Patched Feature Mapping for Accurate Image Anomaly Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional anomaly detection systems in images often produce inaccurate results, including false positives and false negatives, due to the difficulty in distinguishing between normal and anomalous samples, especially when anomalies take various forms and are rare.
Innovation Solution
Enhance anomaly detection by extracting multiple layers of features from an image using a pretrained convolutional neural network, resizing and concatenating these features to form a tensor feature map, performing center cropping, and calculating Mahalanobis distances to a feature distribution mean for each patch, followed by averaging maximum distances to determine an image anomaly score.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional anomaly detection methods are used, then the detection process is simple, but the accuracy is low with many false positives and negatives
Solution Approach 1:
The image is divided into multiple patches that are processed independently through the neural network. Each patch is extracted, fed through the CNN to generate feature maps, and then reassembled to form the complete image representation. This segmentation allows for more granular analysis and improves detection accuracy by capturing local anomalies that might be missed in global processing.
Solution Approach 2:
The patent transforms the 2D image data into a 1D feature vector by flattening the feature maps extracted from multiple CNN layers. This dimensional transformation enables the application of statistical methods like Mahalanobis distance calculation, which operates more effectively in the transformed feature space, thereby improving anomaly detection accuracy.
2Reliability
If multiple layers of features are extracted and concatenated, then the feature representation is more comprehensive, but the computational complexity increases
Solution Approach 1:
The patent uses a pretrained convolutional neural network that has already learned optimal feature extraction patterns during the training phase. By leveraging these prelearned features from multiple layers, the system achieves more reliable anomaly detection without needing to train new complex models, thus reducing the computational energy required at inference time while maintaining high reliability.
3Measurement precision
If center cropping is performed on the tensor feature map, then the focus on central features is improved, but information from peripheral areas is lost
Solution Approach 1:
The center cropping operation is applied selectively to the tensor feature maps extracted from different CNN layers. This local processing approach focuses the analysis on central regions where anomalies are most likely to occur, improving the precision of feature distribution modeling for those critical areas while accepting some information loss in peripheral regions as a necessary trade-off for enhanced detection accuracy.
Data Source
AI summary
One or more systems, devices, computer program products, and/or computer-implemented methods provided herein relate to accurate anomaly detection in images using patched features. According to an embodiment, an extraction component can extract multiple layers of features from one or more patches of an image using a pretrained convolutional neural network (CNN). A feature mapping component can concatenate the features from the multiple layers to generate a tensor feature map comprising a one-dimensional feature vector for respective patches. A cropping component can perform center cropping on the tensor feature map. A calculation component can calculate a distance to a feature distribution mean for respective patches.