly score matrix, generating anomalous area information. This implementation achieves a zero-shot training network and improves the accuracy of locating anomalous areas.