Source-linked AI summary
Road Crack Detection Using Deep Convolutional Neural Network and Adaptive Thresholding
Rui Fan, Mohammud Junaid Bocus, Yilong Zhu, Jianhao Jiao, Li Wang, Fulong Ma, Shanshan Cheng, Ming Liu
TL;DR
Road-crack inspection is important but manual detection is time-consuming, dangerous, labor-intensive, and subjective. The paper combines CNN-based image classification with bilateral filtering and adaptive thresholding for crack segmentation, achieving 99.92% image-classification precision and around 98.70% pixel-level segmentation accuracy, while noisy color images remain challenging.
Problem
Road cracks can reduce road performance and pose safety hazards, while traditional manual detection is time-consuming, dangerous, labor-intensive, and subjective.
Method
The method classifies images with a deep convolutional neural network, filters positive images bilaterally, downsamples them, and computes an adaptive threshold by minimizing within-cluster sum squares.
Results
99.92% image-classification precision and around 98.70% pixel-level segmentation accuracy were reported.
Takeaways & Limitations
The approach successfully classifies crack-containing images and extracts cracks from positive images using adaptive thresholding.
Takeaways & Limitations
Color images with many noisy pixels cannot always be properly segmented by the proposed algorithm.
Abstract
from arXiv · showhide
Crack is one of the most common road distresses which may pose road safety hazards. Generally, crack detection is performed by either certified inspectors or structural engineers. This task is, however, time-consuming, subjective and labor-intensive. In this paper, we propose a novel road crack detection algorithm based on deep learning and adaptive image segmentation. Firstly, a deep convolutional neural network is trained to determine whether an image contains cracks or not. The images containing cracks are then smoothed using bilateral filtering, which greatly minimizes the number of noisy pixels. Finally, we utilize an adaptive thresholding method to extract the cracks from road surface. The experimental results illustrate that our network can classify images with an accuracy of 99.92%, and the cracks can be successfully extracted from the images using our proposed thresholding algorithm.
I. INTRODUCTION
Road cracks can reduce road performance and create safety hazards, while manual detection is time-consuming, labor-intensive, dangerous, and subjective. The paper builds on deep neural networks and automated image analysis to classify crack-containing images and extract cracks.
- Road cracks may reduce road performance and pose potential road safety hazards.
- Traditional manual crack detection is time-consuming, dangerous, labor-intensive, and subjective.
- Automated crack detection can provide fast and reliable road-surface analysis for maintenance planning.
- Earlier approaches used gradient features, binary classification, thresholding, and local features to identify crack regions and suppress noise.
- The paper trains a CNN to identify images containing cracks and proposes thresholding to extract cracks from classified color images.
II. METHODOLOGY
The proposed crack detection method has two stages: classifying images by crack presence and segmenting cracks in the positive images with adaptive thresholding.
- The method first classifies each image as positive or negative using a deep convolutional neural network.
- Only positive images are then processed with adaptive thresholding to extract the cracks.
- Images showing cracks are called positive images, whereas images without cracks are called negative images.
A. Image Classification
The image-classification network uses convolutional feature extraction with ReLU, normalization, pooling, softmax, and a fully connected classification layer. Figure 3 illustrates subsequent filtering and segmentation outputs for positive images.
- A. Image Classification: The proposed CNN uses ReLU activation and hierarchical feature extraction to process image inputs.
- A. Image Classification: Figure 3 compares an original positive image, its bilateral-filtered version, and the resulting segmentation.
- A. Image Classification: Convolutional outputs are batch-normalized, while max pooling downsamples the representations.
- A. Image Classification: Softmax converts the network output into a probability distribution, and a fully connected layer computes class scores.
B. Image Segmentation
The method smooths positive crack images, downsamples them to reduce noise, and uses adaptive thresholding to separate cracks from road surface.
- Bilateral filtering smooths positive images while preserving edges before segmentation.The filter uses spatial-distance and color-similarity weights.
- Downsampling produces an image approximately nine times smaller, which is normalized and used as the segmentation threshold.
- The thresholding method models the downsampled image as foreground cracks and background road surface separated by one threshold δ.
- Each pixel and its neighborhood provide intensity and mean-intensity features stored as vectors in a 2D histogram.The neighborhood size is controlled by τ.
- K-means searches along the principal diagonal, assigning foreground and background regions while excluding noisy regions from threshold selection.The within-cluster sum squares for different δ values guides threshold determination.
III. EXPERIMENTAL RESULTS
The experiments evaluate image classification and pixel-level segmentation, reporting 99.92% classification performance and better segmentation than Otsu’s method, with best performance at τ=1.
- Experimental setup: The experiment used 30,000 randomly selected images for training, with the remaining dataset images used for evaluation.The dataset contained 40,000 RGB images, evenly divided between positive and negative examples.
- Image classification: 99.92% precision, recall, accuracy and F1-measure were achieved by the proposed image classifier.The classifier produced four false positives and four false negatives.
- Image classification: 4.8 ms was required for image classification on a single Intel Core i7-8700K CPU core.
- Image segmentation: Pixel-level segmentation was evaluated against manually labeled crack areas because the dataset lacked pixel-level ground truth.
- Image segmentation: The proposed segmentation method outperformed Otsu’s thresholding in precision, accuracy and F1-measure, achieving its best performance when τ=1.Crack regions smaller than 100 pixels were ignored in the experiments.
IV. CONCLUSION AND FUTURE WORK
The paper combines neural-network image classification with bilateral filtering and adaptive k-means-based segmentation for road cracks, reporting strong classification and pixel-level results but limited performance on noisy color images.
- Conclusion: The proposed approach classifies images as crack-positive or crack-negative, then segments positive images using bilateral filtering and adaptive thresholding.The segmentation threshold is computed by minimizing within-cluster sum squares after downsampling.
- Conclusion: 99.92% classification precision and around 98.70% pixel-level segmentation accuracy were reported experimentally.
- Conclusion: Bilateral filtering minimized noisy pixels while preserving edges between cracks and road surface.
- Future work: Some color images with many noisy pixels could not be properly segmented by the proposed algorithm.The paper proposes training a deep neural network to segment cracks and road surface as future work.