HybridNet: Integrating Multiple Approaches for Aerial Semantic Segmentation

In recent times, semantic segmentation for VHR aerial images has become an emerging research topic due to its widespread applications in disaster management, environmental monitoring, natural resource mapping, etc. The problem of semantic segmentation can be modeled as an image-to-image mapping problem where pixel-level classification is required. Pixel level classification is challenging for the high-resolution aerial image due to the presence of the tiny objects in low-frequency and more information details for such tiny objects required for dense semantic labeling. In general, encoder–decoder based architecture for semantic segmentation suffers from information loss due to the up and downsampling process. To handle this, we extend a high-resolution network with dense connection integration to preserve the original resolution and better parameter sharing. We also incorporate a lightweight self-attention module for positional attention, which results in better segmentation maps. Additionally, we use a generalized Hough transform based deep voting module for pixel dependencies extraction. Experimental results reveal that the proposed model achieves the best mean intersection over union and overall accuracy in local and benchmark evaluation on the Vaihingen and Potsdam datasets.