针对现有基于平视图像目标检测算法在航拍图像上检测精度不高的问题,提出一种具有动态感受野的单阶段目标检测算法。该算法采用SE-ResNeXt作为特征提取网络,在RetinaNet结构中添加bottom-up短连接通路和全局上下文上采样模块,增强了检测层特征的结构性和语义性。构造具有动态感受野的检测子网络,动态选取适当尺度的感受野特征进行目标检测。在实景航拍数据集上进行实验评测,并与相关算法作对比,结果表明改进算法在数据集上表现良好,性能指标具有明显提升,即使在光线昏暗、下视视角、斜下视视角和密集目标等场景图像中,也具有较好的检测效果。

The accuracy of existing image-based methods for aerial imaging of flat-view images is limited. In this paper, a dynamic receptive field-based single-stage object detection algorithm is proposed to address this problem. First, the feature pyramid network is constructed by using SE-ResNeXt. This network is used as the backbone network to extract features efficiently. A bottom-up short connection path and a global context upsampling module are proposed to enhance the structural and semantic features of the detection layer. A dynamic receptive field-based detection subnet is designed to dynamically select the receptive field of an appropriate scale for object detection. Experimental evaluation is conducted on a realistic aerial dataset, and the results are compared with those of other related algorithms. The results show that the improved algorithm performs better on the dataset, and the performance score is evidently increased. It also exhibits good detection capability in scene images such as dim light, down view, oblique view, and dense objects.