Multi-Class Road Defects Detection and Classification System Using Transfer Learning-Based Deep Convolutional Neural Networks

Abstract

The road’s infrastructure is crucial for growth, development, and forming the backbone of any country's economy. The accurate detection and classification of road defects for optimal road maintenance is a challenging task due to the varied types of road defects of different severity. This paper presents a transfer learning model for the detection and classification of road defects based on types of defects (cracks and potholes) and severity. A new local dataset was introduced consisting of road surface images (defects and non-defects) of Kaduna metropolis, Nigeria. The types and severity of the defects were grouped as non-defect, low-pothole, low-crack, moderate-pothole, moderate-crack, high-pothole, and high-crack. The model was developed by extracting the features using pretrained VGG19 and EfficientNetB3. The extracted features were concatenated and evaluated on the locally gathered datasets. The pretrained EfficientNetB3 achieved an accuracy 98.1% higher than the 97.9% accuracy of VGG19. The concatenated model (VGG19+EfficientNetB3) achieved 98.5% accuracy, which outperformed the two pretrained models, VGG19 and EfficientNetB3. This study demonstrated the benefit of combining the strengths of VGG19 and EfficientNetB3, for improved performance and efficiency in road defect detection and classification tasks.