Image Inspection Using Scanned Reconstruction Error Areas
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image classification methods using variational autoencoders struggle to accurately distinguish between normal and abnormal images when the abnormal region is small or has a subtle difference in color tone, leading to decreased identification accuracy due to ambiguous reconstruction error thresholds.
Innovation Solution
An inspection device that calculates reconstruction error by setting a detection area including multiple pixels and scanning this area on the image data, using overlapping detection areas and calculating the error using the square mean value of differences in parameters like hue, brightness, or luminance, allowing for accurate identification of abnormal regions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If reconstruction error is calculated for each unit pixel, then measurement precision is improved for minute abnormalities, but device complexity increases and processing time extends
Solution Approach 1:
The patent divides the image into multiple detection areas (blocks) of predetermined size, and calculates reconstruction errors for each detection area rather than for every single pixel. This segmentation approach reduces the total number of calculations while maintaining effective abnormality detection, as each detection area covers multiple pixels and captures local variations.
Solution Approach 2:
The patent calculates reconstruction errors for all detection areas across the entire image, including areas that may not contain abnormalities. This exhaustive approach ensures that no potential abnormal region is missed, while the predetermined detection area size optimizes the balance between detection coverage and computational load.
2Productivity
If detection area size is increased to cover larger regions, then processing speed is improved, but measurement precision decreases for small abnormalities
Solution Approach 1:
The patent segments the image into multiple detection areas of a predetermined size that is optimized to balance detection precision and processing speed. This segmentation allows the system to process larger images faster while maintaining the ability to detect abnormalities within each segment.
Solution Approach 2:
The patent introduces a hierarchical dimension by first dividing the image into detection areas, then further dividing each detection area into sub-regions for detailed comparison between input and output images. This multi-level division strategy enables efficient processing while preserving detection precision for small abnormalities.
3Measurement precision
If threshold is set to distinguish normal from abnormal images, then identification accuracy is improved, but reliability decreases when abnormal regions are small or subtle
Solution Approach 1:
The patent calculates reconstruction errors for all detection areas across the entire image, ensuring comprehensive coverage. By aggregating error information from multiple detection areas, the system achieves more reliable classification decisions even when individual areas show subtle variations or when abnormal regions are small.
Data Source
AI summary
An inspection device 1 that inspects whether an image to be inspected is a normal image or an abnormal image includes a learning unit 20A that learns so as to enable reconstruction of normal image data, an error calculation unit 23 that calculates a reconstruction error when a plurality of pieces of input image data are input to the learning unit 20A, a threshold calculation unit 24 that calculates a threshold based on the reconstruction error, and an identification unit 25 that performs identification based on the threshold. The error calculation unit 23 sequentially compares input image data DI and output image data DO while scanning a detection area E on an image data.