Semi-Supervised Image Defect Detection with Adaptive Thresholds

Solution Overview

Problem

Existing image defect detection systems require significant manpower and time for image marking and struggle with insufficient data for training, especially in low-defect scenarios, leading to inefficiencies and delayed deployment.

Innovation Solution

A semi-supervised learning model, such as a GANomaly model, is used to generate a threshold value and a standby inference model, reducing the need for manual marking and enabling adaptive adjustment of the threshold based on validation and testing images.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If manual marking of images is performed to train the AOI detection model, then detection accuracy is improved, but manpower requirements and time costs increase significantly

Engineering Contradiction:

Improvedetection accuracyVSAvoidtime costs

Core Design Contradiction:

Measurement precisionVSLoss of time

Solution Approach 1:

The system uses unsupervised learning to enable the model to automatically learn from unmarked images, making the system self-training without requiring manual marking. The model independently identifies patterns and defects in the data, eliminating the need for human annotators to mark images while still achieving effective detection capability.

Inventive Principle:

Principle #25Self-service

Solution Approach 2:

The patent replaces the manual mechanical marking process with an automated computational learning system. Instead of human operators manually marking images, the system uses unsupervised learning algorithms to automatically process and learn from raw image data, substituting human labor with automated intelligent processing.

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

2Measurement precision

If manual marking of images is performed to train the AOI detection model, then detection accuracy is improved, but the amount of labor required increases

Engineering Contradiction:

Improvedetection accuracyVSAvoidmanpower needs

Core Design Contradiction:

Measurement precisionVSExtent of automation

Solution Approach 1:

The system uses unsupervised learning to enable the model to automatically learn from unmarked images, making the system self-training without requiring manual marking. The model independently identifies patterns and defects in the data, eliminating the need for human annotators to mark images while still achieving effective detection capability.

Inventive Principle:

Principle #25Self-service

Solution Approach 2:

The patent replaces the manual mechanical marking process with an automated computational learning system. Instead of human operators manually marking images, the system uses unsupervised learning algorithms to automatically process and learn from raw image data, substituting human labor with automated intelligent processing.

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

3Reliability

If traditional supervised learning is used for training, then the model can be trained with marked images, but it fails when insufficient defective product images are available for training

Engineering Contradiction:

Improvemodel training capabilityVSAvoidadaptability to low-defect scenarios

Core Design Contradiction:

ReliabilityVSAdaptability or versatility

Solution Approach 1:

Instead of using supervised learning that requires marked defective images, the patent inverts the approach by using unsupervised learning that learns from unmarked images. The model learns normal patterns without labels and identifies defects as deviations from these patterns, enabling training in low-defect scenarios where traditional methods fail.

Inventive Principle:

Principle #13The other way round (Inversion)

Solution Approach 2:

The system uses unsupervised learning to enable the model to automatically learn from unmarked images, making the system self-training without requiring manual marking. The model independently identifies patterns and defects in the data, eliminating the need for human annotators to mark images while still achieving effective detection capability.

Inventive Principle:

Principle #25Self-service

Data Source

PatentUS12354255B2Image defect detecting system, generation method of image defect detecting system and non-transitory computer readable medium

Publication Date: 2025.07.08 WISTRON CORP

AI summary

A generation method of an image defect detecting system is provided. Said method includes: obtaining a plurality of validation difference scores respectively associated with a plurality of validation images based on a semi-supervised learning model; calculating a threshold value based on the plurality of validation difference scores; creating a standby inference model based on the plurality of validation difference scores; obtaining a testing difference score associated with a testing image based on the semi-supervised learning model; adjusting the threshold value by the standby inference model in response to the testing difference score and the threshold value; and outputting data comprising the testing difference score.