Dimensionality Reduction for Defect Detection in Product Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Convolutional neural networks face challenges in accurately detecting defect signals in image data due to similar original image data leading to similar extracted features, which reduces detection accuracy.
Innovation Solution
A defect detecting method that includes an electronic device with modules for obtaining, classifying, dimension reduction, and scoring product images using various algorithms like PCA, Isomap, and Gaussian mixture models to enhance feature extraction and classification accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If convolutional neural network performs feature extraction on original image data, then automation of defect detection is achieved, but detection accuracy deteriorates when original image data is similar
Solution Approach 1:
The patent applies dimensionality reduction techniques (PCA, Isomap, LLE, etc.) to transform the extracted features from their original high-dimensional space into a lower-dimensional space. This dimensional transformation enhances the separability of similar defect patterns by reorganizing the feature space, thereby improving detection accuracy while maintaining automation.
Solution Approach 2:
The patent changes the parameters of the feature representation by applying various dimensionality reduction algorithms with different parameters (number of components, neighborhood size, etc.). By optimizing these parameters, the system transforms the feature characteristics to improve distinguishability between similar defects while preserving the automated detection capability.
2Measurement precision
If multiple dimensionality reduction algorithms are applied to product images, then feature extraction accuracy is improved, but processing time increases
Solution Approach 1:
The patent applies multiple dimensionality reduction algorithms but selects only the top-performing ones based on evaluation metrics. This partial application approach ensures sufficient feature extraction accuracy is achieved without unnecessarily applying all possible algorithms, thereby balancing processing time with detection performance.
Solution Approach 2:
The system performs preliminary evaluation of different dimensionality reduction algorithms on training data before actual defect detection. This preliminary action identifies the most effective algorithms for specific defect types, so that during actual detection, only the pre-selected optimal algorithms are applied, reducing processing time while maintaining accuracy.
Data Source
AI summary
A method for detecting product for defects implemented in an electronic device includes classifying a plurality of product images to be detected into linear images or non-linear images; performing dimension reduction processing on the product images after image classification according to a plurality of dimension reduction algorithms to obtain a plurality of dimension reduction data; determining an optimal dimension reduction data of the plurality of dimension reduction data; obtaining score data of the product image by inputting the optimal dimension reduction data into a Gaussian mixture model; comparing the score data with a threshold; determining whether the score data is less than the threshold; and determining that there is at least one defect in the product image in response that the score data is determined to be less than the threshold.