特征增强的单阶段遥感图像目标检测算法

Single-Stage Remote Sensing Image Object Detection Algorithm Based on Feature Enhancement

针对遥感图像目标检测算法在特征融合过程中最高层信息丢失及遥感图像复杂背景难以区分的问题,提出了一个特征增强的单阶段遥感图像目标检测算法。该算法在基线单阶段对齐网络(S2A-Net)中引入残差特征增强处理策略和改进的通道注意力机制,从而保留特征融合过程中最高层信息并使网络抑制复杂背景信息。为了验证所提算法的有效性,分别在DOTA-v1.0数据集和HRSC2016数据集上进行了实验。实验结果表明,所提算法相比于S2A-Net算法,在DOTA-v1.0数据集上的平均精度提升了1.43个百分点,并在HRSC2016数据集上取得了比较有竞争力的结果。

To address the problem that the top-level information of remote sensing image object detection algorithm is lost in the process of feature fusion and the complex background in remote sensing image is difficult to distinguish, a feature-enhanced single-stage remote sensing image object detection algorithm is proposed. The algorithm introduces a residual feature enhancement processing strategy and an improved channel attention mechanism in the baseline single-shot alignment network(S~2A-Net), thereby retaining the highest-level information in the feature fusion process and enabling the network to suppress complex background information. To evaluate the effectiveness of the proposed algorithm, experiments are conducted on the DOTA-v1.0 dataset and HRSC2016 dataset, respectively. The experimental results show that, compared with the S~2A-Net algorithm, the average accuracy of the proposed algorithm on the DOTA-v1.0 dataset has been improved by 1.4 percent, and it has also achieved more competitive results on the HRSC2016 dataset.