Semantic Segmentation of Built-UP Areas in SAR Images Based on Improved YOLOv7 Model

With the continuous development of synthetic aperture radar technology, image segmentation has evolved into a pivotal method for identifying and classifying land cover. Traditional segmentation methods usually suffer from poor segmentation performance and low accuracy when dealing with dense and small targets. To address these issues, this paper proposed an improved YOLOv7 model for building object segmentation. The model was improved as follows: (1) The loss function of the original YOLOv7 model was replacd by SIoU (Scally Intersection over Union). (2) The attention mechanisms was introduced into the network. Through comparison with original YOLOv7 model, and Mask R-CNN model, the improved model achieves significant improvements in accuracy. Furthermore, the improved model outperforms other models in terms of runtime speed and power consumption.