CNN Boundary Feature Suppression in Industrial Inspection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In image processing for factory automation, the feature extraction of boundaries between inspected and uninspected pixel regions in convolutional neural networks can lead to false detection, similar to foreign matter features, due to the influence of boundary features during convolution and pooling processes.
Innovation Solution
An image processing device and method that perform a first masking process on an inspection image, followed by an intermediate process using a convolutional neural network, and then a second masking process on both the inspection and mask images to suppress the extraction of boundary features between inspected and uninspected regions, with the second masking process being performed at a stage with minimal intermediate processes to reduce feature propagation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a mask image is applied to mask the pixel region not to be inspected, then the foreign matter feature extraction is improved, but the boundary feature between inspected and uninspected regions causes false detection
Solution Approach 1:
The patent divides the masking process into two separate stages: first masking the original inspection image before CNN processing, and second masking the feature map after CNN processing. This segmentation of the masking operation prevents boundary features from being extracted and propagated throughout the network, thereby eliminating false detection while maintaining accurate foreign matter feature extraction.
Solution Approach 2:
The patent performs the first masking process as a preliminary action before the inspection image is input to the CNN. By masking the pixel regions not to be inspected in advance, the boundary features between inspected and uninspected regions are prevented from being extracted in the first place, thus avoiding false detection caused by these boundary features during subsequent convolution and pooling operations.
2Object-affected harmful factors
If the mask image is processed through the same intermediate process, then the boundary feature suppression is improved, but the computational complexity increases
Solution Approach 1:
The patent creates a copy of the mask image and processes this copy through the same intermediate processes (convolution and pooling) as the inspection image. This copying approach allows the boundary features to be suppressed in the feature map without requiring modification of the original mask or inspection image processing pipeline, thus adding minimal computational complexity while effectively preventing false detection.
Data Source
AI summary
A convolutional neural network performs: a first masking process of masking a pixel region not to be inspected, by computing pixel values of corresponding pixels of an inspection image and of a mask image; an intermediate process for extracting a feature image from the inspection image that has been subjected to the first masking process; and a second masking process of masking the pixel region not to be inspected, by computing the pixel values of corresponding pixels of the inspection image that has been subjected to the intermediate process and of the mask image that has been subjected to a process identical to the intermediate process.