Visual Inspection Classification to Reduce Defect Overdetection

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing visual inspection methods using machine learning often result in high overdetection of non-defective products and overlooking of defective products due to variations in brightness and shape, leading to increased inspection costs and risks of shipping defective products.

Innovation Solution

A computer-based visual inspection method that includes a processor for image acquisition, pass/fail determination, overdetection determination, non-defect estimation parameter learning, and overdetection determination parameter learning, using non-defect images to accurately distinguish between non-defective and defective products by learning appropriate boundaries in feature space.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Extent of automation

If machine learning is performed using only images of non-defective products, then inspection cost is reduced and automation is improved, but overdetection of non-defective portions increases and measurement precision deteriorates

Engineering Contradiction:

Improveautomation of inspectionVSAvoidaccuracy of defect detection

Core Design Contradiction:

Extent of automationVSMeasurement precision

Solution Approach 1:

The system performs preliminary classification by first determining whether a non-defect candidate image or defect candidate image is input, then applies different processing paths accordingly. This preliminary action allows the system to handle non-defective and defective products differently, improving overall detection accuracy while maintaining automation.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The system applies different determination parameters locally based on the input image type. Non-defect determination parameters are used for non-defect candidate images, while defect determination parameters are used for defect candidate images. This local differentiation improves measurement precision without compromising automation.

Inventive Principle:

Principle #3Local quality

2Ease of manufacture

If machine learning is performed using only images of non-defective products, then collection cost of learning images is reduced, but overlooking of defective products increases and reliability deteriorates

Engineering Contradiction:

Improveease of learning image collectionVSAvoidreliability of defect detection

Core Design Contradiction:

Ease of manufactureVSReliability

Solution Approach 1:

The system performs preliminary determination to identify whether the input image is a non-defect candidate or defect candidate before applying the final determination. This preliminary action enables the system to maintain high reliability by directing defect candidate images through an appropriate processing path, even though learning is performed using only non-defective product images.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The system introduces an intermediary determination step that acts as a mediator between the simplified learning process and the final defect detection. This intermediary layer ensures that defective products are not overlooked while maintaining the ease of learning image collection.

Inventive Principle:

Principle #24Intermediary (Mediator)

3Adaptability or versatility

If variations in brightness and shape among non-defective products are considered, then adaptability is improved, but overdetection increases and measurement precision deteriorates

Engineering Contradiction:

Improveadaptability to product variationsVSAvoidprecision of defect identification

Core Design Contradiction:

Adaptability or versatilityVSMeasurement precision

Solution Approach 1:

The system performs preliminary determination to identify the type of input image before applying strict defect criteria. This preliminary action allows the system to adapt to product variations in brightness and shape while maintaining precise defect identification by applying appropriate determination parameters based on the preliminary classification.

Inventive Principle:

Principle #10Preliminary action

4Productivity

If the number of images detected as defects is reduced, then inspection efficiency is improved, but defect detection capability may be compromised

Engineering Contradiction:

Improveinspection efficiencyVSAvoiddefect detection capability

Core Design Contradiction:

ProductivityVSReliability

Solution Approach 1:

The system performs preliminary determination to classify images as non-defect candidates or defect candidates before final inspection. This preliminary action improves inspection efficiency by streamlining the processing of non-defective products while maintaining defect detection capability through appropriate handling of defect candidate images.

Inventive Principle:

Principle #10Preliminary action

Data Source

PatentUS12499532B2Computer and visual inspection method for identifying defective products

Publication Date: 2025.12.16 HITACHI LTD

AI summary

The present invention includes image acquiring for acquiring an inspection image of a target object, pass/fail determining for determining whether the inspection image acquired in the image acquiring is the inspection image of a non-defect candidate or the inspection image of a defect candidate, overdetection determining for determining whether the inspection image determined as a defect candidate in the pass/fail determining is the inspection image of overdetection or the inspection image of non-overdetection which is not overdetection, non-defect estimation parameter learning for learning a non-defect estimation parameter used in the pass/fail determining using a learning non-defect image acquired in the image acquiring, and overdetection determination parameter learning for learning an overdetection determination parameter used in the overdetection determining using the learning non-defect image.