Remote sensing image segmentation of ground objects based on improved Deeplabv3+

Abstract

Aiming at the problem of unsatisfactory extraction effect caused by unbalanced category distribution and large differences in scene styles in different domains in the process of extracting key feature information from remote sensing images, an improved Deeplabv3+ semantic segmentation network is adopted. The instance-batch normalization (IBN) module is used in the backbone network ResNetl01 to enhance the model’s generalization ability for remote sensing images with large differences in styles. At the same time, in order to further improve the segmentation accuracy of the model, The squeeze-and-excitation (SE) module is added to the network to strengthen important channel information and alleviate the problem of information loss. The loss function uses the joint loss function of Dice+Focal, dice loss can alleviate the impact of imbalanced category distribution on the extraction of small targets. Focal loss can not only make the model pay more attention to objects that are difficult to classify, but it can also improve the instability of network training caused by dice loss. Experimental results show that compared with the original Deeplabv3+model, the improved Deeplabv3+ improves Fl-Score by 7.78% and Intersection over Union(IoU) by 5.78%. Compared with other mainstream semantic segmentation models (including FCN, UNet, and SegNet), the improved Deeplabv3+ also achieves better segmentation accuracy in ground feature extraction.