在目标检测领域,小目标的检测识别一直都是研究的难点,导致模型提取到的特征并不具有良好的表达能力,因此对小目标的检测结果不佳。为此,提出一种基于特征金字塔网络(FPN)的改进算法。在原有基础上增加并行分支,再融合两种不同上采样方法的特征信息以加强小目标特征的表达能力。同时,增加多阈值检测器(Cascade R-CNN)强化小目标定位能力。基于无人机航拍数据集进行实验,实验结果表明,在MS COCO数据集下,所提算法的平均精确率相比原始FPN算法提高了9.7个百分点,具有良好的检测性能。

The detection and recognition of small targets are always difficult for researchers in the field of target detection, resulting in the feature extracted from the model not having good expression ability, so the detection result of small targets is poor. This paper presents a modified algorithm based on feature pyramid network(FPN). Specifically, the parallel branch is devised on the original basis to fuse the feature information of two different up-sampling methods to enhance the representation ability of small objects. Meanwhile, a multiple threshold detector named Cascade R-CNN is added to prompt the localization ability of small objects. Experiments are conducted on UAV aerial image datasets. The experimental results reveal that the average precision of the proposed algorithm under MS COCO dataset increases by 9.7 percentage compared to that of the initial FPN algorithm; hence, the proposed algorithm has a good detection performance.