Arbitrary-oriented object detection algorithm based on cross-layer integrating attention mechanism for remote images

HuiJun Su,Zhihong Chen,Xinwang Shao,Shuanglong Xu,Y. Miao,Jia Lu,Baoqi Guo

DOI: https://doi.org/10.1117/12.2641358

2022-08-04

Abstract:Compared with the classical object detection algorithms with horizontal bounding box, such as Yolo and Faster-R-CNN, the oriented object detector in remote sensing images can be more robust. Considering that the challenge of oriented object detection in remote sensing images, we propose a cross-layer feature fusion improvement strategy. Specifically, to obtain multiple layer feature maps for subsequent object detection, the shallow feature map with texture information merged with the deep feature map with semantic features. The spatial attention mechanism is introduced to enhance our algorithm's attention to the non-local information in these feature maps. Extensive experiments on a public dataset, DOTA, demonstrate the effectiveness of our proposed method. Under the large difference of object scale, arbitrary orientation of object, objects with dense arrangement and complex background, the experimental results show that our method has better performance.

Computer Science,Engineering,Environmental Science