Image Segmentation Defect Detection With Adaptive Mask Sizing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep learning-based defect detection techniques are difficult to train and unreliable due to the need for large amounts of training data and high nuisance detection, as they often label only one feature per image, leading to inaccurate defect prediction and inefficient defect detection processes.
Innovation Solution
A deep learning-based image segmentation model that identifies multiple features of interest per image, using user-labeled training data to generate segmentation masks, adjusts mask sizes based on feature size, and determines difference-based attributes to classify defects based on predetermined thresholds, reducing the need for extensive training data and improving detection accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If existing deep learning-based defect detection techniques are used, then defect detection can be automated, but the model becomes difficult to train and unreliable due to high nuisance detection
Solution Approach 1:
The patent applies segmentation by dividing the defect detection task into multiple components: feature detection, attribute determination, and defect classification. The system segments the image analysis process to first identify potential features, then evaluate their attributes (size, shape, contrast), and finally classify them as defects or nuisance, thereby improving reliability through staged processing
Solution Approach 2:
The patent utilizes parameter changes by adjusting detection parameters dynamically based on the analyzed features. The system modifies detection thresholds and criteria based on feature attributes such as size, shape, and contrast, allowing the automated detection to adapt to different defect types and reduce nuisance detection, thus improving reliability
2Device complexity
If one feature per image is labeled for training, then training data requirements are reduced, but the amount of training data needed increases significantly
Solution Approach 1:
The patent applies universality by creating a training framework that can handle multiple features per image simultaneously. The system is designed to process and train on images with multiple labeled features, making the training process more efficient and reducing the total quantity of training data needed while maintaining high detection accuracy
Solution Approach 2:
The patent merges multiple feature detection and attribute analysis into a unified training approach. By combining the processing of multiple features within single image contexts, the system achieves more effective use of training data, reducing the overall quantity needed while improving model performance
3Measurement precision
If multiple features per image are analyzed, then detection accuracy improves, but the time required for training increases
Solution Approach 1:
The patent applies preliminary action by pre-processing images to identify and segment potential features before the main training process. This preliminary feature extraction and attribute analysis prepares the data in advance, allowing the training to proceed more efficiently with multiple features per image, thereby reducing training time while maintaining high accuracy
Solution Approach 2:
The patent ensures continuity of useful action by implementing a streamlined processing pipeline that continuously refines feature detection and attribute analysis during training. The system maintains consistent and efficient processing of multiple features throughout the training process, preventing time delays while preserving detection accuracy
4Manufacturing precision
If segmentation masks are adjusted to match feature size, then defect detection precision improves, but the processing complexity increases
Solution Approach 1:
The patent applies dynamics by implementing adaptive segmentation mask adjustment that dynamically responds to detected feature characteristics. The system automatically modifies mask sizes and shapes based on the analyzed features, enabling high precision defect detection while managing processing complexity through intelligent, feature-driven adaptation rather than rigid fixed parameters
Data Source
AI summary
A system may be configured to receive training images of features on the sample. The system may be configured to generate a segmentation model based on the training images. The system may be configured to receive product images of product features and identify, using the segmentation model, features of interest based on the product images. The system may be configured to generate a segmentation model image including segmentation masks corresponding to the identified features of interest and adjust a size of the segmentation masks based on a size of the identified features of interest. The system may be configured to determine a difference-based attribute value for each identified feature of interest based on the adjusted size of the segmentation masks. The system may be configured to determine whether each of the identified features of interest are defects based on the difference-based attribute value and a predetermined threshold.