Deep Learning Quality Inspection Using Non-Defective Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In manufacturing fields, especially in injection processes, the low defective rate makes it difficult to collect sufficient defective data for AI training, leading to unstable AI accuracy and limited application of surface inspection AI due to the complexity of image data and varying defect characteristics, which requires clear images and specific classification criteria.
Innovation Solution
A deep learning system that learns only from non-defective manufactured product data, using preprocessing techniques like resizing and padding to enhance image quality, and generating fake defective features by adding Gaussian noise, which stabilizes the model architecture and loss function for accurate defect inspection without requiring extensive labeling or prior defect learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional AI classification methods are used requiring large amounts of defective data, then AI accuracy can be improved, but data collection becomes difficult due to low defective rate
Solution Approach 1:
The patent inverts the conventional approach by training the AI model only on non-defective product images instead of requiring defective product images. The model learns what normal products look like and identifies deviations as defects, eliminating the need to collect scarce defective data while maintaining high detection accuracy
Solution Approach 2:
The patent generates synthetic defective images by copying non-defective images and adding various types of noise (Gaussian noise, salt-and-pepper noise, etc.) to simulate different defect conditions. This creates abundant training data without requiring actual defective products
2Measurement precision
If image classification AI is applied with strict classification criteria, then detection accuracy improves, but application complexity increases due to varying defect characteristics
Solution Approach 1:
The patent changes the fundamental parameter of training data composition from defective images to non-defective images only. This simplifies the system by removing the need for complex defect classification criteria, as the model automatically learns to detect any deviation from the normal state without requiring pre-defined defect categories
Solution Approach 2:
The patent creates a universal defect detection system that can identify various types of defects (scratches, dents, discoloration, etc.) using a single model trained on non-defective images. The system handles diverse defect characteristics without requiring separate classification rules for each defect type
3Productivity
If preprocessing techniques like cropping are used to focus on specific areas, then image processing speed improves, but defects at edges may be missed
Solution Approach 1:
The patent applies data augmentation techniques that perform excessive preprocessing operations (adding noise, rotating, flipping, scaling) on non-defective images during training. This over-processing ensures the model learns robust features that maintain high detection reliability across the entire image area without sacrificing processing speed
Data Source
AI summary
A deep learning-based quality inspection system by learning only non-defective manufactured product data may include: an input unit receiving a non-defective manufactured product image data set; a preprocessor preprocessing a model to learn the images with a same size by not applying a cropping task to cut and process only an area at a specific location within each image for each of the plurality of images included in the image data set, but applying a resizing task of adjusting each image to a desired size and a padding task of adjusting the size of the image while maintaining a ratio of each image as it is; and a controller extracting a non-defective manufactured product feature which becomes a non-defective manufactured product criterion from the preprocessed image, and generating a plurality of fake defective manufactured product features by adding a Gaussian noise feature to the extracted non-defective manufactured product feature.