摘要
针对传统遥感图像数据分析算法存在鲁棒性较差、适应度与计算效率均偏低的问题,文中基于YOLOv3提出了一种轻量化的遥感图像数据分析算法。该算法使用YOLOv3作为神经网络模型的框架,并将内部的Darknet-53多尺度卷积作为主网络。为了减小主网的冗余度,通过SE-Net模型连接网络的深层与浅层卷积,在轻量化的同时也增强了模型的深度特征提取能力。同时,根据改进后网络的权重输出结果,采用剪枝算法对卷积核进行简化,进而完成了模型的轻量化。在实验测试中,轻量化后的模型可显著提升FPS(Frames Per Second)值,且算法的mAP指标为93.25%,在对比算法中为最优,表明了算法模型的有效性及其性能的优越性。
Aiming at the shortcomings of traditional remote sensing image data analysis algorithms,such as poor robustness,low fitness and low computational efficiency,this paper proposes a lightweight remote sensing image data analysis algorithm based on YOLOv3.The algorithm uses YOLOv3 as the neural network model framework,and uses the internal Darknet-53 multi⁃scale convolution as the main network.In order to reduce the redundancy of the main network,the SE Net model is used to connect the deep convolution and shallow convolution of the network,which not only reduces the weight,but also improves the depth feature extraction ability of the model.At the same time,according to the weight output results of the improved network,we use pruning algorithm to simplify the convolution kernel and complete the model lightweight.In the experimental test,the lightweight model can significantly improve the FPS value.At the same time,the algorithm mAP index is 93.25%,which is also the best in the comparison algorithm,indicating the effectiveness and performance superiority of the algorithm model.
作者
李增顺
刘勇
侯雪蕊
吴松
管守标
LI Zengshun;LIU Yong;HOU Xuerui;WU Song;GUAN Shoubiao(Hebei Zhongke SinoStar Information Technology Co.,Ltd.,Shijiazhuang 050000,China;Hebei AiMap Information Technology Co.,Ltd.,Qinhuangdao 066004,China)
出处
《电子设计工程》
2024年第11期183-187,共5页
Electronic Design Engineering
基金
国家国防科技工业局重大专项(67-Y50G05-9001-22/23)。
