Multiclass Road Surface Damage Classification Dataset: Computer Vision

Description

This dataset contains 1,530 high-resolution, three-channel (RGB) images of road surfaces. It is designed to support machine learning and computer vision research for robust, three-class road damage classification, particularly contributing to innovations in automated infrastructure inspection, predictive maintenance, and smart city frameworks. Dataset Composition: The dataset is organized into three distinct, structurally meaningful categories based on pavement degradation. ⦁ Surface Erosion (651 images): Images displaying weathering, raveling, or degradation of the top asphalt layer. ⦁ Crack (478 images): Images exhibiting linear, transverse, or interconnected (alligator) cracks on the road surface. ⦁ Pothole (401 images): Images showing bowl-shaped depressions or significant missing pavement chunks. Preprocessing: ⦁ Images are harmonized to 640 x 640 pixels and loaded as three-channel RGB images. ⦁ A unified preprocessing pipeline was applied, including resizing, aspect-ratio preservation (via optional zero-padding), and per-image color normalization to ensure consistency across varying lighting conditions. ⦁ The dataset is pre-structured to support standard machine learning workflows, typically divided into an 80/10/10 (train/validation/test) split to ensure robust model evaluation and prevent data leakage. This dataset can be effectively used for: ⦁ Multiclass image classification and automated road anomaly recognition tasks. ⦁ Deep learning model development, including Vision Transformers (ViTs), parameter-efficient CNNs, and real-time edge computing architectures. ⦁ Benchmarking automated pavement distress detection systems for municipal asset management. ⦁ Evaluating autonomous vehicle perception systems for obstacle avoidance and path planning. File Information: ⦁ Total Images: 1,530 ⦁ Image Format: RGB (3-channel) images ⦁ Resolution: 640 x 640 pixels ⦁ Data Structure: Organized into distinct directories based on damage class.

Steps to reproduce

1. Download the dataset: Access the dataset files from this Mendeley Data repository and extract the contents into your working directory. 2. Dataset structure: The dataset is organized into distinct directories based on the specific damage class, making it compatible with standard image data loaders: RoadDamage_Classification_Dataset/ ├── Crack/ │ ├── crack_001.jpg │ ├── ... ├── Pothole/ │ ├── pothole_001.jpg │ ├── ... └── Surface_Erosion/ ├── erosion_001.jpg ├── ... 3. Data preprocessing: ⦁ Ensure all images are loaded as three-channel RGB images and verify the resolution is resized to the harmonized 640x640 pixels. ⦁ Apply the unified preprocessing pipeline, utilizing zero-padding where necessary to preserve the aspect ratio, and perform per-image color normalization to account for varying lighting conditions. ⦁ Split the dataset into training (80%), validation (10%), and testing (10%) sets to ensure robust model evaluation and prevent data leakage. 4. Model training example: ⦁ Load the data using standard data generators in PyTorch or TensorFlow. ⦁ Apply data augmentations (such as rotation, zoom, or brightness adjustments) to improve generalization across different road environments. ⦁ Initialize training using an image classification architecture, such as a Vision Transformer (ViT), a parameter-efficient CNN (e.g., MobileNet or EfficientNet), or a custom edge computing architecture. 5. Evaluation: ⦁ Evaluate the model using standard multiclass classification metrics, including accuracy, precision, recall, and F1-score. ⦁ Visualize the results using a confusion matrix to check for misclassifications between structurally similar classes (e.g., severe cracking being misclassified as a pothole). ⦁ Assess inference speed if benchmarking for real-time autonomous vehicle perception or municipal edge-device deployment.

Institutions

• Daffodil International University

  Dhaka Division, Dhaka

Categories

Licence