Densely multiscale framework for segmentation of high resolution remote sensing imagery

Semantic segmentation has gained research attention in recent times, especially within the remote sensing community. The deep neural network has proven to be the most effective approach for segmentation applications due to its automatic feature extraction capability. Research results indicate that the multiscale segmentation frameworks are more suitable for high-level feature extraction, especially from complex remote sensing images. However, most existing multiscale frameworks are either complex or highly parameterized, making them inefficient for real-time remote sensing applications. In this work, we propose an accurate and highly efficient densely multiscale segmentation network specifically for real-time segmentation of remotely sensed imagery. We significantly improve the representation capability of the network by embedding its structure with the dense connection, which allows gradient to flow with ease through the network. The proposed network with few trainable parameters performed significantly on two publicly available challenging datasets, making it suitable for deployment on resource-constrained devices for real-time remote sensing applications.