基于多尺度非对称密集网络的高光谱图像分类

Hyperspectral Image Classification Based on Multi-scale Asymmetric Dense Network

近年来,基于有限标记样本的高光谱图像(HSI)分类方法取得了重大进展。然而,由于高光谱图像的特殊性,冗余的信息和有限的标记样本给提取强判别特征带来了巨大挑战。此外,由于各类别像素分布不均,如何强化中心像素的作用,减弱不同类别的周围像素的负面影响也是提高分类性能的关键。为了克服上述局限性,该文提出一种基于多尺度非对称密集网络(MS-ADNet)的高光谱图像分类方法。首先,提出一个多尺度样本构建模块,通过在每个像素周围提取多个尺度的图像块,并进行反卷积和拼接以构建输入样本,使其既包含详细的结构区域,又包含较大的同质区域;然后,提出一个非对称密集连接结构,在空间和光谱特征联合提取中实现核骨架增强,即增强了方形卷积核的中心十字区域部分提取的特征,有效地促进了特征重用。此外,为了提高光谱特征的鉴别性,提出一种精简的元素光谱注意力机制,并将其置于密集连接网络的前端和后端。在每类仅采用5个样本进行网络训练的情况下,该方法在Indiana Pines, Pavia University和Salinas数据集上的总体准确率分别达到了77.66%, 84.54%和92.39%,取得了极具竞争力的分类结果。

HyperSpectral Image(HSI)classification methods based on limited labeled samples have made significant progress in recent years.However,due to the specificity of hyperspectral images,redundant information and limited labeled samples pose great challenges for extracting highly discriminative features.In addition,owing to the uneven distribution of pixels in each category,how to strengthen the role of central pixels and attenuate the negative impact of surrounding pixels with different categories is also the key to improve the classification performance.To overcome the above limitations,an HSI classification method based on Multi-Scale Asymmetric Dense Network(MS-ADNet)is proposed.Firstly,a multi-scale sample construction module is proposed,which extracts multiple scale patches around each pixel and performs deconvolution and stitching to construct multiscale input samples that contain both detailed structural regions and large homogeneous regions.Next,an asymmetric densely connected structure is proposed to achieve kernel skeleton enhancement in joint spatial and spectral feature extraction,i.e.,enhancement of features extracted from the central cross-skeleton portion of a square convolutional kernel,which effectively facilitates feature reuse.Moreover,to improve the discriminability of spectral features,a streamlined element spectral attention mechanism is proposed and placed at the front and back ends of the densely connected network.With only five samples per class used for network training,the proposed method achieves competitive classification results with overall accuracies of 77.66%,84.54%,and 92.39%on the Indiana Pines,Pavia University,and Salinas datasets,respectively.