Faster and Better Instance Segmentation for Large Scene Remote Sensing Imagery

Abstract

Instance segmentation is difficult to apply into high-precision scenes with limited computational resources, such as unmanned aerial vehicles (UAV), compared to object detection. In this article, a faster and better instance segmentation network (FB-ISNet) based on CondInst is proposed in large scene remote sensing imagery. The FB-ISN et aims to improve efficiency without sacrificing accuracy. First, we adopt the deep layer aggregation as the backbone network to extract feature. Then, the BiFPN is used to overcome the limitation of one-way information flow in FPN and obtain effective multi-scale feature fusion. Next, the detection and instance segmentation heads are optimized to further reduce the amounts of parameters. Finally, the experimental comparisons on the SSDD, NWPU VHR-10 dataset and a large scene remote sensing imagery show that our FB-ISNet can strike a good balance between accuracy and speed, and the comparison with the existing algorithms also demonstrates the superiority of our approach.