融合SPP与FPN的光学遥感图像飞机目标检测

Aircraft Target Detection in Optical Remote Sensing Images Based on SPP and FPN Fusion

针对光学遥感图像飞机目标背景复杂、检测精度与检测速度不平衡、易漏检等问题,提出一种融合不同网络模块的SPSSD模型。首先,采用Resnet50替换SSD300算法中的特征提取网络,并加入可操控的空洞卷积模块,扩大特征感受野获取更多有利检测目标的特征信息;其次,加入FPN和SPP网络,得到浅层特征信息,并将感受野二次放大后的特征与深层特征信息进行融合;然后,送入ECANet网络中获取更加完整且更具有判断力的特征信息;最后,采用NWPU-RESISC45数据集3400张高分辨率飞机遥感图像输入至SPSSD模型中迭代训练,最终改进算法模型mAP值达到92.68%,较改进前的算法模型提升了5.18个百分点,检测速度达到25.1帧/s。实验结果表明,该方法可以有效兼顾飞机目标的检测精度与检测速度,一定程度上降低了目标漏检率。

In order to solve the problems of complex background of aircraft target in optical remote sensing images,the imbalance between detection accuracy and detection speed,and the likelihood of miss detection,an SPSSD model fusing different network modules is proposed.Firstly,Resnet50 is adopted to replace the feature extraction network in SSD300 algorithm,and a controllable dilated convolution module is added to expand the features receptive field to obtain more feature information favorable to target detection.Secondly,shallow feature information is obtained by adding FPN and SPP network,and the features of the expanded receptive field are fused with deep feature information.Then,the feature information is sent to the ECANet to obtain more complete and more judicious feature information.Finally,NWPU-RESISC45 dataset is adopted to input 3400 high-resolution remote sensing images into SPSSD model for iterative training,and mAP of the improved algorithm model finally reaches 92.68%,which is improved by 5.18 percentage point in comparison with that of the previous algorithm,and the detection speed reaches 25.1 frames per second.The experimental results show that the proposed method can effectively balance the detection accuracy and detection speed of aircraft targets,which reduces the miss detection rate to a certain extent.