Defect Detection via Multi-Level IoU Segmentation and 2D-3D Mask Fusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional defect detection in industrial manufacturing is inefficient due to reliance on manual observation, especially for small defects, leading to high costs and low detection efficiency, and existing AI methods struggle with imbalanced sample ratios and 2D-only defect detection, resulting in underkill or overkill.
Innovation Solution
A method and system using a multi-level feature extraction instance segmentation network with increased IoU thresholds and fusion of 2D and 3D defect masks to improve detection accuracy, reducing overkill and underkill probabilities by aligning and filling corresponding masks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If manual observation is used for defect detection, then detection flexibility is maintained, but detection efficiency is low and detection cost is high
Solution Approach 1:
The patent replaces manual mechanical observation with an automated optical detection system using cameras and image processing algorithms. The system captures images of products and uses computer vision technology to automatically identify defects, eliminating the need for manual inspection while significantly improving detection efficiency and consistency.
Solution Approach 2:
The detection system performs self-analysis through automated image processing and defect recognition algorithms. The system independently processes images, identifies defects based on predefined criteria, and generates detection results without requiring continuous human intervention, enabling the system to serve itself in the detection process.
2Measurement precision
If 2D-only defect detection is used, then system complexity is reduced, but detection accuracy decreases due to inability to detect depth-related defects
Solution Approach 1:
The patent transitions from 2D image-based detection to 3D detection by incorporating depth information through stereo vision or time-of-flight cameras. The system captures both 2D images and 3D depth data, then fuses these multi-dimensional information sources to achieve accurate detection of surface defects, including those with depth variations that cannot be detected by 2D methods alone.
Solution Approach 2:
The patent merges 2D image data with 3D depth information to create a comprehensive defect detection system. By combining multiple data dimensions and processing them through fusion algorithms, the system achieves superior detection accuracy that leverages the strengths of both 2D and 3D approaches while maintaining operational efficiency.
3Measurement precision
If traditional instance segmentation network is used, then model simplicity is maintained, but detection accuracy decreases due to imbalanced positive and negative samples
Solution Approach 1:
The patent implements multi-level feature extraction that segments the defect detection process into distinct hierarchical levels. The network extracts features at multiple scales and resolutions, allowing it to effectively handle imbalanced samples by analyzing defects at different granularities. This multi-level segmentation approach enables accurate identification of both small and large defects while maintaining robustness against sample imbalance.
Solution Approach 2:
The patent applies local quality enhancement by using different IoU thresholds at different levels of the feature extraction hierarchy. Each level of the network is optimized with appropriate threshold values that match the characteristics of defects at that scale, allowing the model to adapt locally to varying defect sizes and densities, thereby improving overall segmentation accuracy despite imbalanced training samples.
4Quantity of substance
If low IoU threshold is used for positive sample sampling, then more positive samples are obtained, but feature extraction precision decreases
Solution Approach 1:
The patent divides the feature extraction process into multiple hierarchical levels, each with its own IoU threshold settings. Lower levels use lower thresholds to capture more potential positive samples, while higher levels use higher thresholds to ensure precision. This segmented approach allows the system to benefit from both high recall at lower levels and high precision at higher levels, resolving the trade-off between quantity and quality of samples.
Solution Approach 2:
The patent implements dynamic threshold adjustment where IoU thresholds vary across different levels of the network and can be adapted based on the specific detection task. This dynamic approach allows the system to optimize the balance between sample quantity and precision for each level, rather than using a fixed threshold throughout, thereby achieving both adequate sample coverage and high feature extraction accuracy.
Data Source
AI summary
The present application provides a method and system for defect detection. The method includes: acquiring a two-dimensional (2D) picture of an object to be detected; inputting the acquired 2D picture to a trained defect segmentation model to obtain a segmented 2D defect mask, where the defect segmentation model is trained based on a multi-level feature extraction instance segmentation network with intersection over union (IoU) thresholds being increased level by level, and the 2D defect mask includes information about a defect type, a defect size, and a defect location of a segmented defect region; and determining the segmented 2D defect mask based on a predefined defect rule to output a defect detection result.