Image Inspection With Divided-Image Context Restoration
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image inspection systems using prelearned models struggle to accurately restore special patterns in good-article images and may erroneously generate defective-article patterns, leading to incorrect defect detection.
Innovation Solution
The system divides inspection images into multiple parts and incorporates surrounding images to generate restored images using a prelearned model trained on good-article images, enhancing pattern restoration and suppressing defective-article patterns.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a prelearned model with low expressive ability is used to restore images, then the model complexity is reduced and processing speed is improved, but the special patterns in good-article images cannot be accurately restored
Solution Approach 1:
The inspection image is divided into multiple divided images, and the prelearned model processes each divided image separately. This segmentation allows the model to focus on local patterns while the surrounding image information provides contextual guidance, improving restoration accuracy without requiring an overly complex global model.
Solution Approach 2:
The surrounding image acts as an intermediary that provides contextual information to the prelearned model. By inputting both the divided image and its surrounding context, the model can accurately restore special patterns while suppressing defective-article patterns, effectively using the surrounding area as a mediator to guide the restoration process.
2Reliability
If only the divided inspection image is used as input to the prelearned model, then the processing is simpler and faster, but the restored image may contain erroneous defective-article patterns
Solution Approach 1:
The input data is segmented into two distinct components: the divided inspection image and the surrounding image. This segmentation allows the system to process local details and global context separately, improving reliability by preventing erroneous pattern generation while maintaining a manageable input structure.
Solution Approach 2:
The solution adds a spatial dimension by incorporating the surrounding image context around each divided image. This dimensional expansion from single-image input to multi-region input provides the model with contextual information necessary for accurate pattern recognition and defect suppression.
3Productivity
If the prelearned model processes each channel separately, then the processing parallelism is improved, but the reconstructed image may lose channel-specific pattern information
Solution Approach 1:
The image processing is segmented into multiple channels, with each channel processed independently by the prelearned model. This segmentation enables parallel processing while the surrounding image context ensures that channel-specific pattern information is preserved through contextual guidance from adjacent regions.
Data Source
AI summary
An image inspection device includes: a divided image generation part for inputting a divided inspection image obtained by dividing an image of an inspection object and a surrounding-containing image that includes an image based on at least some of the surrounding images of the divided inspection image to a prelearned model having been trained so as to accept as inputs a divided good-article image obtained by dividing an image of a good-article inspection object and an image including an image based on at least some of the surrounding images of the divided good-article image and output a restored divided image; and an inspection part for performing inspection of the inspection object on the basis of the restored divided image generated by the divided image generation part.