基于多任务学习的高光谱图像目标分类和分割方法

Multi-task learning hyperspectral image target classification and segmentation method based on CNN

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,所提出的多任务学习方法实现了最好的性能。

Abstract:

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 combine different tasks by sharing a unified network. The framework uses a shared encoder to extract multi-scale features efficiently, 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 the proposed multi-task learning method achieves the best performance compared with other methods.