Roof Defect Segmentation on Aerial Images Using Neural Networks

The paper describes usage of deep neural networks for flat roof defect segmentation on aerial images. Such architectures as U-Net, DeepLabV3+ and HRNet+ OCR are studied for recognition five categories of roof defects: “hollows”, “swelling”, “folds”, “patches” and “breaks”. Paper introduces RoofD dataset containing 6400 image pairs: aerial photos and corresponding ground truth masks. Based on this dataset different approaches to neural networks training are analyzed. New SDice coefficient with categorical cross-entropy is studied for precise training of U-Net and proposed light U-NetMCT architecture. Weighted categorical cross-entropy is studied for DeepLabV3+ and HRNet+ OCR training. It is shown that these training methods allow correctly recognize rare categories of defects. The state-of-the-art model multi-scale HRNet+ OCR achieves the best quality metric of 0.44 mean IoU. In sense of inference time the fastest model is U-NetMCT and DeeplabV3+ with worse quality of 0.33–0.37 mean IoU. The most difficult category for segmentation is “patches” because of small amount of images with this category in the dataset. Paper also demonstrates the possibility of implementation of the obtained models in the special software for automation of the roof state examination in industry, housing and communal services.