Modified MobileNetV2 transfer learning model to detect road potholes

Area of Science:

• Civil Engineering
• Computer Science
• Artificial Intelligence

Background:

• Potholes and other road damages degrade pavement infrastructure, impacting safety and maintenance costs.
• Accurate detection of pavement distress is essential for timely repairs and infrastructure management.
• Current methods for road damage assessment can be labor-intensive and subjective.

Purpose of the Study:

• To develop and evaluate deep learning (DL) models for enhanced pothole detection in pavement images.
• To introduce a novel modified MobileNetV2 (MMNV2) model optimized for efficient feature extraction and accurate classification.
• To compare the performance of MMNV2 against fourteen other established DL models for pavement assessment.

Main Methods:

• Utilized transfer learning and deep learning techniques to preprocess digital pavement images.
• Evaluated fourteen DL models including MobileNetV2, DenseNet, ResNet, and EfficientNet variants.
• Developed a modified MobileNetV2 (MMNV2) model by integrating a five-layer pre-trained network for improved performance.
• Trained and tested models on a dataset of 5,000 pavement images with a learning rate of 0.001.

Main Results:

• The MMNV2 model achieved superior performance in classification, detection, and prediction accuracy compared to other evaluated models.
• Achieved 99.95% accuracy in classifying images as 'normal' or 'pothole'.
• Reported 100% recall, 99.90% precision, 99.95% F1-score, and a minimal 0.05% error rate.

Conclusions:

• The MMNV2 model offers a highly accurate and efficient solution for automated pothole detection.
• The model's reduced parameter count and superior performance make it suitable for real-world pavement assessment applications.
• This deep learning approach enhances infrastructure maintenance strategies and contributes to public safety.