The method and implementation of a Taiwan building recognition model based on YOLOX-S and illustration enhancement

Abstract

Earthquakes pose significant risks in Taiwan, necessitating effective risk assessment and preventive measures to reduce damage. Obtaining complete building structure data is crucial for the accurate evaluation of earthquake-induced losses. However, manual annotation of building structures is time-consuming and inefficient, resulting in incomplete data. To address this, we propose YOLOX-CS, an object detection model, combined with the Convolutional Block Attention Module (CBAM), to enhance recognition capabilities for small structures and reduce background interference. Additionally, we introduce the Illustration Enhancement data augmentation method to improve the recognition of obscured buildings. We collected diverse building images and manually annotated them, resulting in a dataset for training the model. YOLOX-CS with CBAM significantly improves recognition accuracy, particularly for small objects, and Illustration Enhancement enhances the recognition of occluded buildings. Our proposed approach advances building structure recognition, contributing to more effective earthquake risk assessment systems in Taiwan and beyond.