FACENet: A Fusion Atrous and Channel Enhancement Network for Remote Sensing Image Instance Segmentation

Abstract

The instance segmentation task has been widely used in remote sensing. However, existing remote sensing instance segmentation models may lead to incomplete mask segmentation in complex and diverse background environments. In addition, commonly used feature fusion methods struggle to handle instances of different sizes well and predominantly suffer from loss of semantic information, failing to segment the mask accurately. To solve these problems, we propose a fusion atrous and channel enhancement network (FACENet) for the remote sensing image (RSI) instance segmentation. Specifically, we first replace the FPN with the FACE-FPN, which produces a more detailed pyramid by increasing the receptive field at the feature level. Second, we propose a semantic enhancement module for mining the rich semantic information of the underlying features. Then, we enhance the model's adaptability to complex object deformations by introducing deformable convolution. Experiments on the iSAID, NWPU VHR-10, and HRSID datasets demonstrate that our proposed FACENet outperforms SOLOv2 in terms of average accuracy by 5.1%, 12.9%, and 7.6%, respectively, and beats other instance segmentation models.