Student-Teacher Neural Network Defect Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In manufacturing processes, defect detection by machine learning-based systems is challenging due to the rarity of defective samples, making it difficult to assemble a labeled training set for supervised training, and it is often more advantageous to reserve defective samples for verification rather than training.
Innovation Solution
A method involving a student-teacher neural network approach, where a student neural network is trained with a cost function that rewards similarity with a teacher neural network for normal images and dissimilarity for defective images, using multiple neural networks for different image processing tasks such as classification, reconstruction, and super-resolution, to effectively detect defects in manufacturing images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If supervised training with labeled defective samples is used, then detection accuracy is improved, but the scarcity of defective samples limits the training set size
Solution Approach 1:
The patent uses teacher neural networks to generate synthetic defective samples by copying and transforming features from normal samples. The teacher networks learn from scarce real defective samples and generate artificial defective images that mimic real defect patterns, effectively multiplying the training data available for student networks.
Solution Approach 2:
The teacher neural networks are trained in advance on the limited real defective samples to learn defect characteristics. This preliminary training enables the teacher networks to subsequently generate synthetic defective samples that can be used for training student networks, preparing the system before actual defect detection is needed.
2Quantity of substance
If defective samples are used for training, then training data availability is improved, but verification accuracy deteriorates due to sample exhaustion
Solution Approach 1:
The patent divides the neural network system into teacher networks and student networks with distinct roles. Teacher networks are trained on limited real defective samples for verification, while student networks are trained on synthetic defective samples generated by teachers for detection tasks, segmenting the data usage to preserve real samples for verification.
Solution Approach 2:
Student networks use copied synthetic defective samples generated by teacher networks for training, rather than directly using real defective samples. This copying mechanism allows extensive training data availability for students while preserving the original real defective samples for teacher network verification tasks.
3Adaptability or versatility
If multiple neural networks are deployed for different tasks, then detection versatility is improved, but system complexity increases
Solution Approach 1:
The patent merges multiple neural network functions into a unified teacher-student framework where teacher networks perform both verification and synthetic sample generation, while student networks handle detection tasks. This combining of functions reduces overall system complexity compared to having separate networks for each task.
Solution Approach 2:
Teacher neural networks are designed with multi-functionality, serving both as verification models and as generators of synthetic training samples. This universal design allows a single network component to fulfill multiple roles, reducing the total number of networks needed and simplifying the system architecture.
Data Source
AI summary
A system and method for defect detection. The method may include training, with a first set of images, a first neural network including a first student neural network, and a first teacher neural network. The training of the first neural network may include introducing defects into a first subset of the first set of images, and training the first student neural network with the first set of images. The training of the first student neural network may include using a first cost function, that: for an image of the first set and not of the first subset, rewards similarity between a feature map of the first student neural network and a feature map of the first teacher neural network, and for an image of the first subset, rewards dissimilarity between a feature map of the first student neural network and a feature map of the first teacher neural network.