Deep Learning-Based Detection of Paint Peeling: An Evaluation of CNN Architectures for Automated Building Inspections

Building facades are exposed to environmental and anthropogenic agents that accelerate degradation and compromise performance, durability, and service life. Paint peeling is a recurrent defect associated with loss of coating adhesion, moisture, material aging, and inadequate maintenance. This study evaluated convolutional neural networks for automatic paint-peeling detection on building facades. The dataset comprised 8,400 manually annotated image patches (448 × 448 pixels), predominantly from author-acquired images. Most source images were captured manually with a handheld RGB camera, while a smaller complementary portion was acquired using a DJI Air 2S UAV equipped with an RGB camera at variable distances of up to 5 m from the facades. Because manual acquisition predominated and capture distances varied, the dataset lacks a uniform flight-height range or spatial resolution. As standardized augmentation, each image and its annotations were rotated about the image center by 0°, 90°, 180°, and 270°. The dataset was divided into training, validation, and test subsets at a 70:20:10 ratio. Images were resized to 640 × 640 pixels during training and inference. Pre-trained YOLOv3, YOLOv5, YOLOv8, YOLOv12, Faster R-CNN, RetinaNet, and Cascade R-CNN models were retrained through transfer learning and fine-tuning. No systematic hyperparameter optimization was performed; predefined settings and implementation-default hyperparameters were used. A maximum of 150 epochs was established. YOLO models used native early stopping, whereas the remaining models were manually stopped based on learning-curve stagnation or overfitting. Test performance was assessed using Precision, Recall, mAP50, and mAP50–95. Except for Faster R-CNN, all architectures showed potential. YOLOv12 achieved the best metric balance, with an mAP50 of 0.820 and mAP50–95 of 0.630, while YOLO models offered a favorable trade-off between performance and training efficiency. Limitations included undetected early-stage peeling and scale effects: broader views favored advanced defects while masking smaller manifestations. Overall, YOLO architectures are promising for facade inspection.