Cascaded panoptic segmentation method for high resolution remote sensing image

Great progress has been made for remote sensing image segmentation with the development of Deep Convolutional Neural Networks. However, Multiple convolutions significantly reduce the resolution and lead to the loss of many key information, the prediction accuracy of pixel categories is reduced. And DCNN accumulate context information on a large receptive field, which leads to blurred boundary segmentation of objects. This paper proposes a cascaded panoptic segmentation network to target the aforementioned problems. Firstly, a shared feature pyramid network backbone and a new hybrid task cascade framework are designed, which share the features and integrate the complementary features of different tasks in different stages, which can extract rich context information. Then, a functional module is designed to learn the mask quality of predicted instances in Mask R-CNN to calibrate the inconsistency between mask quality and mask score, thus to deal with the scale change of the object. Finally, a new Visual-saliency ranking module is designed to overcome the mutual occlusion problem between the prediction results, and strengthen robustness to illumination. The experimental results prove that our method still has significant advantages even compared with the most advanced methods, and ablation experiments also verify the effectiveness of our designed strategies. (C) 2021 Elsevier B.V. All rights reserved.

Automated summary

This paper proposes a cascaded panoptic segmentation network to address the issues existing in remote sensing image segmentation with deep convolutional neural networks. Experimental results demonstrate the effectiveness of the method by designing a shared feature pyramid network backbone, a new hybrid task cascade framework, and strategies such as learning mask quality.