Defect Detection via Target Convolutional Layer Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Supervised learning classification models for defect inspection in industrial products face challenges due to the difficulty in obtaining sufficient training data, leading to inaccurate defect identification in images.
Innovation Solution
A method utilizing a computing device that determines a target convolutional layer from a convolutional neural network, extracts features from test images using this layer, and employs a pre-trained Gaussian mixture model to accurately quantify and detect defects, with data enhancement algorithms to prevent poor generalization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If a supervised learning classification model is used for defect inspection, then the inspection process can be automated, but the accuracy of defect identification deteriorates due to insufficient training data
Solution Approach 1:
The patent applies preliminary action by pre-training the convolutional neural network on a large dataset of general image features before fine-tuning it for specific defect detection tasks. This pre-training ensures the model has already learned robust feature extraction capabilities, allowing it to achieve high accuracy even when training data for the specific defect type is limited. The model is then adapted to the specific inspection task with minimal additional training data.
2Measurement precision
If more training data is collected to improve model accuracy, then defect identification accuracy improves, but the time and resources required for data collection and model training increase
Solution Approach 1:
The patent changes the parameter of training data requirements by using data augmentation techniques that generate synthetic training samples through transformations such as rotation, scaling, and flipping of existing defect images. This approach effectively increases the training dataset size without requiring additional physical data collection, thereby improving model accuracy while minimizing time loss. The system also optimizes training parameters such as learning rate and batch size to accelerate convergence.
3Measurement precision
If a complex convolutional neural network is used to extract features, then feature extraction accuracy improves, but the computational complexity and training time increase
Solution Approach 1:
The patent applies segmentation by dividing the convolutional neural network into modular components, where a pre-trained backbone network handles general feature extraction and a smaller, task-specific head network handles defect classification. This segmentation allows the system to maintain high feature extraction accuracy while reducing the overall computational complexity and training time, as the complex feature extraction is done once during pre-training and the lightweight head network trains quickly on the specific defect data.
Data Source
AI summary
A method for detecting data defects and a computing device applying the method obtains a test image for analysis. A field to which the test image relates is determined. Based on the field, a target convolutional layer is determined from a convolutional neural network. The target convolutional layer is used to extract features of the test image. A target score of the test image and a score threshold corresponding to the field are determined. If the target score is less than the score threshold, it is determined that the test image reveals defects, thereby improving an accuracy of defect detection.