Patch-Embedding Image Clustering for Anomaly Type Separation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional anomaly detection models provide only binary classification outputs, failing to differentiate between various types of anomalous data, which limits their utility in identifying specific defects and requires manual sorting to isolate relevant anomalies.
Innovation Solution
An image clustering anomaly detector extracts patch embeddings from images, determines weighted average distances between these embeddings using a neural network model, and applies hierarchical or spectral clustering to assign images into groups based on anomaly types, enabling differentiated anomaly detection without labeled data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional anomaly detection models are used, then binary classification output is achieved, but differentiation between various types of anomalous data is lost
Solution Approach 1:
The patent segments the anomaly detection task into multiple clusters, where each cluster represents a different type of anomaly. Instead of a single binary classification, the model divides the anomaly space into distinct groups, allowing differentiation between various anomaly types while maintaining the binary classification framework.
Solution Approach 2:
The patent introduces a new dimension to the output space by adding cluster assignments to the binary classification. This transforms the output from a single binary label to a combination of binary classification and cluster membership, enabling anomaly type differentiation without complicating the base detection mechanism.
2Measurement precision
If manual sorting is applied to isolate particular anomaly types, then specific anomaly types can be identified, but time and productivity are reduced
Solution Approach 1:
The patent enables the anomaly detection system to automatically perform the sorting and classification task that would otherwise require manual intervention. By integrating clustering algorithms, the system self-organizes the anomaly data into meaningful groups, eliminating the need for manual sorting while maintaining anomaly type identification capability.
Solution Approach 2:
The patent replaces the mechanical manual sorting process with an automated computational clustering system. The clustering algorithms automatically group anomalies by type, substituting human labor with algorithmic processing, thereby maintaining precision while dramatically improving productivity.
3Ease of operation
If binary classification is used, then simple output is achieved, but meaningful insights about different anomaly types are lost
Solution Approach 1:
The patent merges binary classification with clustering results to create a comprehensive anomaly detection system. By combining the simplicity of binary classification with the informative power of cluster assignments, the system maintains ease of operation while providing meaningful insights about different anomaly types through the additional cluster dimension.
Data Source
AI summary
A computer-implemented method includes receiving an anomaly clustering request that requests data processing hardware to assign each image of a plurality of images into one of a plurality of groups. The method also includes obtaining a plurality of images. For each respective image, the method includes extracting a respective set of patch embeddings from the respective image, determining a distance between the respective set of patch embeddings and each other set of patch embeddings, and assigning the respective image into one of the plurality of groups using the distances between the respective set of patch embeddings and each other set of patch embeddings.