Deep Transformer-based Semantic Segmentation using Remote Sense Data

Abstract

Semantic segmentation is an important job in computer vision, and its applications have grown in popularity over the last decade. When we talk about remote sense imaging, different aspects are there which proved to be impacted factors while processing RS Data. Many survey applications, such as traffic monitoring, forest identification, and other natural calamities, employ these imagery. A large variety of models have been presented, the majority of them are based on CNN models and have obtained the best results so far. In recent years, models based on Transformers have gained popularity. With a few tweaks, we employ a transformer-based swin model. Because the fundamental disadvantage of transformer-based models is that they demand a lot of memory and processing resources. We offer the swin-model with a decreased number of blocks and MLP-Head, which is utilized to speed up the model and to address these concerns. In the case of RS-Data, edge improvement necessitates extra care. "Explicit Edge-Enhancement" and "Implicit-Edge-Enhancement" are two sophisticated strategies we deploy. The datasets Vaihingen and Potsdam are used in the TrSeg-RS model. The model improved its accuracy and found the best between Flops and mIoU. The vivid improvements in findings show that the TrSeg-RS can play a key role in RS issues