摘要
目前现有的大部分方法对细粒度遥感船只检测识别精度较低,并且星载计算机算力有限,常用的浮点精度数据类型所带来的大量计算和存储需求使其难以满足模型在轨部署的需求。面向这些挑战,提出了一种基于模型量化的细粒度遥感船只快速目标检测方法。首先设计了一种基于融合智能的检测网络,解决了“类内差异大、类间差异小”的难题,可有效提高细粒度船只检测识别的准确度。在此基础上,进一步提出了一种高精度的模型量化方法对裁剪边界实现了优化,可有效提升在轨遥感图像检测识别速度。在多个数据集上的测试表明,所提出检测方法相比于现有研究实现了超过5.9%的最大精度提升,同时量化方法可实现1.2%的最大性能提升,可在降低模型计算量的同时保持较高的精度,可适用于星载计算机的应用。
Most existing methods have low recognition accuracy for fine-grained remote sensing vessel detection,and the large computation and storage requirements associated with the commonly used floating-point precision data types make it difficult to meet the needs of model in-orbit deployment due to the limited power of on-board devices.To address these challenges,this paper proposed a fast target detection method for fine-grained remote sensing vessels based on model quantization.Firstly,a fusion intelligence-based detection network was designed to solve the problem of“large intra-class differences and small inter-class differences”,which can effectively improve the accuracy of fine-grained vessel detection and identification.On this basis,a high-precision model quantization method was proposed to optimize the clipping boundary,which could effectively improve the inference speed.Experimental test results show that the proposed method achieves a maximum accuracy improvement of more than 5.9% compared with existing studies,while the quantization method can achieve a maximum performance improvement of 1.2%.It can effectively reduce the calculation load while maintaining a high accuracy,thus can be easily applied to satellite-based computing units.
作者
王海涛
贺治钧
周天启
马岳
WANG Haitao;HE Zhijun;ZHOU Tianqi;MA Yue(Satellite Application Department of China Academy of Space Technology,Beijing 100094,China;Institute of Telecommunication and Navigation Satellites,China Academy of Space Techology,Beijing 100094,China)
出处
《中国空间科学技术(中英文)》
北大核心
2025年第1期153-161,共9页
Chinese Space Science and Technology
关键词
卫星遥感船只检测
快速目标检测
CNN模型量化
卫星应用
深度神经网络
satellite remote sensing vessel detection
fast target detection
CNN model quantification
satellite applications
deep neural networks
