Peningkatan Kinerja Klasifikasi Kerusakan Jalan Menggunakan VGG-19 Berbasis Spatial Attention

Abstract

Road infrastructure damage is a serious challenge that threatens the safety of road users. This study aims to improve the accuracy of road damage classification using Deep Learning with a Visual Geometry Group-19 architecture modified with a spatial attention mechanism to select relevant features. The study utilizes a dataset of 3,760 road damage images consisting of four classes (Crack, Pothole, Rutting, and Normal). The dataset is divided into 70% training data and 30% testing data, followed by preprocessing stages including resizing, mean normalization, and standard deviation normalization.The dataset is trained using the Adam and SGD optimizers, along with variations in batch size. The experimental results show that the VGG-19 model with spatial attention using the Adam optimizer and a batch size of 8 achieved the highest accuracy of 91.67%, outperforming the model without attention, which obtained an accuracy of 89.89%.It can be concluded that the addition of the attention mechanism effectively improves the model’s performance in recognizing damage patterns. Therefore, it is recommended for the development of more accurate intelligent infrastructure monitoring systems. This study contributes by demonstrating that the integration of a spatial attention mechanism into VGG-19 consistently enhances classification accuracy compared to the standard VGG-19 model without attention, particularly in small batch size scenarios.