低分辨率暗弱光斑图像的目标识别技术研究

Research on target recognition technology for low resolution dim spot images

针对新一代激光雷达对远距离、高速运动目标实现超快发现、检测与识别的需求,解决自然环境多变、目标暗弱且高速运动导致图像分辨率低的问题,鉴于传统光学和传统网络无法对目标实现高精准的识别,提出低分辨率暗弱光斑图像的深度层次轮廓识别网络LRDSI-DLCRN,该网络引入全局权重编码模块,采用子像素卷积进行上采样,丰富了不同层次边缘结构特征的相关性,在公开数据集PASCAL VOC 2012和真实环境采集的Spotcraf数据集上的效果都优于其它流行算法。

In response to the demand of the new generation of LiDAR for ultra fast detection,detection,and recognition of long-distance and high-speed moving targets,and to solve the problem of low image resolution caused by the changing natural envi-ronment,dim targets,and high-speed motion,traditional optics and networks cannot achieve high-precision recognition of targets.Therefore,a deep level contour recognition network LRDSI-DLCRN for low resolution dim spot images is proposed,The network in-troduces a global weight encoding module and uses sub pixel convolution for upsampling,enriching the correlation of edge struc-ture features at different levels.The performance on the public dataset PASCAL VOC 2012 and the real environment collected Spotcraf dataset is superior to other popular algorithms.