Surface Defect Detection Model Training Without Defect Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep learning-based surface defect detection models face challenges in acquiring sufficient product images representing defects due to their rarity, leading to high collection costs and instability, and require cumbersome threshold settings for reliable detection.
Innovation Solution
Training a surface defect detection model using normal product images and external images irrelevant to the product, employing a deep neural network with a loss function based on distance matrices of feature maps to enhance reliability and reduce the need for defect images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If product images representing defects are collected for training the defect detection model, then the model can learn to detect defects, but the collection cost is high and the number of defect images is insufficient
Solution Approach 1:
Instead of training the model to directly detect defects using defect images, the patent inverts the approach by training the model to recognize normal product images and external images, then using the trained model to identify abnormalities by comparing against the learned normal patterns. This eliminates the need for scarce defect images while still achieving defect detection capability.
Solution Approach 2:
The patent introduces external images as an intermediary element in the training process. These external images serve as a mediator between the normal product images and the defect detection task, helping the model learn to distinguish normal from abnormal patterns without requiring actual defect images during training.
2Ease of manufacture
If a threshold is set to determine whether the product image represents defects or not, then the defect detection model can be trained, but the stability of the trained model varies with the threshold
Solution Approach 1:
The patent removes the traditional threshold-setting step from the training process by inverting the detection logic. Instead of training the model to classify images as defective or normal with a threshold, the model learns to generate similarity scores between input images and normal images, eliminating the need for arbitrary threshold selection and improving model stability.
3Ease of operation
If manual visual observation is used to detect surface defects, then the detection process is simple, but it is time-consuming and laborious with low detection accuracy
Solution Approach 1:
The patent replaces the manual visual observation system with an automated machine vision system based on deep learning. The trained model automatically processes product images to detect defects, substituting human operators while maintaining simplicity in operation and dramatically improving detection efficiency and accuracy.
Data Source
AI summary
A method for training a surface defect detection model as well as a method and a system for detecting a surface defect are provided. The method and the system for detecting a surface defect adopt the trained surface defect detection model. The method for training a surface defect detection model includes acquiring a normal image of a product and an external image that is irrelevant to the product and inputting the normal image and the external image into a deep neural network-based surface defect detection model, and training the deep neural network-based surface defect detection model to obtain the trained surface defect detection model. The normal image represents the product surface of no defect on this surface.