SAR多尺度海洋大气环境与目标分类  

SAR Multi-scale Marine Atmospheric Environment and Target Classification

作　　者：冯运 Feng Yun(Faculty of Engineering,Anhui Sanlian University,Hefei 230601,China)

机构地区：[1]安徽三联学院工学部,合肥230601

出　　处：《塑料包装》2025年第2期76-79,83,共5页Plastics Packaging

基　　金：安徽三联学院2024年度校级自然科学研究项目:“基于轻量级深度学习的SAR图像目标分类模型研究”(项目编号:KJYB2024001);安徽省教育厅自然基金重点项目:基于WiFi信道状态信息的人体呼吸波形感知算法研究(项目编号:2024AH050518)。

摘　　要：海洋是人们赖以生存和发展的空间,为了能更好地利用和保护海洋,需要对海洋大气环境和目标进行可靠有效的监测。近年来,基于深度神经网络DNN的方法在海洋大气环境分类中发挥越来越重要的作用。然而,现实环境复杂多变且存在未知类别,为了更好的对海洋大气环境中的目标进行监测分类,经过大量调研,了解到基于DNN的方法通常以高置信度将未知类样本分类为已知类别之一,从而提高分类性能。基于此,提出一种GESAR方法,具体来说,提出一种GeneralizedENtropyScore(GEN),可以提高模型预测未知类的分类性能,增加模型泛化性;同时,为了解决类内相似性小,类间差异性大的问题,还引入了梯度增强损失引导网络关注类内相似性和类间差异性,引导模型学习细粒度鉴别性特征。实验结果表明,在SAR多尺度海洋大气环境与目标分类数据集上验证了本文所提方法的有效性。The ocean is the space that people rely on for survival and development.In order to better utilize and protect the ocean,reliable and effective monitoring of the marine atmospheric environment and targets is needed.In recent years,methods based on deep neural network s(DNNs)have played an increasingly important role in the classification of marine atmospheric environments.However,the real environment is complex and ever-changing,and there are unknown categories.In order to better monitor and classify targets in the marine atmospheric environment,after extensive research,it has been found that DNN based methods usually classify unknown class samples into one of the known categories with high confidence,thereby improving classification performance.Based on this,a GESAR method is propose d,specifically a Generalized ENtropy Score,which can improve the classification performance of the model in predicting unknown classes and increase the model's generalization ability;At the same time,in order to solve the problem of small intra class similarity and large inter class difference,gradient enhancement loss is introduced to guide the network to focus on intra class similarity and inter class difference,and guide the model to learn fine-grained discriminative features.The experimental results demonstrate the effectiveness of the proposed method on the SAR multi-scale ocean atmospheric environment and target classification dataset.

关 键 词：SAR 海洋环境 目标分类 深度神经网络 细粒度特征 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]