An improved DeepLabv3+ lightweight network for remote-sensing image semantic segmentation

To improve the accuracy of remote-sensing image semantic segmentation in complex scenario, an improved DeepLabv3+ lightweight neural network is proposed. Specifically, the lightweight network MobileNetv2 is used as the backbone network. In atrous spatial pyramid pooling (ASPP), to alleviate the gridding effect, the Dilated Convolution in original DeepLabv3+ network is replaced with the Hybrid Dilated Convolution (HDC) module. In addition, the traditional spatial mean pooling is replaced by the strip pooling module (SPN) to improve the local segmentation effect. In the decoder, to obtain the rich low-level target edge information, the ResNet50 residual network is added after the low-level feature fusion. To enhance the shallow semantic information, the efficient and lightweight Normalization-based Attention Module (NAM) is added to capture the feature information of small target objects. The results show that, under the INRIA Aerial Image Dataset and same parameter setting, the Mean Pixel Accuracy (MPA) and Mean Intersection over Union (MIoU) are generally best than DeepLabv3+ , U-Net, and PSP-Net, which are respectively improved by 1.22%, − 0.22%, and 2.22% and 2.17%, 1.35%, and 3.42%. Our proposed method has also a good performance on the small object segmentation and multi-object segmentation. What’s more, it significantly converges faster with fewer model parameters and stronger computing power while ensuring the segmentation effect. It is proved to be robust and can provide a methodological reference for high-precision remote-sensing image semantic segmentation.