Neural Network Defect Detection Without Precise Positioning

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing image inspection methods require precise positioning and frequent relearning when product specifications change, leading to reduced accuracy and increased workload due to the need for new reference images.

Innovation Solution

A processing device and method utilizing a neural network that combines inspection and reference images, allowing for defect detection without the need for precise positioning and relearning, by learning defect patterns rather than entire reference images, thus enabling efficient adaptation to changes in product specifications.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If a neural network is used to detect abnormalities by learning a new reference image when product specifications change, then the system becomes more adaptable to different products, but the volume of work required for learning increases

Engineering Contradiction:

Improveadaptability to product specification changesVSAvoidvolume of work required for learning

Core Design Contradiction:

Adaptability or versatilityVSLoss of time

Solution Approach 1:

The patent extracts only the defect patterns from the reference image rather than learning the entire reference image. The neural network is trained to detect specific defect characteristics (such as scratches, dents, or contamination) by separating these defect patterns from the normal product features. This extraction approach allows the system to adapt to different products by learning only the relevant defect patterns rather than relearning entire product specifications.

Inventive Principle:

Principle #2Taking out (Extraction)

Solution Approach 2:

The patent segments the image processing task into distinct components: normal product features and defect patterns. By dividing the reference image into these separate elements and training the neural network to recognize only the defect segment, the system reduces the learning workload while maintaining adaptability. The segmentation allows independent learning of defect characteristics that can be applied across different product types.

Inventive Principle:

Principle #1Segmentation

2Measurement precision

If traditional image inspection methods are used with reference images, then positioning accuracy is improved, but the system becomes less adaptable when product specifications change

Engineering Contradiction:

Improvepositioning accuracyVSAvoidadaptability to specification changes

Core Design Contradiction:

Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The patent replaces the mechanical positioning system with a neural network-based pattern recognition system. Instead of relying on precise physical alignment between inspection images and reference images, the system uses learned defect patterns to identify abnormalities regardless of positioning variations. This substitution of mechanical alignment with intelligent pattern recognition maintains measurement precision while dramatically improving adaptability to specification changes.

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

Data Source

PatentUS11436717B2Image processing device and image processing method

Publication Date: 2022.09.06 PANASONIC INTELLECTUAL PROPERTY MANAGEMENT CO LTD

AI summary

A first input unit receives an inspection image subject to inspection. A second input unit receives a normal reference image that should be referred to. A processing unit subjects the inspection image input to the first input unit and the reference image input to the second input unit to a process in a neural network. An output unit outputs information that results from the process in the processing unit and relates to an item of defect included in the inspection image.