Cross-Architecture Performance Evaluation of Transfer Learning Models for Multi-Class Vehicle Damage Severity Classification

Abstract

Automated vehicle damage classification supports objectivity and scalability in insurance claim processing and digital inspection systems; however, prior studies often report performance improvements without controlled experimental settings or statistical validation, limiting methodological reliability. This study establishes a statistically controlled cross-architecture evaluation framework to determine whether pretrained convolutional neural networks significantly outperform a custom baseline model in multi-class vehicle damage classification. A dataset of 891 labeled vehicle images categorized into heavy, medium, light, and normal damage was partitioned using stratified sampling (70% training, 15% validation, 15% testing). Four architectures Baseline (CustomCNN), VGG16, ResNet50, and MobileNetV2 were trained under identical preprocessing and optimization settings with two training durations (30 and 50 epochs). Five-fold cross-validation and paired t-test analysis were applied to assess statistical significance. At 30 epochs, MobileNetV2 achieved the highest accuracy (75.76%), while at 50 epochs VGG16 obtained the best performance (78.03%). Extending training duration did not produce statistically significant improvement (p > 0.05). Pretrained architectures significantly outperformed the baseline model, whereas ResNet50 did not demonstrate superior performance. The novelty of this study lies in its statistically validated comparative framework. Although limited by moderate dataset size and single-source imagery, the findings provide practical guidance for selecting efficient convolutional neural networks in vehicle damage classification systems.

GS Cited Analysis