[Paper Review] Pavement Distress Detection and Segmentation using YOLOv4 and DeepLabv3 on Pavements in the Philippines

James-Andrew Sarmiento|arXiv (Cornell University)|Mar 11, 2021

Infrastructure Maintenance and Monitoring17 citations

TL;DR

This study proposes a deep learning pipeline for automated pavement distress detection and segmentation in the Philippines using YOLOv4 for object detection and DeepLabv3 for semantic segmentation on a small RGB image dataset. Despite limited data, DeepLabv3 achieved a mean IoU of 0.56 and Dice coefficient of 0.58, while YOLOv4 yielded a low mAP of 0.286, highlighting challenges in detecting fine cracks and scaling with bounding boxes.

ABSTRACT

Road transport infrastructure is critical for safe, fast, economical, and reliable mobility within the whole country that is conducive to a productive society. However, roads tend to deteriorate over time due to natural causes in the environment and repeated traffic loads. Pavement Distress (PD) detection is essential in monitoring the current conditions of the public roads to enable targeted rehabilitation and preventive maintenance. Nonetheless, distress detection surveys are still done via manual inspection for developing countries such as the Philippines. This study proposed the use of deep learning for two ways of recording pavement distresses from 2D RGB images - detection and segmentation. YOLOv4 is used for pavement distress detection while DeepLabv3 is employed for pavement distress segmentation on a small dataset of pavement images in the Philippines. This study aims to provide a basis to potentially spark solutions in building a cheap, scalable, and automated end-to-end solution for PD detection in the country.

Motivation & Objective

• Address the lack of automated pavement distress detection in developing countries like the Philippines, where manual surveys are costly and inconsistent.
• Develop a scalable, low-cost, and automated solution for monitoring road infrastructure health using deep learning on 2D RGB images.
• Evaluate the performance of YOLOv4 for object detection and DeepLabv3 for semantic segmentation on a small, real-world pavement image dataset from the Philippines.
• Provide a foundational framework for future end-to-end automated pavement condition assessment systems in resource-constrained settings.

Proposed method

• Employed YOLOv4 for instance-level detection of pavement distresses (e.g., alligator cracks, delamination) using bounding boxes on 2D RGB images.
• Applied DeepLabv3 with atrous spatial pyramid pooling (ASPP) and encoder-decoder architecture for pixel-wise semantic segmentation of distress types.
• Preprocessed images using CLAHE to enhance contrast and improve feature visibility for both models.
• Trained models on a small dataset of 305 images with five distress classes: alligator cracks, delamination, bowl-shaped depressions, scaling, and disintegration.
• Evaluated detection performance using mean Average Precision (mAP) and segmentation performance using mean Intersection over Union (mIoU) and Dice coefficient.
• Used data augmentation and transfer learning to mitigate overfitting due to limited training samples.

Experimental results

Research questions

• RQ1 Can YOLOv4 achieve reliable object detection of pavement distresses in a small, real-world dataset from the Philippines?
• RQ2 Can DeepLabv3 produce accurate semantic segmentation of complex pavement defects such as overlapping cracks and scaling with limited training data?
• RQ3 How do preprocessing techniques like CLAHE affect the performance of deep learning models in pavement distress detection and segmentation?
• RQ4 What are the primary failure modes of YOLOv4 and DeepLabv3 in detecting and segmenting pavement faults in this context?
• RQ5 To what extent can a small-scale, low-cost deep learning pipeline serve as a foundation for scalable, automated pavement condition monitoring in developing countries?

Key findings

• DeepLabv3 achieved a mean Intersection over Union (mIoU) of 0.56 and a Dice coefficient of 0.58 on the DPWH dataset, indicating moderate to good performance in semantic segmentation despite limited training data.
• YOLOv4 achieved a mean Average Precision (mAP) of 0.286 on pre-processed images and 0.272 on original images, indicating poor detection performance, particularly for fine or complex distresses.
• The models struggled with class confusion, misclassifying delamination as bowl-shaped depressions and vice versa, likely due to texture similarities and CLAHE-enhanced feature sensitivity.
• DeepLabv3 exhibited high false positives by segmenting normal pavement surfaces as scaling, especially in textured or uneven regions.
• YOLOv4 suffered from high false negative rates, particularly missing small or thin cracks and scaling defects, due to the limitations of bounding box detection for irregular, elongated, or fragmented distresses.
• Despite low metric scores, both models demonstrated feasibility in detecting major distress types like alligator cracks and delamination, suggesting potential for improvement with larger, more uniform datasets.