Few-Shot Defect Detection via Metric Learning and GAN Augmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current defect detection methods for large assembly components in high-end equipment face challenges with low precision, high labor intensity, and inefficiency due to the scarcity of defect data, leading to unreliable and subjective visual detection processes.
Innovation Solution
A few-shot defect detection method based on metric learning, utilizing a G2-Generative adversarial network (G2-GAN) for data enhancement and a convolutional neural network with a selective kernel module (SKM-CNN) for feature extraction, followed by transfer learning to a few-shot defect detection network (S2D2N), which performs target feature extraction and metric learning to identify and locate defects efficiently.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning algorithms are used for defect detection, then detection precision and robustness are improved, but the requirement for large amounts of labeled training data increases, which cannot be satisfied in few-shot scenarios
Solution Approach 1:
The method applies preliminary data enhancement using G2-GAN to generate synthetic defect samples before the actual detection task. This preliminary action creates an expanded training dataset that enables deep learning models to achieve high detection precision without requiring large amounts of real defect data.
Solution Approach 2:
The G2-GAN generates synthetic copies of defect samples by learning from limited real defect data and producing realistic artificial defect images. These copied samples augment the training dataset, allowing the model to learn effective defect features without needing extensive real defect data.
2Device complexity
If manual feature extraction methods are used for defect detection, then the system complexity is reduced, but detection efficiency and intelligence are significantly attenuated
Solution Approach 1:
The method replaces manual feature extraction (mechanical approach) with automatic deep learning-based feature extraction. The SKM-CNN and S2D2N networks automatically learn optimal defect features from images, eliminating the need for hand-crafted features while significantly improving detection efficiency and intelligence.
3Adaptability or versatility
If visual detection by human eyes is used, then flexibility in handling various defect types is maintained, but labor intensity increases and detection accuracy is affected by subjective experience
Solution Approach 1:
The S2D2N detection network is designed with universal feature extraction capabilities that can handle multiple defect types and scenarios. The metric learning component learns a unified feature space that accommodates various defect characteristics, enabling the system to maintain flexibility across different defect types while operating automatically without human intervention.
Data Source
AI summary
A few-shot defect detection method based on metric learning, including: (S1) performing data enhancement on a to-be-detected few-shot defect data set through a G2-Generative adversarial network (G2-GAN); (S2) extracting features of a defect data set similar to the to-be-detected few-shot defect data set based on an adaptive convolution kernel-based convolutional neural network (SKM-CNN) to generate a pre-training model; and (S3) transferring the pre-training model to a few-shot defect detection network (S2D2N) based on metric learning; and performing target feature extraction and metric learning in sequence to realize rapid identification and location of defects.