Machine Vision Anomaly Localization with Hybrid UAD and Segmentation

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Conventional machine vision systems struggle to accurately and consistently detect anomalies on objects with high variation or missing part defects, as supervised and unsupervised learning models excel in different scenarios, leading to inefficiencies in anomaly detection.

Innovation Solution

A hybrid model combining supervised and unsupervised machine learning models for anomaly detection, utilizing a convolutional autoencoder for unsupervised anomaly detection and a convolutional neural network for supervised segmentation, to generate a hybrid output mask identifying localized anomalies.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a supervised machine learning model is used for anomaly detection, then accuracy is improved for objects with high variation, but the model fails to detect missing part defects

Engineering Contradiction:

Improveanomaly detection accuracyVSAvoiddetection capability across different defect types

Core Design Contradiction:

Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The patent combines supervised and unsupervised machine learning models into a hybrid system. The supervised model (e.g., CNN) processes images to detect anomalies where high variation exists, while the unsupervised model (e.g., autoencoder) simultaneously processes the same images to detect missing part defects. The outputs of both models are merged to produce a comprehensive anomaly detection result, thereby achieving both accuracy for high variation objects and versatility across different defect types.

Inventive Principle:

Principle #5Merging (Combining)

2Adaptability or versatility

If an unsupervised machine learning model is used for anomaly detection, then detection capability for missing part defects is improved, but accuracy decreases for objects with high variation

Engineering Contradiction:

Improvedetection capability for missing part defectsVSAvoidanomaly detection accuracy

Core Design Contradiction:

Adaptability or versatilityVSMeasurement precision

Solution Approach 1:

The hybrid model merges the strengths of unsupervised learning (detecting missing parts) with supervised learning (accurate anomaly localization). The unsupervised model's output is integrated with the supervised model's precise anomaly detection capabilities, creating a system that maintains high accuracy while gaining versatility for missing part detection.

Inventive Principle:

Principle #5Merging (Combining)

3Adaptability or versatility

If separate supervised and unsupervised models are used for different scenarios, then specialized detection capability is improved, but system complexity increases

Engineering Contradiction:

Improvescenario-specific detection capabilityVSAvoidsystem architecture complexity

Core Design Contradiction:

Adaptability or versatilityVSDevice complexity

Solution Approach 1:

Instead of maintaining separate systems for supervised and unsupervised detection, the patent merges both approaches into a single hybrid model architecture. This unified system processes images through both supervised and unsupervised pathways simultaneously, reducing system complexity while maintaining the specialized detection capabilities of each approach.

Inventive Principle:

Principle #5Merging (Combining)

Solution Approach 2:

The hybrid model serves multiple functions within a single system: it performs supervised anomaly detection for high variation objects, unsupervised detection for missing parts, and combines both results for comprehensive anomaly identification. This multi-functionality eliminates the need for separate specialized systems.

Inventive Principle:

Principle #6Universality (Multi-functionality)

4Device complexity

If conventional single-model approaches are used, then system simplicity is maintained, but anomaly detection accuracy and consistency deteriorate

Engineering Contradiction:

Improvesystem architecture simplicityVSAvoidanomaly detection accuracy

Core Design Contradiction:

Device complexityVSMeasurement precision

Solution Approach 1:

The patent merges two complementary machine learning approaches into a unified hybrid model, where the supervised component provides accurate anomaly localization and the unsupervised component ensures comprehensive defect detection. This combination improves both accuracy and consistency while maintaining reasonable system complexity through integrated architecture.

Inventive Principle:

Principle #5Merging (Combining)

Data Source

PatentUS12380565B2Systems and methods for precise anomaly localization within content captured by a machine vision camera

Publication Date: 2025.08.05 ZEBRA TECHNOLOGIES CORP

AI summary

Systems and methods for anomaly localization within content captured by a machine vision camera are disclosed herein. An example method includes receiving, at an application executed on a user computing device communicatively coupled to a machine vision camera, an image captured by the machine vision camera, the image including a target object with one or more localized anomalies. The example method further includes generating, by applying an unsupervised anomaly detection (UAD) module to the image, an anomaly heatmap corresponding to the image; and concatenating, by a concatenation module, the image and the anomaly heatmap into a multi-channel image. The example method also includes generating, by applying a supervised segmentation (SS) module to the multi-channel image, a hybrid output mask identifying the one or more localized anomalies within the image.