Weakly Supervised Anomaly Detection via Attention Mapping
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current anomaly detection methods in industrial inspection, such as manual inspection and fully supervised machine vision, are inefficient due to low automation and the need for extensive manual labeling, making them unsuitable for real-time applications and difficult to scale.
Innovation Solution
A weakly supervised network using a Convolutional Neural Network (CNN) or Recurrent Neural Network that processes images to detect anomalies through gradient-based backpropagation, generating attention mappings to identify anomalous regions without prior knowledge of their shape or form, and using color models to segment anomalies from normal regions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If fully supervised machine vision approaches are used with pixel-wise labeling or bounding box annotation, then anomaly detection accuracy is improved, but manual labeling effort and time consumption increase significantly
Solution Approach 1:
The patent applies partial supervision by using only image-level anomaly labels instead of complete pixel-wise annotations. The weakly supervised approach performs anomaly detection with reduced labeling effort (partial action) while still achieving effective anomaly identification through attention mapping generation and gradient-based localization, sacrificing some precision for significant time savings.
Solution Approach 2:
The patent introduces attention mappings as an intermediary mechanism between the weakly supervised classifier and the final anomaly detection result. The attention maps serve as a mediator that translates coarse image-level labels into fine-grained anomaly localization without requiring direct pixel-wise annotations, enabling accurate detection with minimal manual labeling.
2Measurement precision
If manual inspection by specialists is used, then anomaly detection accuracy is improved, but automation level and productivity decrease
Solution Approach 1:
The system enables self-service anomaly detection by training the classifier to automatically generate attention mappings that highlight anomalous regions. The model serves itself by learning from weakly labeled data and autonomously performing anomaly localization without requiring specialist intervention, thereby maintaining accuracy while dramatically improving productivity and automation level.
Solution Approach 2:
The patent replaces the mechanical inspection process performed by human specialists with an automated neural network-based system. The CNN classifier with attention mapping generation substitutes the human visual inspection mechanism, achieving comparable accuracy while enabling high-speed automated processing and significantly increasing inspection throughput.
3Reliability
If fully supervised methods with extensive manual labeling are used, then detection reliability is improved, but scalability and ease of deployment worsen
Solution Approach 1:
The patent applies partial supervision by using only image-level anomaly labels instead of complete pixel-wise annotations. The weakly supervised approach performs anomaly detection with reduced labeling effort (partial action) while still achieving effective anomaly identification through attention mapping generation and gradient-based localization, sacrificing some precision for significant time savings.
Solution Approach 2:
The patent creates a universal weakly supervised framework that can be applied across different industrial inspection scenarios without requiring scenario-specific pixel-wise labeling. The attention mapping mechanism and gradient-based localization work universally across various anomaly types and imaging conditions, enabling scalable deployment across multiple applications with consistent reliability.
Data Source
AI summary
Systems and methods are disclosed for processing an image to detect anomalous pixels. An image classification is received from a trained convolutional neural network (CNN) for an input image with a positive classification being defined to represent detection of an anomaly in the image and a negative classification being defined to represent absence of an anomaly. A backward propagation analysis of the input image for each layer of the CNN generates an attention mapping that includes a positive attention map and a negative attention map. A positive mask is generated based on intensity thresholds of the positive attention map and a negative mask is generated based on intensity thresholds of the negative attention map. An image of segmented anomalous pixels is generated based on an aggregation of the positive mask and the negative mask.