CNN Crack Segmentation with Separable Convolutions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing computer vision methods for automated infrastructure assessment are susceptible to environmental changes and struggle with extracting sensitive features, particularly failing to accurately segment cracks in images with complex backgrounds, due to their reliance on classification or object detection approaches rather than pixel-level segmentation.
Innovation Solution
A computer-implemented method using a convolutional neural network (CNN) with separable convolution modules, atrous spatial pyramid pooling, and a decoder module, specifically designed for semantic damage detection, capable of segmenting cracks in images with complex backgrounds by processing images of varying sizes and resolutions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If computer vision algorithms are used for automated infrastructure assessment, then human visual inspections can be replenished, but the algorithms are highly susceptible to changes in environmental conditions and cannot extract sensitive features
Solution Approach 1:
The patent replaces traditional computer vision algorithms with a deep learning-based convolutional neural network (CNN) system. This substitution enables the system to automatically adapt to environmental variations and extract sensitive crack features through learned representations rather than hand-crafted features, resolving the reliability issue while maintaining automation.
Solution Approach 2:
The patent transforms the problem from traditional image classification to pixel-level segmentation, fundamentally changing the output parameter from class labels to per-pixel crack probability maps. This parameter change enables precise localization and quantification of cracks while making the system more robust to environmental conditions through the use of loss functions optimized for segmentation tasks.
2Ease of manufacture
If classification or object detection approaches are used, then simple crack detection is possible, but accurate segmentation of cracks in images with complex backgrounds fails
Solution Approach 1:
The patent divides the image into individual pixels and assigns a crack probability to each pixel independently through the CNN. This pixel-level segmentation approach allows precise delineation of crack boundaries even in complex backgrounds, overcoming the limitations of object detection methods that treat cracks as single bounding boxes.
Solution Approach 2:
The patent transitions from 2D spatial processing in traditional object detection to 3D feature space processing by incorporating depthwise separable convolutions and atrous spatial pyramid pooling. These operations extract features at multiple scales and depths, enabling the network to distinguish cracks from background textures by analyzing patterns across different dimensionalities.
3Measurement precision
If deep learning models with multiple convolution layers are used, then feature extraction capability improves, but computational resources and processing time increase
Solution Approach 1:
The patent replaces standard convolution operations with depthwise separable convolutions, which decompose the convolution into two separate operations: depthwise convolution (applying a single filter per input channel) and pointwise convolution (1x1 convolutions to combine features). This substitution reduces the number of parameters and computational operations while maintaining feature extraction capability.
Solution Approach 2:
The patent segments the convolutional processing into multiple independent depthwise convolution layers, each handling specific feature extraction tasks. This modular approach allows the network to process features at different depths independently, reducing redundant computations and lowering overall computational complexity compared to dense convolutional connections.
4Measurement precision
If standard convolution operations are used, then feature extraction is comprehensive, but processing speed and computational efficiency decrease
Solution Approach 1:
The patent substitutes standard convolutions with atrous (dilated) convolutions that introduce gaps between kernel elements. This substitution allows the network to capture spatial patterns at multiple scales simultaneously without increasing the number of parameters or computational operations, thereby maintaining feature extraction comprehensiveness while improving processing efficiency.
Solution Approach 2:
The patent employs atrous spatial pyramid pooling that applies convolutions with different dilation rates (1, 2, 4, 8) in a periodic manner across different pooling units. This periodic application of convolutions at varying scales enables comprehensive feature extraction at multiple resolutions, improving both processing speed and feature comprehensiveness through efficient multi-scale analysis.
Data Source
AI summary
A computer-implemented method of analyzing an image to segment an article of interest in the image comprises (i) receiving the image having a width of n1 pixels, a height of n2 pixels and a depth of d channels; (ii) processing the image using a machine learning algorithm configured to segment the article of interest, the machine learning algorithm comprising a convolutional neural network including: at least one convolution layer; after said at least one convolution layer, at least one separable convolution module comprising a series of separable convolutions, each separable convolution comprising a depthwise convolution and a pointwise convolution; after said at least one separable convolution module, a pooling module; and a decoder module after the pooling module; and (iii) displaying the image with location of the article of interest being indicated if determined to be present by the machine learning algorithm.