基于改进EfficientDet算法的可见光遥感舰船目标检测  

Visible light remote sensing ship target detection based on improved EfficientDet algorithm

作　　者：刘浪[1] 刘国栋[1] 刘佳[1] LIU Lang;LIU Guodong;LIU Jia(School of Civil Engineering,Chongqing Jiaotong University,Chongqing 400074,China)

机构地区：[1]重庆交通大学土木工程学院,重庆400074

出　　处：《现代电子技术》2022年第22期28-32,共5页Modern Electronics Technique

摘　　要：为了有效地解决可见光遥感影像中舰船目标难检测、易错检等问题,文中提出一种基于改进EfficientDet的舰船目标检测算法E-EfficientDet。首先利用K-means++聚类算法对舰船数据集中的目标物体的长宽信息进行聚类分析,得到适合舰船目标的Anchors;然后,针对SENet模块参数量大和降维处理使特征图通道之间的依赖性减小的问题,使用ECANet模块代替EfficientNet网络中SENet模块;其次,为了解决遥感图像中相似目标对船只检测的干扰及小目标难检的问题,采用EfficientNet-B0、EfficientNetV2网络作为E-EfficientDet模型的主干特征提取网络,将获取到的特征送入到改进的BiFPN网络中进行特征融合,获取更加丰富有效的船只目标特征信息;最后,为了扩大图像的响应区域并避免骨干网络对输入图片进行特征提取时出现的信息丢失情况,利用SPPNet网络对特征图进行不同尺度的最大池化。实验结果表明,E-EfficientDet算法对文中数据检测的平均精度(AP)达到90.18%,每张图像的检测时间为0.06 s。另外,将E-EfficientDet算法与Faster RCNN、SSD、YOLOv3算法进行对比,得到所提算法的AP精度均高于其余算法,说明所提算法对实际遥感场景下舰船的检测性能较好。A ship target detection algorithm E-EfficientDet based on improved EfficientDet is proposed to effectively solve the problems of difficult detection and error pickup of ship targets in visible remote sensing images.The K-means++clutsering algorithm is used to cluster the length and width information of the target object in the ship data set to obtain the Anchors suitable for the ship target.In allusion to the large number of parameters of the SENet module and the reduced dependence between feature map channels caused by the dimension processing,the ECANet module is used to replace the SENet module in the EfficientNet network.In order to solve the problem of the interference of similar targets in remote sensing images on ship detection and the difficulty of small target detection,EfficientNet-B0 and EfficientNetV2 networks are used as the backbone feature extraction network of the E-EfficientDet model to sent the acquired features to the improved BiFPN network for feature fusion,so as to obtain more abundant and effective ship target feature information.In order to expand the response area of the image and avoid the information loss of the backbone network for feature extraction of the input image,the feature map is maximally pooled at different scales by means of SPPNet network.The experimental results show that the average accuracy(AP)of the E-Efficient Det algorithm for detecting the data in the text reaches 90.18%,and the detection time of each image is 0.06 s.the comparison expriments between E-Efficient Det algorithm and Faster RCNN,SSD and YOLOV3 algorithms show that the AP accuracy of the proposed algorithm are higher than those of other algorithms,which verifies that the proposed algorithm has a good performance for ship detection in real remote sensing scenes.

关 键 词：K-means++ E-EfficientDet 舰船目标检测 ECANet SPPNet 数据集增强 AP 

分 类 号：TN929.1-34[电子电信—通信与信息系统]