针对SAR图像船舶检测任务在船舶组合和船舶融合场景下低检测精度的问题,提出了一种轻量化船舶检测算法——RGDET-Ship,有效提高了SAR图像在复杂场景下的船舶检测精度.该算法的创新点包括:①构建基于改进ResNet的基础主干网络,增强深浅网络早特征融合,保留更丰富的有效特征图,并利用RegNet进行模型搜索得到一簇最优结构子网络RegNet and Early-Add(RGEA),实现模型的轻量化;②在FPN Neck基础上,结合EA-fusion策略设计出FPN and Early Add Fusion(FEAF)Neck网络,进一步加强深浅特征晚融合,提高中大船舶目标特征的提取;③通过细粒度分析改进RPN网络得到Two-RPN(TRPN)网络,提高模型的检测粒度和预测框准确性;④引入多任务损失函数——Cross Entropy Loss and Smooth Ll Loss(CE_S),包括分类任务和回归任务,进一步提升检测性能.通过在标准基准数据集SSDD上进行大量实验,验证了RGDET-Ship模型的有效性和健壮性.实验结果表明,相较于Faster RCNN和Cascade RCNN,RGDET-Ship在mAP_0.5:0.95上分别提升了5.6％和3.3％,在AR上分别提升了9.8％和7.6％.
Lightweight SAR Image Ship Detection Algorithm with Enhanced Feature Fusion and Refinement
To solve the problem of low detection accuracy of SAR image ship detection tasks in ship combination and fusion scenarios,a lightweight ship detection algorithm known as RGDET-Ship is proposed to effectively improve the ship detection accuracy of SAR images in complex scenes.The innovations of the algorithm include:① constructing a foundational backbone network based on an enhanced ResNet to enhance early feature fusion of deep and shallow networks and retain richer and more effective feature maps.Furthermore,employing RegNet to search for the model and obtain a cluster of optimal structural subnetworks known as RegNet and Early-Add(RGEA),achieving lightweighting of the model;② building upon the FPN Neck,the FPN and Early Add Fusion(FEAF)Neck network is designed by integrating the EA-fusion strategy,further enhancing late fusion of deep and shallow features.This augmentation leads to improved extraction of features from medium to large-sized ship targets;③ the RPN network is enhanced by means of fine-grained analysis to obtain the Two-RPN(TRPN)network,thus improving the model's detection granularity and prediction box accuracy;④ introducing a multi-task loss function known as Cross Entropy Loss and Smooth Ll Loss(CE_S),including classification and regression tasks,to further enhance detection performance.The effectiveness and robustness of the RGDET-Ship model are validated through extensive experiments on the standard benchmark dataset SSDD.The experimental results show that compared to Faster RCNN and Cascade RCNN,RGDET-Ship achieves improvements of 5.6％and 3.3％in mAP_0.5:0.95,and enhancements of 9.8％and 7.6％in terms of AR.
ship detectiondeep and shallow feature fusionfine-grained designRGDET-Ship
万方数据
维普
