Object-Enhanced Semantic Segmentation Model for High-Resolution Remote Sensing Images

Abstract

The semantic segmentation of high-resolution remote sensing images is of paramount importance for applications such as land use, land cover, and cartography, and it holds great potential for widespread use. The characteristics of high-resolution remote sensing images are addressed in this study, and an end-to-end semantic segmentation network model is devised. An improved Resnet was employed to extract image features, and an efficient channel attention module was introduced to enhance the weights of important channels, with only a marginal increase in parameters. Due to significant scale variations among objects in high-resolution remote sensing images, a multi-scale algorithm was utilized to extract more comprehensive information. An object-enhancement algorithm was introduced, which, by incorporating object features, enhanced the contextual information of pixels in the high-level feature space. Through an analysis of the ultimate predictive performance, the model proposed in this paper demonstrated a certain advantage over popular semantic segmentation models such as Unet, DAnet, DeeplabV3, in predicting high-resolution remote sensing images.