Dual Learning Model for Image Defect Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing techniques for automated appearance inspection using artificial intelligence or deep learning struggle to accurately detect both high-frequency and low-frequency image defects in production settings, as excessive compression or segmentation can lead to missed detections of specific types of defects.
Innovation Solution
A learning model generating device and type identification system that employs two learning models: one for high-frequency defect detection through segmentation and another for low-frequency detection through compression, allowing for the generation and use of segmented and compressed images to infer the type of image defects effectively.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If image compression is applied to reduce processing load, then processing efficiency is improved, but detection accuracy of high-frequency defects deteriorates
Solution Approach 1:
The patent divides the inspection process into two separate pathways: one for high-frequency defects using segmented images without compression, and another for low-frequency defects using compressed images. This segmentation allows each pathway to be optimized for its specific defect type, resolving the contradiction between compression efficiency and detection accuracy.
Solution Approach 2:
Different image processing qualities are applied to different defect detection tasks. High-frequency defect detection uses high-quality uncompressed segmented images, while low-frequency defect detection uses compressed images. This local quality differentiation ensures optimal detection accuracy for each defect type while maintaining overall processing efficiency.
2Measurement precision
If image segmentation is applied to detect high-frequency defects, then detection accuracy is improved, but processing complexity increases
Solution Approach 1:
The patent creates two independent processing pathways with dedicated learning models: one for segmented images targeting high-frequency defects, and another for compressed images targeting low-frequency defects. This segmentation of the inspection system itself reduces the complexity burden on each individual pathway while maintaining high detection accuracy.
Solution Approach 2:
The patent adds a dimensional separation by creating parallel processing pathways instead of trying to handle all defect types in a single complex pathway. This dimensional change from a single complex process to multiple specialized processes reduces overall system complexity while improving detection accuracy for each defect type.
3Device complexity
If single learning model is used for all defect types, then system complexity is reduced, but detection accuracy for specific defect types deteriorates
Solution Approach 1:
The patent applies local quality by training different learning models with different qualities of input images for different defect types. The first learning model processes high-quality segmented images for high-frequency defects, while the second learning model processes compressed images for low-frequency defects, ensuring optimal detection accuracy for each specific defect type.
Solution Approach 2:
The patent segments the learning model system into two specialized models rather than using a single general-purpose model. This segmentation allows each model to be specialized for its specific defect type and image processing requirements, improving detection accuracy without significantly increasing overall system complexity.
Data Source
AI summary
A learning model generating device includes a first image reading device and a first control device. The first control device includes a processor and functions, through the processor executing a first control program, as a first segmenter, a learning model generator, and a first compressor. The first segmenter segments each of images of training prints obtained by reading performed by the first image reading device. The learning model generator learns segmented images to generate a first learning model for use in inferring a type of an image defect. The first compressor compresses each of the images of the training prints. The first segmenter segments each of compressed images obtained by compression performed by the first compressor. The learning model generator learns compressed and segmented images to generate a second learning model for use in inferring a type of an image defect.