Remote sensing image segmentation based on improved Mask-RCNN

Abstract

Satellite remote sensing images have the problems of large image scale, dense arrangement of segmentation targets and different directions. And the specifications and clarity are far from the natural images, resulting in difficulty in feature extraction. Therefore, the detection accuracy of Mask-RCNN is poor when applied to remote sensing image instance segmentation. In this regard, an improved Mask-RCNN algorithm is proposed. First, a deformable convolution kernel is introduced into the back bone network to adaptively change the theoretical receptive field. On this basis, the FPN module is modified, and feature layered fusion is introduced to further improve the feature extraction capability of the model. At the same time, the Soft-NMS algorithm is used to screen the target candidate frame. Validated using the iSAID dataset. The experimental results based on the data set show that the improved algorithm has higher detection accuracy than the original algorithm, and the missed detection rate is reduced.