ENHANCED VEHICLE DAMAGE CLASSIFICATION USING ATTENTION-AUGMENTED DEEP LEARNING ONTHREE-QUARTERS VIEW CAR DAMAGE DATASET

Abstract

The manual detection process for vehicle damage is monotonous and prone to errors. Especially when dealing with massive-scale scenarios like insurance claims assessment and repair costs evaluation, the process becomes cumbersome. In this report, we propose a classification framework that automatically identifies damage using image processing techniques and deep learning algorithms. We evaluate our proposed system using the TQVCD dataset that includes images from a three quarter view showing both frontal and lateral parts of vehicle damage. All images go through a pre-processing stage where we resize, normalize, and randomly augment them by rotating, horizontal flipping, and adjusting their bright ness levels. We compare three architectures in terms of performance on TQVCD dataset: E!cientNet-B3, DenseNet169, and ResNet50 augmented with an attention mechanism. In the latter, the attention module allows the network to attend only to the relevant part of the image where vehicle damage is located. On TQVCD, ResNet50 with Attention achieves 87% accuracy, DenseNet169 achieves 85%, and E!cientNet-B3 achieves 82%. Another dataset used to verify the e”ectiveness of our proposed architecture in di”erent vehicle categories is CarDD, which includes many types of damages. On CarDD, ResNet50 with Attention achieves 85% accuracy, while DenseNet169 and E!cientNet-B3 achieve 83% and 80% respectively. Lastly, on TartesiaDS, an insurance-oriented dataset with annotations made by professionals under various lighting conditions, ResNet50 with Attention achieves 89% accuracy, which is higher than the baseline ResNet50 architecture tested on the same dataset reported by P´erez-Zarate et al. (2024) achieving 85.8%. On TartesiaDS, DenseNet169 achieves 87% and E!cientNet-B3 achieves 83 Overall, our study proves the reliability and e”ectiveness of the attention-augmented deep learning models in the process of automating damage detection in vehicles. In particular, on the three presented benchmarks, architectures with attention mod ules perform better than their counterparts without such mechanisms. Our future research will focus on ensemble fusion of multiple models, grading the degree of damage, and utilizing explainable AI methods like Grad-CAM.