针对基于深度学习的遥感目标检测算法参数冗余、计算量大且实时检测性能较差的问题, 提出了一种基于深度可分离卷积的实时遥感目标检测算法。首先通过K-means++算法对数据集进行锚框(Anchor)聚类分析, 使锚框参数更加符合遥感检测场景。为了降低模型参数量、提升检测速度, 以轻量级网络MobileNetv3作为主干网络进行特征提取; 此外, 基于深度可分离卷积的PANet(Path Aggregation Network)结构的设计, 使网络参数量进一步降低。改进后模型参数量仅为原来的18.3%, 检测速度提升2.19倍, 在UCAS_AOD, RSOD, DIOR这3个遥感数据集上进行测试, 实验结果表明, 算法鲁棒性强, 能够在保证模型检测精度的同时有效提高检测实时性。

Remote sensing target detection algorithms based on deep learning have redundant parameters, large computation amount and poor real-time detection performance.To solve the above problems, a real-time remote sensing target detection algorithm based on depth separable convolution is proposed.The datasets are analyzed by anchor box (Anchor) clustering via the K-means++ algorithm, making the anchor box parameters more compliant with the remote sensing detection scenario.In order to reduce the quantity of model parameters and improve the detection speed, feature extraction is performed with the lightweight network MobileNetv3 as the backbone network.In addition, the design of PANet (Path Aggregation Network) structure based on depth separable convolution makes the quantity of network parameters further reduced.The quantity of model parameters after improvement is only 18.3% of the original, and the detection speed is 2.19 times faster than the original.Tests are conducted on three remote sensing datasets, that is, UCAS_AOD, RSOD and DIOR, and the experimental results show that the algorithm is robust and can effectively improve the real-time detection performance while ensuring the model detection accuracy.