[Paper Review] Road Crack Detection Using Deep Convolutional Neural Network and Adaptive Thresholding

Rui Fan, Mohammud Junaid Bocus+6 more|arXiv (Cornell University)|Apr 18, 2019

Infrastructure Maintenance and Monitoring43 references17 citations

TL;DR

This paper proposes a two-stage deep learning approach for road crack detection: first using a deep convolutional neural network (DCNN) to classify images as containing cracks (99.92% accuracy), followed by bilateral filtering and adaptive thresholding to segment cracks from road surfaces. The method achieves 98.70% pixel-level segmentation accuracy, outperforming Otsu’s thresholding in precision and F1-score.

ABSTRACT

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.

Motivation & Objective

• To address the limitations of traditional, manual, and noise-sensitive crack detection methods that are time-consuming and subjective.
• To develop an automated, robust, and accurate system for detecting road cracks in real-world, low-quality images.
• To improve crack segmentation by combining deep learning classification with adaptive thresholding that reduces noise while preserving crack edges.
• To outperform conventional thresholding methods like Otsu’s in pixel-level precision and F1-score for crack segmentation.

Proposed method

• A deep convolutional neural network (DCNN) is trained to classify input images as positive (crack present) or negative (crack absent), using ReLU activation and max-pooling layers.
• Positive images are processed with bilateral filtering to suppress noise while preserving edge details between cracks and the road surface.
• The filtered images are downsampled, and an adaptive threshold is computed by minimizing within-cluster sum of squares using k-means clustering.
• Crack regions are extracted by segmenting the filtered images using the computed adaptive threshold, which dynamically adjusts based on local image characteristics.
• Performance is evaluated using standard metrics: precision, recall, accuracy, and F1-score at both image and pixel levels.
• The method is compared against Otsu’s thresholding using the same evaluation metrics on a manually annotated dataset with cracks below 100 pixels ignored.

Experimental results

Research questions

• RQ1 Can a deep convolutional neural network achieve high-accuracy image-level classification of road crack presence with minimal false positives and negatives?
• RQ2 Does bilateral filtering effectively reduce noise while preserving crack edge information prior to segmentation?
• RQ3 Can an adaptive thresholding method based on k-means clustering outperform fixed-threshold methods like Otsu’s in segmenting cracks from complex backgrounds?
• RQ4 What is the pixel-level segmentation accuracy of the proposed method compared to baseline thresholding techniques?

Key findings

• The proposed DCNN achieved a classification accuracy of 99.92%, with only four false positives and four false negatives across the test set.
• The adaptive thresholding method achieved a pixel-level F1-score of 0.9548 when τ was set to 1, outperforming Otsu’s method (F1 = 0.9462).
• The proposed method recorded a precision of 0.9774 and accuracy of 0.9870 at τ = 1, exceeding Otsu’s precision (0.9590) and accuracy (0.9848).
• The method maintained high performance across different τ values, with F1-scores consistently above 0.94, indicating robustness to threshold parameter variation.
• Despite improvements, some images with high noise levels still led to imperfect segmentation, indicating a need for further refinement.
• Image classification using the DCNN took only 4.8 ms per image on a single-core CPU, demonstrating computational efficiency.