Fusion-Based Classifier for Industrial Defect Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current neural network-based image recognition methods in industrial production fail to effectively reduce the miss rate of critical defects while increasing the precision rate for less impactful defects, leading to inefficiencies in quality control and increased labor costs.
Innovation Solution
A fusion-based classifier system that includes a sub-classifier, a fusion layer, and an output layer, which generates and selects probability vectors to prioritize classes of interest, using extremum values to improve classification accuracy and reduce error rates.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If current neural network methods focus on reducing overall error rate, then the overall classification accuracy is improved, but the miss rate of critical defects cannot be effectively reduced and the precision rate for less impactful defects is not improved
Solution Approach 1:
The patent segments the classification task by introducing a fusion layer that processes probability vectors for different defect classes separately. The fusion layer divides the overall classification problem into class-specific sub-problems, allowing different fusion strategies (max fusion for critical defects, mean fusion for less impactful defects) to be applied to different defect classes, thereby simultaneously improving miss rate reduction for critical defects and precision rate for less impactful defects.
Solution Approach 2:
The patent changes the parameter fusion strategy based on defect class importance. For critical defects, it uses max fusion (taking the maximum probability value across multiple neural networks) to reduce miss rates. For less impactful defects, it uses mean fusion (averaging probability values) to improve precision rates. This parameter change in the fusion operation allows simultaneous optimization of both contradictory objectives.
2Productivity
If traditional classification methods are used, then the system structure is simple, but the quality control efficiency is low and labor costs increase
Solution Approach 1:
The patent introduces a fusion layer as an intermediary component between multiple neural networks and the final output. This fusion layer acts as a mediator that combines probability vectors from multiple networks using different fusion strategies. The addition of this intermediary component increases system capability and quality control efficiency while maintaining manageable structural complexity through its modular design.
Solution Approach 2:
The patent merges multiple neural network outputs (probability vectors) into a unified classification decision through the fusion layer. By combining the strengths of multiple networks using different fusion operations (max and mean fusion), the system achieves higher quality control efficiency and accuracy without requiring complete redesign of the classification architecture.
3Measurement precision
If multiple probability vectors are generated and fused, then the classification accuracy for different defect types is improved, but the computational complexity increases
Solution Approach 1:
The patent applies partial action by using different fusion strategies for different defect classes rather than applying a uniform expensive operation to all classes. For critical defects requiring high sensitivity, it uses max fusion which is computationally efficient. For less impactful defects, it uses mean fusion which provides balanced accuracy. This selective approach improves overall classification accuracy while controlling computational energy consumption.
Data Source
AI summary
A fusion-based classifier, classification method, and classification system, wherein the classification method includes: generating a plurality of probability vectors according to input data, wherein each of the plurality of probability vectors includes a plurality of elements corresponding to a plurality of class respectively; selecting, from the plurality of probability vectors, a first probability vector having an extremum value corresponding to a first class-of-interest according to the first class-of-interest; and determining a class of the input data according to the first probability vector.