基于多任务学习的高光谱图像目标分类和分割方法  

Hyperspectral Image Target Classification and Segmentation Method Based on Multi-task Learning

作　　者：韩宇霖 刘凯新 陈平[1,2] HAN Yulin;LIU Kaixin;CHEN Ping(The State Key Lab for Electronic Testing Technology,North University of China,Taiyuan 030051,China;Department of Information and Communication Engineering,North University of China,Taiyuan 030051,China)

机构地区：[1]中北大学电子测试技术国家重点实验室,太原030051 [2]中北大学信息与通信工程学院,太原030051

出　　处：《计算机测量与控制》2025年第4期241-246,254,共7页Computer Measurement &Control

基　　金：国家优秀青年科学基金(62122070);国家自然科学基金-青年基金(62301507)。

摘　　要：高光谱图像具有丰富的光谱特征和空间特征,针对传统的基于卷积神经网络的高光谱图像分类方法只能完成单一的任务,缺乏对其他任务的泛化能力的问题,提出基于多任务学习的高光谱图像分类和分割方法,通过共享统一的网络来组合完成不同的任务;该框架采用共享编码器来有效提取多尺度特征,同时在主干网络中引入光谱通道注意力以实现空间和光谱特征的联合提取,并使用两个特定于任务的解码器来获取不同任务的结果;所提出的方法在高光谱图像数据集上进行两项任务的对比实验,结果表明,与其他方法相比,该方法实现了最好的性能。Hyperspectral images have rich spectral and spatial features.Traditional hyperspectral image classification methods based on convolutional neural networks can only complete a single task and lack generalization ability for other tasks.In this paper,a multi-task learning hyperspectral image classification and segmentation method is proposed to achieve different tasks by sharing a unified network.The framework uses a shared encoder to efficiently extract multi-scale features,while introducing spectral attention in the backbone network to achieve joint extraction of spatial and spectral features,and uses two task-specific decoders to obtain results for different tasks.The proposed method performs comparative experiments of two tasks on hyperspectral image datasets.The results show that compared with other methods,the proposed multi-task learning method achieves the best performance.

关 键 词：高光谱图像 多任务学习 语义分割 特征融合 注意力机制 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]