TL;DR
This paper introduces a novel system for automatically detecting and quantifying cracks in concrete structures using binocular video and an improved deep learning model. By combining a lightweight classification network with precise segmentation and 3D depth information, the method achieves high accuracy in measuring crack length, width, and orientation, paving the way for more robust structural health monitoring.
What Problem Does This Paper Address?
Assessing the structural integrity of concrete infrastructure like bridges and buildings is crucial, and cracks are a primary indicator of damage. Traditional visual inspection methods are subjective and time-consuming. While automated methods exist, they often struggle with real-time performance, accuracy, and comprehensive quantification of crack features. Existing approaches may oversimplify crack geometry, leading to inaccurate measurements, especially for cracks with varying widths or complex patterns. This paper aims to bridge this gap by providing a precise, automated, and real-time system for detecting and quantifying cracks, including their length, maximum width, and orientation.
Methods
The proposed system is built upon four key modules:
- Lightweight Classification Algorithm: An enhanced EfficientNetV2 model is used for initial crack detection in video frames. The novelty here lies in modifying the MBConv module by incorporating channel attention (CAM) and spatial attention (SAM) mechanisms. This aims to improve detection accuracy without significantly increasing computational cost.
- High-Precision Segmentation Algorithm: A U-Net semantic segmentation model is employed to accurately delineate crack regions from the background at a pixel level. This is crucial for detailed geometric analysis.
- Semi-Global Block Matching (SGBM): This algorithm is used to compute depth information from the binocular camera's stereo images. By analyzing parallax between the left and right camera views, it enables the conversion of pixel measurements to real-world dimensions.
- Crack Quantification Technique: Geometric parameters like crack length, maximum width, and orientation angle are extracted from the segmented crack masks and depth information.
- Crack Length: A skeleton line-based algorithm is used, where the length is calculated by summing pixel lengths along the crack's central line.
- Maximum Crack Width: A plumb line method based on the median axis is proposed, which is claimed to be more accurate than traditional edge-based minimum distance methods, especially for cracks with abrupt width changes.
- Crack Angle: The angle is determined by the line connecting the crack's initiation and termination points relative to the horizontal.
Datasets
The paper utilizes the Concrete Crack Images for Classification dataset from Mendeley Data for training and testing the crack classification model. This dataset contains 20,000 images of cracks and non-cracks. For segmentation, a dataset of 777 crack images was collected. For the overall quantification tests, 20 crack images with varying widths were collected from real-world environments (indoor/outdoor concrete, walls, roads).
Key Insights and Ideas
- Attention Mechanisms Boost Accuracy: Integrating channel and spatial attention into the EfficientNetV2's MBConv module significantly improves crack detection accuracy (1.6% increase) without a substantial performance penalty. This highlights the power of attention for fine-grained feature extraction in crack detection.
- Binocular Vision for 3D Crack Geometry: Leveraging binocular vision and SGBM for depth estimation is key to converting 2D pixel measurements into accurate 3D geometric properties of cracks, a crucial step for meaningful structural assessment.
- Novel Width Quantification Method: The proposed plumb line method for maximum crack width quantification offers a more robust solution compared to traditional edge-based methods, particularly for complex crack shapes.
- Comprehensive Quantification is Achievable: The system demonstrates the feasibility of simultaneously quantifying crack length, maximum width, and orientation angle with high precision, offering a more complete picture of structural damage.
Key Results
- The improved EfficientNetV2 achieved a 1.6% higher detection accuracy compared to the original model.
- The proposed quantification method achieved low quantification errors: 2% for crack length, 4.5% for maximum width, and 1% for crack angle of orientation.
- The pixel-to-actual distance conversion method showed absolute errors of no more than 0.2 pixels.
- The maximum crack width quantification method based on the median plumb line had an average absolute error of 0.6 pixels, significantly lower than the traditional minimum distance method's average error of 4.4 pixels.
- The overall system demonstrated high robustness and reliability, with average relative errors for crack length, width, and orientation angle not exceeding 2%, 4.5%, and 1%, respectively.
Limitations and Future Work
- Single Crack Focus: The current quantization module is designed for single cracks and cannot effectively quantify complex multi-branched cracks (e.g., Y-shaped, X-shaped, mesh cracks).
- Real-time Performance: While improved, the method's real-time performance is still limited by the computational load of segmentation and quantization, with a processing speed of approximately 0.5 seconds per frame for a 640x480 image.
- Future Work: The authors suggest future work will focus on identifying and quantifying multi-crack regions and further optimizing the algorithm for real-time applications. They also note that applying the method to specific environments and training with relevant data can further improve detection accuracy.
Summary updated