A Hybrid Multi-scale and Attention Mechanism Approach for Pothole Image Classification: A Comparative Study

Potholes are a severe threat to transport infrastructure, safety, and mobility in urban areas, requiring efficient and automated classification techniques. In this paper, we present a multiscale hybrid deep learning model that integrates both Convolutional Neural Networks (CNNs) and Transformer-based models for the classification of multiclass pothole images. We compare different architectures, including a self-designed CNN, VGG16, VGG19, DenseNet169, DenseNet201, Xception, InceptionV3, ResNet50V2, and InceptionResNetV2. The suggested hybrid model combines DenseNet201 for deep feature extraction with a Transformer block for improving contextual comprehension based on multiscale features and self-attention mechanisms. Experimental evaluation using real-road surface image datasets—diverse for variability—demonstrates the hybrid DenseNet201-Transformer model performs significantly better than alternatives, with 98% classification accuracy and encouraging precision-recall measures for three classes: Normal, Pothole, and Bump. The results confirm the effectiveness of the combination of convolutional and attention-based architectures in enhancing robustness, generalizability, and performance of road damage classification models.