A Target Detection Algorithm for Remote Sensing Images Based on a Combination of Feature Fusion and Improved Anchor
Wenqing Zhao, Yijin Kang, Hao Chen, Zhenbing Zhao, Yongjie Zhai, Panpan Yang
Abstract
Aiming at the low average target accuracy in remote sensing images with complex background and multiple small targets, we propose a new algorithm for small target detection based on a combination of feature fusion and improved anchor. Firstly, a data enhancement method is presented to increase the number of small targets in remote sensing images. Secondly, the high level features and low level features are efficiently fused using pixel-by-pixel summation and lightweight feature extraction, which can extract features that are more favorable for small target detection. Finally, by adjusting the aspect ratio of the anchor, the misdetection rate of small targets is reduced. The algorithm proposed in this paper is applied on the PASCAL VOC data set and UCAS-AOD data set, with the mean average accuracy (mAP) of 82.5% and 87.9%, respectively. The test results demonstrate that the proposed method can not only achieve a comparable detection speed but also demonstrates considerable accuracy superiority over traditional algorithms.