An improved algorithm for semantic segmentation of remote sensing images based on DeepLabv3+

Abstract

Remote sensing image segmentation is a more and more popular topic currently, and obviously it depends on the improvement of semantic segmentation. Encoder-decoder structure is an effective architecture in semantic segmentation. Since the encoder network helps to obtain various scale feature from the deep convolution layers, and the decoder network often help to recover the spatial resolution and location information in detail. Drawing on this idea, Google proposed the DeepLabv3+ [9] after DeepLabv3[8], which reused DeepLabv3 as its encoder, and added the decoder part to help the network to recover accurate location information. However, the decoder part of DeepLabv3+ is simple and sometimes it's difficult to obtain enough details from the encoder, and achieves not so good results on remote sensing images. Therefore, we design our decoder by adding more skip connections and convolution layers, which improves the result of building detection in the dataset SpaceNet [1].