基于改进YOLOv5n的轻量化光学遥感图像目标检测

Lightweight Optical Remote Sensing Image Object Detection Based on Improved YOLOv5n

针对光学遥感图像检测时存在背景复杂、目标密集、检测模型参数量和浮点计算量大的问题,提出了基于改进YOLOv5n的轻量化光学遥感图像目标检测方法。在主干网络中通过减少部分模块数量、引入无参数的SimAM注意力机制和Bottleneck Transformer结构,在参数量降低的情况下增强了网络的特征提取能力;引入Ghost卷积和使用同层跨越连接设计特征融合网络,降低模型复杂度和融合更多的特征信息;使用SIoU损失函数加快收敛速度和提升模型精度。通过在NWPU VHR-10和RSOD公开数据集上的实验结果表明,改进后的算法与原算法相比,平均精度均值(mAP0.5)分别提升了2.4%和1.5%,并且参数量减少25.4%,GFLOPs减少了21.4%。

To deal with the problems of complex background,dense objects,and large amount of detection model parameters and floating-point computation in optical remote sensing image detection,a lightweight optical remote sensing image object detection method based on improved YOLOv5n is proposed.Firstly,by reducing the number of modules in the backbone network and introducing the non-parameter SimAM attention mechanism and Bottleneck Transformer structure,the feature extraction capability of the network is enhanced under the condition of reduced number of parameters.Secondly,Ghost convolution and same-layer spanning connections are introduced to design a feature fusion network and reduce model complexity and fuse more feature information.Finally,SIoU loss function is used to speed up convergence and improve model accuracy.The experimental results on the NWPU VHR-10 and RSOD public datasets show that,compared with the original algorithm,the average precision(mAP0.5)of the improved algorithm is increased by 2.4%and 1.5%respectively,and the number of parameters is reduced by 25.4%,and FLOPs decreased by 21.4%.