Analysis Device for Visualizing Pseudo-Defect Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image inspection devices generate pseudo-defective product images that may incorrectly classify non-defective products as defective, and lack verification of defect severity, leading to inaccurate determination results.
Innovation Solution
An analysis device that generates composite images by synthesizing pseudo-defective parts with non-defective product images, outputs these images to a trained determination device, and displays labels in an array based on feature quantities to visualize accuracy, using an acquisition, extraction, generation, compositing, and display control units.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If all pseudo-defective product images are trained as defective product images, then the determination device can be trained to detect defects, but the determination accuracy decreases because some pseudo-defective images actually correspond to non-defective products
Solution Approach 1:
The patent applies partial action by generating multiple pseudo-defective images with varying defect magnitudes and selectively training only those that represent actual defects. The system generates a range of pseudo-defective images with different feature quantities, then uses a determination device to evaluate which ones truly represent defective products, thereby avoiding the excessive action of training all generated images regardless of their actual defect representation.
Solution Approach 2:
The patent changes parameters by varying the feature quantities of pseudo-defective parts across multiple generated images. By adjusting parameters such as defect size, position, and magnitude, the system creates a diverse set of training images with different characteristics, allowing the determination device to learn to distinguish between actual defects and artifacts of image generation.
2Device complexity
If the image inspection device cannot verify the degree of defect, then the training process is simplified, but the ability to understand and adjust classification thresholds is reduced
Solution Approach 1:
The patent adds another dimension to the training process by generating multiple pseudo-defective images with varying feature quantities (different magnitudes and types of defects). This dimensional expansion allows the determination device to learn not just binary defect detection but also to understand defect severity levels, enabling better threshold adjustment without significantly increasing overall system complexity.
Data Source
AI summary
An analysis device for visualizing an accuracy of a trained determination device includes an acquisition unit acquiring an image pair of a non-defective product image and a defective product image, an extraction unit extracting an image region of a defective part of the defective product, a generation unit generating a plurality of image regions of pseudo-defective parts, a compositing unit synthesizing each of the image regions of the plurality of pseudo-defective parts with the non-defective product image to generate a plurality of composite images having different feature quantities, an unit outputting the plurality of composite images to the determination device and acquiring a label corresponding to each of the plurality of composite images from the determination device, and a display control unit displaying an object indicating the label corresponding to each of the plurality of composite images in an array based on the feature quantities.