Image Anomaly Detection Using Pixel-Level Restored Comparisons
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current neural network models, such as the YOLO algorithm, are limited to detecting anomalies at a rectangular level and cannot effectively detect anomalies at a non-rectangular level, resulting in low detection capability.
Innovation Solution
An image anomaly detection method that involves restoring predicted distributions of multiple images, including the original image and scaled versions, and comparing these restored images with target distributions to identify abnormal pixels, enabling pixel-level anomaly detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If the YOLO algorithm is used for anomaly detection, then the detection process is simple and fast, but the detection capability is limited to rectangular level only
Solution Approach 1:
The patent segments the anomaly detection process into multiple stages: first using YOLO for rapid rectangular boundary box prediction, then applying flow model restoration to generate restored images from predicted distributions, and finally performing pixel-level comparison to identify abnormal pixels. This segmentation allows each stage to specialize - YOLO handles speed while the restoration and comparison stages handle precision.
Solution Approach 2:
The patent transitions from rectangular-level detection (2D bounding boxes) to pixel-level detection (individual pixel analysis). By restoring predicted distributions to generated images and comparing them pixel-by-pixel with original images, the system elevates the detection dimension from coarse rectangular boundaries to fine-grained pixel precision, enabling detection of non-rectangular anomaly shapes.
2Measurement precision
If pixel-level anomaly detection is implemented, then the detection capability is enhanced, but the computational complexity increases
Solution Approach 1:
The patent performs preliminary action by using the flow model to restore predicted distributions into generated images before comparison. This preprocessing step transforms abstract distribution predictions into visual images that can be directly compared with original images, simplifying the subsequent pixel-level analysis and reducing computational burden during the detection phase.
Solution Approach 2:
The patent introduces restored images as an intermediary between the YOLO predictions and the final anomaly detection. The flow model acts as a mediator that converts predicted distributions into restored images, which then serve as the basis for pixel-level comparison. This intermediary structure simplifies the complexity by providing a visual intermediate representation that bridges the gap between rectangular predictions and pixel-level analysis.
3Ease of manufacture
If only rectangular boundary boxes are used for anomaly detection, then the algorithm is simple to implement, but non-rectangular anomalies cannot be detected
Solution Approach 1:
The patent segments the anomaly representation into two parts: rectangular boundary boxes from YOLO for localization, and pixel-level restored image comparisons for shape characterization. This segmentation allows the system to maintain the simplicity of rectangular bounding for initial detection while adding pixel-level analysis to capture arbitrary anomaly shapes, achieving both simplicity and versatility.
Solution Approach 2:
The patent adds a new dimension to anomaly representation by transitioning from 2D rectangular boundaries to pixel-level intensity comparisons. The restored images provide an additional dimensional layer that captures the actual shape and extent of anomalies, enabling detection of non-rectangular shapes while building upon the existing rectangular boundary box framework.
Data Source
AI summary
An image anomaly detection method includes: restoring predicted distributions of N images to obtain N restored images, where the N images include a to-be-detected image and/or an image obtained by scaling the to-be-detected image; and obtaining an abnormal pixel of the to-be-detected image based on N comparison results of the N restored images and N images obtained by restoring target distributions. The abnormal pixel of the to-be-detected image can be obtained, and pixel-level detection is realized, to aid detection capability.