摘要
水下目标识别在海洋探索过程中扮演着非常重要的角色,高光谱图像(hyperspectral image,HSI)通过叠加数百个连续的波段提供了丰富的光谱空间信息,如何在丰富的高光谱水下信息中准确提取目标信息成为一项挑战。通常通过使用具有固定大小感受野(receptive field,RF)的卷积神经网络(convolutional neural network,CNN)来解决,然而,当使用前向传播和后向传播来优化网络时,这些解决方案无法使神经元有效调整感受野大小并且具有跨通道依赖性。文章提出了一种基于空-谱残差网络的无人机高光谱水下目标分类识别算法,该网络具有频谱注意力,实现自适应感受野,能够以端到端的训练方式捕捉用于人机交互分类的辨别性频谱空间特征。首先,采用SG(Savitzky-Golay)平滑处理,消除噪声所引起的光谱曲线高频抖动,保留光谱曲线有效峰谷形貌,提升后续高光谱处理精度。之后,对降噪后的光谱图像采用主成分分析法(principal components analysis,PCA)进行数据降维,提取有效数据。最后,采用空谱transformer网络(spectral-spatial transformer,SST)进行像素分类,确定水下目标所在的像素位置。使用高光谱无人机采集水下目标数据,分别在0.1 m、2 m、3 m和5 m的目标深度采集了无人机高光谱数据。实验结果表明,所提出的算法能够准确地识别水下目标。
Underwater target recognition plays a very important role in the process of ocean exploration.Hyperspectral image(HSI)provides rich spectral spatial information by stacking hundreds of continuous bands.How to accurately extract target information from rich hyperspectral underwater information has become a challenge.This is usually solved by using convolutional neural network(CNN)with a fixed size receptive field(RF).However,when using forward and back propagation to optimize the network,these solutions fail to enable neurons to efficiently adjust receptive field sizes and cross-channel dependencies.In this paper,a hyperspectral underwater target classification and recognition algorithm for UAV based on spatial-spectral residual network is proposed.The network has spectral attention,realizes adaptive receptive field,and can capture discriminative spectral spatial features for human-computer interaction classification in an end-to-end training manner.Firstly,the SG(Savitzky-Golay)smoothing process is used to eliminate the high-frequency jitter of the spectral curve caused by noise,retain the effective peak and valley morphology of the spectral curve,and improve the accuracy of subsequent hyperspectral processing.Then,the principal components analysis(PCA)method is used to reduce the dimension of the denoised spectral image and extract the effective data.Finally,the spectral-spatial transformer(SST)network is used for pixel classification to determine the pixel position of the underwater target.In this paper,a hyperspectral UAV is used to collect underwater target data,and the UAV hyperspectral data are collected at target depths of 0.1 m,2 m,3 m,and 5 m,respectively.The experimental results show that the proposed algorithm can accurately identify underwater targets.
作者
顾程鑫
张彬
祝江敏
潘明忠
GU Chengxin;ZHANG Bin;ZHU Jiangmin;PAN Mingzhong(School of Physics and Optoelectronic Engineering,Hangzhou Institute for Advanced Study,USTC,Hangzhou 340024,China;College of Aviation Combat Service,Naval Aeronautical University,Yantai,Shandong 264000,China)
出处
《遥感信息》
CSCD
北大核心
2024年第3期128-135,共8页
Remote Sensing Information
基金
国科大杭州高等研究院自主立项项目(2022ZZ01008)。
作者简介
顾程鑫(2000-),男,硕士研究生,主要研究方向为高光谱降维、分类算法及硬件加速。E-mail:602151111@qq.com;通信作者:潘明忠(1982-),男,研究员,主要研究方向为高光谱成像技术、电子学技术。E-mail:mzpan@ucas.ac.cn。
