Inference Model Training for Small Foreign Object Detection

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for detecting foreign objects in images, such as those obtained by imaging packaged items with X-rays, often fail to accurately identify small foreign objects within a plurality of similar items, particularly when the foreign object is smaller than each individual item.

Innovation Solution

A model generation method that trains an inference model using four types of images: a first normal image, a second normal image with a foreign object added, a second foreign object image, and a third normal image with a foreign object not assumed to be a detection target, to enhance the model's ability to distinguish between the target object and foreign objects.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If an inference model is trained using only normal images and pseudo-defect images, then the model can learn basic defect patterns, but it fails to accurately detect small foreign objects among multiple similar items

Engineering Contradiction:
Improveforeign object detection accuracyVSAvoidmodel ability to distinguish target objects from foreign objects
Core Design Contradiction:
Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The training data is segmented into three distinct categories: normal images, foreign object images, and target object images. This segmentation allows the model to learn distinct features for each category, improving its ability to differentiate between small foreign objects and similar-looking target items during detection

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The model performs preliminary learning by separately training on target object images before final detection. This preliminary action of learning target object characteristics in advance enables the model to better distinguish between target items and foreign objects during the actual detection process

Inventive Principle:
Principle #10Preliminary action

2Measurement precision

If the inference model is trained with diverse foreign object images, then detection accuracy improves, but the training complexity and data preparation effort increase

Engineering Contradiction:
Improveforeign object detection accuracyVSAvoidtraining data preparation complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

Instead of manually creating complex foreign object images by adding defects to normal images, the approach is inverted by using actual captured images of foreign objects and target objects as training data. This inversion simplifies data preparation while maintaining high detection accuracy

Inventive Principle:
Principle #13The other way round (Inversion)

Data Source

PatentEP4625313A1Model generation method, model generation system, model generation program, foreign substance detection method, foreign substance detection system, foreign substance detection program, and inference model
Publication Date: 2025.10.01 HAMAMATSU PHOTONICS KK
  • EP4625313A1 patent drawingFigure 1(a)~1(b)
  • EP4625313A1 patent drawingFigure 2
  • EP4625313A1 patent drawingFigure 3

AI summary

A foreign object is detected with high accuracy. A model generation method is a method for generating an inference model used for detecting a foreign object included in an image with a target object captured, and includes: an acquisition step for training (S01) for acquiring, as images for training, a first normal image with a target object for training captured, a second normal image with a target object for training captured, a second foreign object image obtained by adding a foreign object assumed to be a detection target to the second normal image, a third normal image with a target object for training captured, and a third foreign object image obtained by adding a foreign object not assumed to be a detection target to the third normal image; and a model generation step (S02) for generating the inference model by performing training using the acquired images for training.