Damage Detection of Red Bricks on Building Facades Based on an Improved YOLOv8: A Case Study of the Wuhan Lifen Architecture

Automated detection of surface damage on historical red brick facades can support preliminary assessments for architectural heritage conservation. However, complex masonry textures and indistinct damage boundaries remain challenging for conventional object detectors. This study presents YOLO-DINO-ECSA, a YOLOv8-based detection framework that integrates pretrained DINOv2 features, an ECSA attention module, and a hybrid loss combining Focal Loss and SIoU. DINOv2 provides contextual features, while ECSA refines local feature representations. The modified loss functions are designed to address difficult samples and bounding box localization. Experiments were conducted on a dataset of historical red brick facades collected in Wuhan, China. Across three randomized image-level partitions, YOLO-DINO-ECSA achieved a mean mAP@0.5 of 77.50 ± 0.46%, compared with 73.45 ± 0.88% for YOLOv8n. The proposed model also achieved higher mean Precision but lower mean Recall and F1-score. Its inference throughput was 59.53 FPS on an NVIDIA RTX 4090 GPU, although the integrated framework required substantially greater computational resources than YOLOv8n. These results demonstrate the performance of the proposed system on the evaluated dataset and support its potential use as a preliminary visual screening tool for masonry conservation.