Fast Attention CNN for Fine-Grained Crack Segmentation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Materials Science

Background:

• Deep learning excels at image segmentation for pixel-level crack detection.
• Accurate detection of fine-grained cracks and elimination of crack-like noise are critical challenges.
• Existing methods struggle with subtle cracks and noise like grooving.

Purpose of the Study:

• To develop a fast encoder-decoder network with scaling attention for enhanced crack detection.
• To improve the accuracy of detecting tiny cracks and suppress irrelevant crack-like noise.
• To reduce computational complexity while maintaining high detection performance.

Main Methods:

• Proposed a fast encoder-decoder network incorporating an Atrous Spatial Pyramid Pooling (ASPP) layer.
• Introduced a novel scaling attention mechanism (AG+) to suppress non-semantic regions and noise.
• Generated a comprehensive crack dataset with 11,226 image-mask pairs for training and evaluation.

Main Results:

• The model demonstrated improved detection accuracy for fine-grained cracks.
• The novel AG+ attention mechanism effectively reduced crack-like noise.
• Achieved performance close to state-of-the-art models with a mean Dice coefficient (mDice) difference of only 1.2%.
• Reduced computational complexity by achieving two times fewer FLOPs (floating-point operations).

Conclusions:

• The proposed deep learning model offers a computationally efficient and accurate solution for pixel-level crack detection.
• The integration of ASPP and scaling attention (AG+) significantly enhances the ability to detect fine cracks and filter noise.
• This approach provides a promising advancement for structural health monitoring and material defect analysis.