Spatial-Coordinate Attention and Multi-Path Residual Block Based Oriented Object Detection in Remote Sensing Images

Abstract

ABSTRACT Remote sensing images have been widely used in various fields due to the great success of deep learning. However, as the basis of other tasks, remote sensing image object detection is still challenging. In this paper, we propose a Spatial-Coordinate Attention (SCA) and Multi-Path Residual Block (MPRB) guided oriented object detection algorithm based on an extended ResNet-18 backbone network. For specially, we first modify the coordinate attention module and spatial attention module through a multi-branch manner, and then combine them to generate the new Spatial-Coordinate Attention module. In this module, the channel information, location information and spatial information are integrated to enhance the representative capability of the object. After this, in order to further improve the feature representative power of a shallow CNN, such as ResNet-18. We increase the residual connection in the Residual Block and designed a Multi-Path Residual Block to make it more powerful. Moreover, we also introduce a new skip-connection architecture into the upsampling process to ensure that the predicted feature map can integrate high-level and low-level semantic information. We conduct experiments on the optical remote sensing image dataset UCAS-AOD, HRSC2016 and the SAR aerial dataset SSDD. Experiments show the effectiveness and feasibility of the algorithm in this paper.