基于轻量化Yolact的遥感图像检测研究  

Research on remote sensing image detection based on lightweight Yolact

作　　者：方俞泽 宋亚男[1] 徐荣华[1] 戴昂 Fang Yuze;Song Yanan;Xu Ronghua;Dai Ang(School of Automation,Guangdong University of Technology,Guangzhou 510006,China)

机构地区：[1]广东工业大学自动化学院,广东广州510006

出　　处：《无线互联科技》2023年第10期117-121,共5页Wireless Internet Technology

基　　金：广东工业大学高水平大学建设研究生教育创新计划项目,项目名称:实践创新能力培养模式研究,项目编号:2018JGMS-09;广东省本科高校在线开放课程指导委员会研究课题,项目名称:多模态数据融合的学习分析与评价创新,项目编号:2022ZXKC143。

摘　　要：基于深度学习的遥感图像检测在农业生产、军事打击等领域都有所应用。但深度学习模型有计算复杂度高和参数量大的问题,而实际部署深度模型的边缘设备计算性能有限。文章以高分辨率遥感图像作为研究对象,对单阶段的实例分割网络算法进行改进,在Yolact(You Only Look At CoefficienTs)网络的基础上提出一种融入注意力机制和可变形卷积的轻量级实例分割算法。使用NWPU VHR-10遥感图像数据集对所提算法进行性能评估,实验结果表明,该算法能在保持性能的情况下减少计算复杂度和参数量。Remote sensing image detection based on deep learning has been applied in agricultural production,military attack and other fields.However,the deep learning model has the problems of high computational complexity and large number of parameters,while the actual deployment depth model has limited computational performance of edge devices.Taking high-resolution remote sensing images as the research object,this paper improves the one-stage instance segmentation network algorithm.Based on the Yolact(You Only Look At CoefficienTs)network,a lightweight instance segmentation algorithm is proposed which integrates the attentional mechanism and deformations convolution.The performance of the proposed algorithm is evaluated using the NWPU VHR-10 remote sensing image dataset,and the experimental results show that the algorithm can reduce the computational complexity and the number of parameters while maintaining the performance.

关 键 词：深度学习 Yolact 轻量化 注意力机制 可变形卷积 

分 类 号：TP39[自动化与计算机技术—计算机应用技术]