Source-linked AI summary
Ensemble of Deep Convolutional Neural Networks for Automatic Pavement Crack Detection and Measurement
Zhun Fan, Chong Li, Ying Chen, Paola Di Mascio, Xiaopeng Chen, Guijie Zhu, Giuseppe Loprencipe
TL;DR
Pavement crack detection and measurement are important for road safety, but conventional methods are time-consuming and inefficient. The paper proposes an ensemble CNN without pooling layers that fuses probability outputs and extracts crack morphology from skeletons. On the CFD and AigleRN databases, the method outperforms other methods and measures crack length and width across several crack types.
Problem
Conventional pavement crack detection and measurement methods are time-consuming and inefficient, despite the importance of these tasks for road safety.
Method
The method ensembles CNNs without pooling layers, averages their outputs into pixelwise crack probabilities, and uses segmentation and skeleton extraction for morphological measurement.
Results
The method outperforms other algorithms, achieving precision, recall, and F1 scores of 0.9552, 0.9521, and 0.9533 on CFD and 0.9302, 0.9166, and 0.9238 on AigleRN.
Takeaways & Limitations
Crack length and width can be measured with satisfactory accuracy for complex, common, thin, and intersecting cracks.
Takeaways & Limitations
The method cannot perform end-to-end crack detection and is limited to static images.
Abstract
from arXiv · showhide
Automated pavement crack detection and measurement are important road issues. Agencies have to guarantee the improvement of road safety. Conventional crack detection and measurement algorithms can be extremely time-consuming and low efficiency. Therefore, recently, innovative algorithms have received increased attention from researchers. In this paper, we propose an ensemble of convolutional neural networks (without a pooling layer) based on probability fusion for automated pavement crack detection and measurement. Specifically, an ensemble of convolutional neural networks was employed to identify the structure of small cracks with raw images. Secondly, outputs of the individual convolutional neural network model for the ensemble were averaged to produce the final crack probability value of each pixel, which can obtain a predicted probability map. Finally, the predicted morphological features of the cracks were measured by using the skeleton extraction algorithm. To validate the proposed method, some experiments were performed on two public crack databases (CFD and AigleRN) and the results of the different state-of-the-art methods were compared. The experimental results show that the proposed method outperforms the other methods. For crack measurement, the crack length and width can be measure based on different crack types (complex, common, thin, and intersecting cracks.). The results show that the proposed algorithm can be effectively applied for crack measurement.
1. Introduction
Pavement cracks threaten road safety, while conventional detection and measurement methods are time-consuming and inefficient. The paper introduces an ensemble CNN approach using probability fusion to detect cracks and support morphological measurement.
- Pavement cracks threaten road and highway safety, making automated detection and measurement relevant to road management.
- Related work: Earlier approaches used thresholding, edge detection, morphology, wavelets, minimal paths, and machine learning, but noise can prevent complete crack detection.
- Proposed method: The resulting crack skeleton supports measurement of crack width and length from predicted crack maps.
- Proposed method: The proposed ensemble network uses probability fusion and CNNs without pooling layers to improve crack detection, including small-crack structure recognition.
- Proposed method: Individual CNN outputs are averaged into a per-pixel crack probability map, followed by segmentation and medial-axis skeleton extraction.
- Evaluation: The paper evaluates the proposed system against other algorithms through experiments on public pavement-crack databases.
2. Methods
The method combines structured-prediction CNNs without pooling, probability-fused ensemble outputs, and morphological processing to detect pavement cracks and measure their geometry.
- Convolutional neural network: The individual CNN uses convolutional feature extraction, activation functions, and structured prediction to produce crack outputs from image patches.The architecture maps a 27 × 27 input patch to a 5 × 5 structured output, with sigmoid outputs for binary crack classification.
- Convolutional neural network: Pooling is discarded because downsampling can lose input information, while zero padding preserves feature-map spatial resolution.The CNN uses 3 × 3 kernels and zero padding during convolution.
- Ensemble network learning: The ensemble averages outputs from multiple CNN models to produce a final crack-probability prediction for each pixel.Ensemble learning combines multiple classifiers and is reported to improve performance over individual models.
- Crack segmentation and skeletonization: Binary crack images are segmented through hole filling, noise removal, and connected-crack labeling before skeletonization extracts morphological features.Closing fills small holes, opening removes noisy pixels, and medial-axis processing produces single-pixel-wide crack skeletons.
- Crack measurement: Crack length and average width are calculated from skeleton geometry, calibrated pixel displacements, and crack-element areas.The method assumes no geometric distortion and converts image measurements into physical crack dimensions using image resolution.
3. Experimental Results
Experiments evaluated the ensemble on CFD and AigleRN using precision, recall, and F1 across model counts and thresholds. On CFD, the ensemble achieved higher reported accuracy than other algorithms.
- Experimental setup: Experiments evaluated the ensemble on CFD and AigleRN using precision, recall, and F1 metrics.The evaluation also accepted a two-pixel distance between prediction images and ground truth when calculating TP, FP, and FN.
- Parameter analysis: The experiments varied ensemble size n = 1, 3, 5, 7 and decision threshold t = 0.4, 0.5, 0.6, 0.7.Figures 3 and 4 report these variations for AigleRN and CFD.
- Parameter analysis: AigleRN performed best at t = 0.4 with n = 3, whereas CFD performed best at t = 0.6 with n = 3.The project therefore used three ensemble models and thresholds of 0.4 for AigleRN and 0.6 for CFD.
- Comparison with other methods: On CFD, traditional Canny and local-threshold methods were negatively affected by noise, while CrackForest overestimated crack width.The structured prediction method produced wrong detections, and U-net showed missed detections.
- Comparison with other methods: Pr: 0.9552, Re: 0.9521, and F1: 0.9533 were reported for the ensemble networks on CFD.The paper states that these results were more satisfactory than those of other algorithms.
Pr Re F1
On AigleRN, the ensemble method outperformed comparison algorithms for crack detection and supported crack segmentation, skeleton extraction, and morphological measurement, with known measurement deviations.
- Detection performance: The ensemble detected more continuous cracks than comparison methods and extracted crack skeletons while maintaining good accuracy.Canny, local threshold, and FFA missed continuous cracks; MPS found them but did not extract skeletons, while structured prediction had missed detections.
- Detection performance: Pr:0.9302, Re:0.9166, and F1:0.9238 were reported for the proposed ensemble method on AigleRN.The authors report that the method outperformed other algorithms.
- Crack segmentation and skeleton: CFD results showed more accurate crack segmentation and skeletonization, while AigleRN results still captured complex, disconnected, and small cracks.AigleRN’s complex structures caused some deficiencies relative to the public database reference results.
- Crack measurements: The method tended to overestimate crack length when disconnected cracks were eliminated during detection.The authors attribute this discrepancy to missed disconnected cracks reducing the number of detected cracks.
- Crack measurements: Morphological opening and closing operations could produce predicted crack lengths above ground truth and larger mean widths by filling holes and neighboring pixels.These operations also eliminate isolated pixels.
4. Conclusions
The paper presents a probability-fusion ensemble of CNNs without pooling layers for automated pavement crack detection and measurement. It reports strong CFD and AigleRN results across detection metrics and crack types, while limiting the method to static images rather than end-to-end detection.
- 4. Conclusions: The method uses an ensemble of CNNs with probability fusion, discards pooling layers, and measures crack width and length from predicted crack maps.The approach targets automated pavement surface evaluation.
- 4. Conclusions: CFD achieved Pr 0.9552, Re 0.9521, and F1 0.9533, while AigleRN achieved Pr 0.9302, Re 0.9166, and F1 0.9238.The results were compared with existing databases and methods.
- 4. Conclusions: The algorithm measured the length and width of complex, common, thin, and intersecting cracks with satisfactory accuracy.The conclusion covers multiple crack types.
- 4. Conclusions: The proposed method cannot perform end-to-end crack detection and can only be used with static images.Future work targets end-to-end deep learning and crack detection in video streams.