Remote Sensing Image Object Detection Algorithm Combining Attention

Xu Yang, Dan Qu, Nianwen Si, Shangde Xin

Abstract: To address the issue of the complex background being difficult to differentiate and the object percentage, rotation angle, and position affecting the identification accuracy, a better object detection algorithm for remote sensing images is given. This algorithm uses the improved frequency channel attention to make the network take more notice of the foreground information, in order to achieve the effect of suppressing the complex background information. It also introduces a multiple attention mechanism in the baseline, which is conducive to alleviating the issue of increasing the detection difficulty due to the different proportion of objects in remote sensing images, arbitrary rotation angle, and position. Experiments on the DIOR, NWPUVHR-10, and RSOD datasets, respectively, are conducted to confirm the efficacy of the proposed algorithm. The suggested method's average accuracy on the DIOR dataset is 1.46% higher than that of the single-stage ATSS algorithm, and it has also produced results that are competitive on the NWPUVHR-10 and RSOD datasets.