基于改进YOLOv3的空间非合作目标部件识别算法

Spatial Non-cooperative Target Components Recognition Algorithm Based on Improved YOLOv3

在使用深度学习方法进行空间非合作目标部件识别时,由于神经网络参数量大且嵌入式设备计算能力不足,难以将神经网络有效地部署在嵌入式平台上。针对该问题,文中提出了一种改进的轻量化目标检测网络,在保证检测精度的同时,有效降低网络参数量,提升了网络检测速度。所提网络模型在YOLOv3的基础上借鉴深度可分离卷积的思想,引入Bottleneck模块降低了模型参数量,提升了检测速度,同时引入Res2Net残差模块来增加模型的感受野尺度丰富性和结构深度,提高了网络对于小目标的检测能力。设计了一个新的轻量化特征提取主干网络Res2-MobileNet,并结合多尺度检测方法进行空间非合作目标部件识别。实验结果表明,相比YOLOv3,所提模型在参数量上降低了55.5%,检测速度由34fps提高到65 fps,同时对于小目标的检测效果也有显著提升。

Due to the large amount of neural network parameters and insufficient computing power of embedded devices,it is difficult to effectively deploy neural networks on embedded platforms when using deep learning methods to identify spatial non-cooperative target components.Aiming at this problem,an improved lightweight target detection network is proposed in this paper.On the basis of YOLOv3,a new lightweight feature extraction backbone network Res2-MobileNet is designed,drawing on the ideaof Depth Separable Convolution,introducing the Bottleneck module to reduce the amount of model parameters to improve the detection speed,and introducing the Res2 Net residual module to increase the sensitivity of network to small targets by increasing the model’s receptive field scale richness and structural depth,and combines multi-scale detection methods to recognize spatial non-cooperative target components.The experimental results show that compared with the YOLOv3 model,the size of this model is reduced by 55.5%,the detection speed is increased from 34 fps to 65 fps,and the detection effect for small targets is also significantly improved.