Defect Detection System Using Active Learning Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning models for defect detection require large amounts of labeled data, which is time-consuming and costly to obtain, especially when dealing with diverse and rare defect classes.
Innovation Solution
A multi-staged approach that uses a deep learning segmentation model to prioritize images for human labeling based on their contribution to the training of the prediction model, focusing on data sets with randomly distributed objects and varying morphology.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If manual labeling of all images is performed to obtain sufficient training data, then the prediction model can be trained effectively, but the time and cost required increase significantly
Solution Approach 1:
The patent segments the image dataset into multiple batches and uses a deep learning segmentation model to identify and prioritize specific images that are most informative for training. Instead of manually labeling all images, the system segments the work into labeled and unlabeled portions, selectively processing only the most valuable images through manual labeling while using automated labeling for the remainder.
Solution Approach 2:
The deep learning segmentation model automatically identifies and selects images that would be most beneficial to label, performing the preliminary work of curation without human intervention. The system serves itself by using its own predictions to guide the manual labeling process, automatically generating priority queues and selecting images that maximize training effectiveness.
2Loss of time
If automatic labeling is used to reduce manual work, then labeling time decreases, but the precision and accuracy of labels deteriorate
Solution Approach 1:
The patent segments the labeling task into two parts: high-priority images that require manual labeling for accuracy, and lower-priority images that can be automatically labeled. The segmentation model identifies which images fall into each category based on their informativeness and uncertainty, ensuring that manual effort is concentrated where it matters most while automated labeling handles the remainder.
Solution Approach 2:
The deep learning segmentation model acts as an intermediary between complete manual labeling and complete automatic labeling. It evaluates each image and determines the optimal labeling approach, serving as a mediator that directs which images go to manual labelers and which can be processed automatically, thereby optimizing the balance between accuracy and efficiency.
3Adaptability or versatility
If diverse defect classes with varying morphology are included in training data, then the prediction model becomes more versatile, but the complexity of data collection and labeling increases
Solution Approach 1:
The segmentation model automatically adapts to diverse defect classes by evaluating images based on their informativeness rather than requiring pre-categorization. The system self-adjusts to handle varying morphologies and defect types without increasing manual complexity, as the automated model identifies relevant features and prioritizes images across all defect categories uniformly.
Solution Approach 2:
The patent changes the parameter used for image selection from manual expert judgment to automated metrics such as uncertainty, informativeness, and prediction confidence. This parameter change allows the system to handle diverse defect classes efficiently by using quantitative measures that work across all morphologies and defect types, reducing the complexity of data management.
Data Source
AI summary
A computing system generates a training data set for training the prediction model to detect defects present in a target surface of a target specimen and training the prediction model to detect defects present in the target surface of the target specimen based on the training data set. The computing system generates the training data set by identifying a set of images for training the prediction model, the set of images comprising a first subset of images. A deep learning network generates a second subset of images for subsequent labelling based on the set of images comprising the first subset of images. The deep learning network generates a third subset of images for labelling based on the set of images comprising the first subset of images and the labeled second subset of images. The computing system continues the process until a threshold number of labeled images is generated.