Deep Learning Vision Inspection for Module Line Defect Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current vision inspection systems for module lines have high error detection rates, requiring module lines to stop operations for manual defect identification and achieving only a 5% NG error detection rate.
Innovation Solution
A chassis module vision inspection system based on deep learning, utilizing a GT generation module, learning module, and interface module to rapidly process images and reduce detection errors to 0%, comprising a GT generation module for generating ground truth, a learning module for deep learning, and an interface module for defect detection using a YOLO detection model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional vision inspection systems use pattern, blob and color tools to check module quality, then the inspection process can be automated, but the error detection rate remains high at approximately 5% and requires manual intervention
Solution Approach 1:
The patent replaces traditional vision inspection tools (pattern, blob, and color tools) with a deep learning-based detection model. This substitution transforms the mechanical rule-based inspection system into an intelligent system that uses neural networks to automatically identify defects, thereby reducing the error detection rate from 5% to 0% and eliminating the need for manual intervention.
Solution Approach 2:
The patent changes the fundamental parameters of the inspection system by transitioning from traditional image processing algorithms to deep learning models. This parameter change involves using trained neural network weights and configurations that can adaptively learn defect patterns, significantly improving detection accuracy and reliability compared to fixed-rule-based systems.
2Measurement precision
If the vision program determines a module is defective using traditional methods, then the module line must stop immediately for manual defect identification, but this causes production downtime and reduces productivity
Solution Approach 1:
The patent replaces manual defect identification with an automated deep learning-based vision inspection system. The system provides real-time defect detection and classification, eliminating the need for production line stoppages and manual intervention, thereby maintaining high productivity while achieving accurate defect identification.
Solution Approach 2:
The patent enables continuous operation of the module line by implementing real-time automated defect detection. The deep learning model continuously processes images and identifies defects without requiring production stoppages, ensuring the useful action of manufacturing continues uninterrupted while maintaining quality control.
3Extent of automation
If traditional vision inspection systems are used, then the inspection process can be performed, but the error detection rate is only 5% requiring frequent manual checks
Solution Approach 1:
The patent replaces traditional automated vision inspection tools with a deep learning-based system. This substitution maintains the automated nature of the inspection process while dramatically improving defect detection accuracy from 5% to 0% error rate, eliminating the need for manual checks.
Solution Approach 2:
The patent performs preliminary action by training deep learning models with extensive defect data before deployment. This preliminary training enables the system to achieve high detection accuracy from the start, eliminating the need for subsequent manual corrections and retraining that would be required with traditional systems.
Data Source
AI summary
The present disclosure relates to a vision inspection system based on deep learning and a vision inspection method of. The vision inspection system based on deep learning according to the present disclosure includes a GT generation module that generates a GT for a region of interest of a car part image, a learning module that receives learning data from the GT generation module, performs learning based on deep learning, and outputs a weight file, and an interface module that detects a defect with respect to an image file received from a vision program by using the weight file, and returns a defect detection result to the vision program.