DenseNet121-Enabled Automated Pavement Distress Detection: A Novel Deep Learning Framework for Granular Road Damage Assessment and Comparative Architectural Analysis

The accelerated deterioration of pavements in developing economies like Bangladesh is mainly due to heavy axle loads combined with the tropical climatic stressors and thus presents a pressing need for scalable and objective methods for pavement condition assessment beyond the usual manual condition inspections. While deep learning techniques have improved the recognition of road damages, the systematic benchmarking on region-specific tropical imagery is still quite sparse, limiting reproducibility and reducing the practical guidance about backbone selection. To address this issue (methodological gap), we present a leakage safe and unified transfer learning benchmarking pipeline that is adapted for three class pavement distress classification (Asphalt, Crack, Pothole) using Bangladesh specific RSD-BD dataset. In contrast to previous regional studies, our framework compares all 7 of the ImageNet pretrained CNN backbones under identical stratified data splits, fixed classifier head, and predetermined training budget, and, as a result, separates the architectural effects. While testing on a held-out test set, DenseNet121 performs best with an accuracy of 94.70 per cent and a weighted F1 score of 94.69 per cent, with results that are balanced between the three classes. The proposed protocol therefore allows reproducible backbone selection and is especially well suited for use in lightweight image-level screening tasks in the context of routine pavement management workflows.