Image Feature Classification for Low-Sample Defect Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge in camera module production is the low accuracy and efficiency of deep learning-based target detection due to insufficient samples for model training, particularly for defects like solder overflow and pin bending, which are difficult to obtain given the routine yield control and production line optimization.
Innovation Solution
A data detection method that involves obtaining image samples, extracting sample features, determining a target sample feature with the maximum nearest-neighbor feature distance, and training an initial feature classification model based on (N+1) sample feature categories to improve detection accuracy and efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning-based target detection is used, then detection capability is improved, but sample sufficiency deteriorates
Solution Approach 1:
The patent generates synthetic defect samples by copying and transforming normal connector images through artificial defect introduction. This creates virtual defective samples without requiring physical defective products, thus resolving the contradiction between needing sufficient training samples and the scarcity of actual defect occurrences in optimized production lines.
Solution Approach 2:
The patent transforms normal images into defective samples by changing parameters such as introducing solder overflow, pin bending, or pin missing conditions through image processing. This parameter transformation allows generation of diverse defect samples from limited normal samples, improving detection capability without requiring abundant physical defect samples.
2Reliability
If more samples are collected for model training, then model robustness is improved, but time consumption increases
Solution Approach 1:
The patent performs preliminary action by pre-generating a comprehensive dataset of synthetic defective samples before actual detection tasks. This pre-generated dataset can be stored and reused for multiple training iterations, avoiding repeated time-consuming data collection and processing while maintaining model robustness through diverse synthetic samples.
Solution Approach 2:
By copying normal samples and transforming them into various defect types through systematic image processing, the patent creates large volumes of training data quickly without the time cost of collecting actual defective samples from production lines, thus improving model robustness efficiently.
3Adaptability or versatility
If traditional defect grouping method is used, then detection coverage is improved, but data collection complexity increases
Solution Approach 1:
Instead of the traditional approach of collecting different defect types separately and grouping them, the patent inverts the process by starting with normal samples and systematically transforming them into various defect types through controlled image manipulation. This simplifies data collection by requiring only normal samples while achieving comprehensive defect coverage through transformation.
Solution Approach 2:
The patent makes normal connector images serve multiple functions by transforming them into various defect types (solder overflow, pin bending, pin missing, etc.). This multi-functionality approach allows a single set of normal samples to generate comprehensive training data for multiple defect categories, reducing data collection complexity while maintaining broad detection coverage.
Data Source
AI summary
A data detection method includes obtaining an image sample, extracting sample features from the image sample, determining a target sample feature with a maximum nearest-neighbor feature distance, determining, based on the nearest-neighbor feature distances of the sample features, (N+1) sample feature categories corresponding to the sample features, N being a positive integer, and training an initial feature classification model based on the (N+1) sample feature categories and the sample features to obtain a trained feature classification model. The nearest-neighbor feature distance of one sample feature is a minimum value of feature distances between the one sample feature and other ones of the plurality of sample features except the one sample feature. The (N+1) sample feature categories include N first feature categories indicating detected data being normal data and a second feature category indicating to perform secondary detection on the detected data.