A deep-learning semantic segmentation approach for road segmentation of UAV images

Road image segmentation is important for a variety of reasons, including road maintenance, intelligent transportation systems, and urban planning. The method has recently been adapted to road images captured by unmanned aerial vehicles (UAV), because of its low cost, easy manoeuvrability and wide field of view. Due to the complexity of the backgrounds in these images, high-precision road segmentation from UAV images remains difficult. This research proposed an automated road segmentation model utilising a deep learning technique called DeepLab V3+ semantic segmentation to solve the problem. A UAV is utilised to capture and collect road images in the state of Kedah and Selangor, Malaysia. Then, the DeepLab V3+ with Resnet-50 backbone is developed and trained to segment the road from the background. The performance is then evaluated by comparing the segmented images using deep learning to images that have been labelled manually. For the evaluation, three measures are used: pixel accuracy (PA), mean area intersection by union (mIoU), and mean F1-score (MeanF1). Additionally, for benchmarking purposes, the research compares the segmentation performance with two different backbones of DeepLab V3+ called Resnet-18 and Mobile NetV2 backbone. According to simulation findings, the DeepLab V3+ with Resnet-50 outperformed the DeepLab V3+ with Resnet-18 and Mobile NetV2 techniques. Experiments on various road images as well as comparisons with different backbones demonstrate the effectiveness and robustness of the proposed DeepLab V3+ with Resnet-50. The method outperformed the Resnet-18 and Mobile NetV2 by at least 0.22%, 0.14%, and 1.39%, for PA, mIoU and MeanF1 respectively.