Metal Part Position Detection Under Surface Abnormalities

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing inspection devices struggle to accurately detect the position of inspection targets with surface abnormalities such as rust, distortion, or deposits, as these abnormalities affect the matching between captured images and reference patterns, leading to position detection errors.

Innovation Solution

The inspection device employs machine learning to generate a position detection model using image information of training images with surface abnormalities, allowing it to specify the position of inspection targets by applying the inspection image to this model, thereby enhancing detection accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If conventional image matching methods are used to detect the position of metal terminals, then the inspection device can detect the position under normal conditions, but the detection accuracy decreases when surface abnormalities such as rust, distortion, or deposits are present

Engineering Contradiction:
Improveposition detection accuracyVSAvoiddetection reliability under abnormal conditions
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The patent changes the fundamental parameter of the detection method from conventional image matching to machine learning-based classification. The system trains a classification model using multiple types of training images (normal images, abnormal images, and synthetic images with added abnormalities) to enable accurate position detection regardless of surface conditions. This parameter change allows the system to maintain high detection accuracy even when surface abnormalities are present.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent implements preliminary action by pre-training the classification model with diverse training images that include various abnormal conditions before actual inspection. The system prepares a comprehensive dataset containing normal metal terminal images, images with various abnormalities (rust, distortion, deposits), and synthetic images with added abnormality patterns. This preliminary training enables the model to recognize and accurately detect positions under diverse conditions without requiring real-time adaptation.

Inventive Principle:
Principle #10Preliminary action

2Measurement precision

If the inspection device uses reference pattern matching to identify metal terminals, then it can achieve accurate position detection for normal terminals, but the glossy portion changes with surface abnormalities causing matching failures

Engineering Contradiction:
Improveposition detection accuracyVSAvoidadaptability to surface abnormalities
Core Design Contradiction:
Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The patent fundamentally changes the detection parameter from reference pattern matching to machine learning-based classification. Instead of comparing images against a fixed reference pattern, the system uses a trained classification model that has learned to identify metal terminals based on their inherent features. This parameter change provides adaptability to surface abnormalities because the model was trained on diverse images including various abnormal conditions, enabling it to recognize terminals regardless of rust, distortion, or deposits.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent uses synthetic image generation as a form of copying to create additional training data. The system generates synthetic images by superimposing abnormality patterns (rust, distortion, deposit patterns) onto normal metal terminal images. These copied and modified images serve as training samples that teach the classification model to recognize terminals under various abnormal conditions, thereby improving the model's adaptability without requiring physical samples of every possible defect type.

Inventive Principle:
Principle #26Copying

3Measurement precision

If the inspection device excludes pixels near glossy portions to reduce inspection errors, then it can improve matching accuracy under normal conditions, but it cannot detect position when surface abnormalities cause glossy portion changes

Engineering Contradiction:
Improvematching accuracyVSAvoiddetection reliability with surface defects
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The patent changes the approach from selective pixel exclusion to comprehensive image classification. Instead of trying to identify and exclude problematic pixels (which becomes difficult when surface abnormalities change the glossy portion characteristics), the system uses a classification model that processes the entire image. The model learns to identify metal terminals based on multiple features across the whole image, making it robust to changes in glossy portion characteristics caused by surface abnormalities.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent employs synthetic image copying and transformation to create diverse training samples that include various glossy portion variations. By generating synthetic images with different abnormality patterns superimposed on normal images, the system creates a comprehensive training dataset that teaches the classification model to recognize terminals regardless of glossy portion changes. This copying approach allows the model to learn from simulated abnormal conditions without requiring actual defective samples.

Inventive Principle:
Principle #26Copying

Data Source

PatentEP4001831B1Inspection device, inspection method, and program
Publication Date: 2026.02.18 MITSUBISHI ELECTRIC CORP
  • EP4001831B1 patent drawingFigure 1
  • EP4001831B1 patent drawingFigure 2A
  • EP4001831B1 patent drawingFigure 2B

AI summary

A position detector (312) applies an image to be inspected to a position detection model generated by machine learning using image information of a training image to calculate a score for a portion of an image to be inspected. The position detector (312) specifies a portion of the image where the calculated score is a score that satisfies a predetermined relationship to detect the position of a metal part in the image to be inspected. A position outputter (314) outputs the position of the metal part detected by the position detector (312).