Road Crack Detection Using Deep Convolutional Neural Network and Adaptive Thresholding (1904.08582v1)

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.

• The paper presents a two-step method using a deep CNN and adaptive thresholding that achieves 99.92% classification accuracy and 98.70% segmentation precision.
• The methodology leverages hierarchical feature extraction, bilateral filtering, and k-means clustering to robustly detect road cracks under varied conditions.
• This approach automates crack detection in roads, reducing maintenance costs and paving the way for future enhancements with advanced neural network architectures.

Road Crack Detection Using Deep Convolutional Neural Network and Adaptive Thresholding: An Analysis

The paper "Road Crack Detection Using Deep Convolutional Neural Network and Adaptive Thresholding" presents a structured approach to addressing road safety hazards posed by surface cracks. The work introduces a two-step method combining a deep convolutional neural network (DCNN) for image classification with an adaptive thresholding technique for image segmentation.

Image Classification

The initial phase involves a deep convolutional neural network specifically designed for the binary classification of road images into crack-present (positive) and crack-absent (negative) categories. The architecture, employing layers such as ReLU for activation, batch normalization, max pooling, and a fully connected layer for classification, achieves an impressive accuracy rate of 99.92%. This efficacy is facilitated by hierarchical feature extraction, which optimizes the network's ability to discern features indicative of road cracks. Training is conducted on a robust dataset comprising 40,000 images, divided equally between positive and negative samples, ensuring comprehensive coverage of potential road conditions.

Image Segmentation

Following classification, the methodology employs bilateral filtering to smooth positive images, aiming to preserve edge details while reducing noise. The paper introduces an adaptive thresholding technique, framed as a 2D vector quantization problem. By leveraging a downsampled representation of the filtered images, the method utilizes $k$-means clustering to bifurcate the image into foreground (cracks) and background (road surface). This approach ensures that segmentation remains robust against varying lighting and textural conditions of the road surface.

Empirical Results

The experimental results underscore the method's effectiveness, with pixel-level segmentation achieving a precision of 98.70% under optimal parameter settings. Compared to traditional techniques like Otsu's thresholding, the proposed adaptive strategy enhances segmentation precision, accuracy, and the F$_1$-measure. This performance is critical in real-world applications where accuracy dictates the reliability of automated inspection systems.

Implications and Future Directions

The integration of DCNN for classification and adaptive thresholding for segmentation demonstrates potential utility in intelligent transportation systems (ITS), where timely and accurate detection of road surface conditions is paramount. By automating the traditionally labor-intensive and subjective process of crack detection, this approach can significantly reduce maintenance costs while improving road safety.

Future research could explore the incorporation of more advanced neural network architectures, possibly leveraging transfer learning or transformer-based models to further refine segmentation accuracy. Additionally, expanding the dataset to include diverse environmental conditions and types of road surfaces could bolster the model's generalizability.

Overall, this work provides a solid foundation for enhancing automated road maintenance systems and facilitates further advancements in the field of image-based structural health monitoring.