Source-linked AI summary
Automatic Crack Detection on Road Pavements Using Encoder Decoder Architecture
Zhun Fan, Chong Li, Ying Chen, Jiahong Wei, Giuseppe Loprencipe, Xiaopeng Chen, Paola Di Mascio
TL;DR
Pavement crack detection supports pavement management because cracks can reduce structural capacity and threaten road safety. U-HDN addresses this task with an end-to-end encoder-decoder network combining multi-dilation and hierarchical feature learning. It reports strong results on public databases, including precision 0.945, recall 0.936, and F1 0.939 on CFD, while remaining limited by computational cost and inability to detect crack widths.
Problem
Pavement cracks can reduce load-bearing capacity and threaten road safety, making automatic crack detection important for pavement management.
Method
U-HDN is an end-to-end U-net-based encoder-decoder that fuses multiple dilation contexts with hierarchical high- and low-level feature maps for pixel-wise crack detection.
Results
U-HDN reports precision 0.945, recall 0.936, and F1 0.939 on CFD, and superior ODS and OIS performance compared with other algorithms.
Takeaways & Limitations
The reported results support U-HDN as a high-performing automatic pavement-crack detection method across the evaluated public databases.
Takeaways & Limitations
The method does not detect crack widths, and its redundant feature maps increase computational cost and reduce efficiency.
Abstract
from arXiv · showhide
Inspired by the development of deep learning in computer vision and object detection, the proposed algorithm considers an encoder-decoder architecture with hierarchical feature learning and dilated convolution, named U-Hierarchical Dilated Network (U-HDN), to perform crack detection in an end-to-end method. Crack characteristics with multiple context information are automatically able to learn and perform end-to-end crack detection. Then, a multi-dilation module embedded in an encoder-decoder architecture is proposed. The crack features of multiple context sizes can be integrated into the multi-dilation module by dilation convolution with different dilatation rates, which can obtain much more cracks information. Finally, the hierarchical feature learning module is designed to obtain a multi-scale features from the high to low-level convolutional layers, which are integrated to predict pixel-wise crack detection. Some experiments on public crack databases using 118 images were performed and the results were compared with those obtained with other methods on the same images. The results show that the proposed U-HDN method achieves high performance because it can extract and fuse different context sizes and different levels of feature maps than other algorithms.
1. Introduction
The paper presents U-HDN, an encoder-decoder crack-detection method that combines multi-dilation and hierarchical feature learning to fuse crack information across context sizes and feature levels. It targets automatic, efficient pavement-crack analysis for pavement management.
- 1.1. Motivation: Automatic crack detection is important for pavement management because cracks can reduce pavement load-bearing capacity, increase surface unevenness, and threaten road safety.The paper contrasts automatic surveys with visual inspection as more efficient, faster, and lower-cost.
- 1. Introduction: Existing visual crack-detection algorithms include traditional image-processing methods and artificial-intelligence approaches, with earlier methods relying heavily on manually designed feature descriptors.The introduction discusses wavelet transforms, thresholding, and feature-based methods before motivating deep learning.
- 1. Introduction: U-HDN combines U-net, a multi-dilation module, and hierarchical feature learning to perform end-to-end, pixel-wise pavement crack detection.The architecture fuses multi-scale and multi-context crack features for prediction.
- 1. Introduction: The proposed MDM uses dilation convolution to integrate crack features from multiple context sizes within an encoder-decoder architecture.The HF module complements this by extracting high- to low-level features for side-output prediction, whose results are concatenated and deeply supervised.
- 1.4. Contribution: The paper reports U-HDN as an automatic road-crack detection method designed for end-to-end training, hierarchical feature learning, and multi-context feature integration.The contribution list also describes modifications to the U-net architecture for end-to-end operation.
2. Methods
The method uses a modified U-net encoder-decoder with multi-dilation and hierarchical feature-learning modules. Dilated convolutions capture multiple context sizes, while side outputs and fused features support pixel-wise crack classification.
- 2. Methods: The U-HDN backbone is a modified U-net with contracting encoder and expansive decoder paths trained end to end for road-crack detection.The modification removes selected U-net components and uses zero-padding during convolution and up-convolution.
- 2.2. Multi-Dilation Module (MDM): Dilated convolution expands the effective filter context by inserting r−1 zeros between filter values without increasing the amount of calculation through larger filters.The dilation rate controls the sampling stride and context size; standard convolution corresponds to r=1.
- 2.2. Multi-Dilation Module (MDM): The MDM addresses standard convolution’s single-context limitation for thin, simple cracks and wide, complex cracks in images with varied topologies and widths.Its multiple context features are fused to better represent different crack structures.
- 2.2. Multi-Dilation Module (MDM): The multi-dilation module uses dilation rates {2, 4, 8, 16}, concatenates five crack-feature maps, and applies a 1 × 1 convolution to produce 1024 output features.The module is designed to improve detection across various crack types by fusing different context sizes.
- 2. Methods: Hierarchical feature learning predicts crack maps from multiple feature levels, while concatenated side outputs form a final fused prediction supervised by deeply supervised networks.The pipeline uses sigmoid classification to distinguish crack from non-crack pixels.
3. Experiments and Results
Experiments evaluated U-HDN on the CFD and AigleRN public databases using precision, recall, F1, ODS, and OIS, including dilation-rate analysis and comparisons with existing methods. U-HDN achieved superior reported performance while supporting end-to-end training and fusing multi-context and multi-level features.
- Evaluate Metrics: Evaluation used precision, recall, F1, ODS, and OIS, with a 0.5 decision threshold and acceptance of transitional regions within 2 pixels.F1 was used as the harmonic average of precision and recall; ODS used a fixed threshold and OIS used the best threshold per image.
- Discussion for Multi-Dilation Module (MDM): The {2, 4, 8, 16} dilation-rate group produced the highest accuracy on both CFD and AigleRN.The authors attribute this result to larger dilation rates providing more crack context for relatively wide or thin crack structures.
- Experimental Results on CFD: On CFD, Canny and local thresholding were sensitive to noise, while CrackForest over-measured cracks and achieved recall 0.9514.Structured prediction and U-net detected wrong non-crack pixels, whereas the ensemble network showed missed detections despite high reported metrics.
- Experimental Results: U-HDN achieved superior performance on CFD and AigleRN compared with other algorithms, reporting Pr: 0.945, Re: 0.936, F1: 0.939 on CFD and Pr: 0.921, Re: 0.931, F1: 0.924 on AigleRN.The paper also states that U-HDN outperformed other algorithms on both ODS and OIS.
- Experimental Results: U-HDN performs end-to-end training and fuses different context sizes through MDM with different feature levels through hierarchical feature learning.These design choices are presented as the main reasons for its reported comparative performance.
- Experimental Results on AigleRN: On AigleRN, traditional methods and several learned baselines struggled to extract continuous crack skeletons, while U-HDN was reported to achieve superior comparative performance.FFA detected thicker cracks but did not extract the crack skeleton and was associated with lower precision.
4. Conclusions
U-HDN combines multi-dilation and hierarchical-feature modules within U-net to support pavement crack detection, while the authors identify computational cost and efficiency as limitations for future work.
- U-HDN combines multi-dilation and hierarchical-feature modules in U-net to obtain multi-context and multi-scale features for pixel-wise crack detection.The MDM extracts feature maps at different context sizes through different dilation rates, while the HF module integrates high- and low-level maps at side outputs.
- The neural network contains redundant feature maps that increase computational cost and reduce efficiency.The authors state that these issues will be addressed in future work.
- Future work will explore channel pruning and automatic neural-network design to improve computational efficiency and accuracy.
- Future directions include video-streaming crack detection, cement-concrete crack detection, waterproofing evaluation, and integrated crack-detection systems.