基于改进的胶囊网络模型的高光谱图像分类方法

针对卷积神经网络无法有效提取高光谱图像光谱与空间特征以及识别特征之间的空间位置问题,提出一种基于胶囊网络的改进神经网络模型.采用1×1卷积核对高光谱图像块进行降维处理;利用双通道卷积神经网络提取降维图像的初级特征,进而在PrimaryCaps层将初级特征信息封装为胶囊向量;通过DigitCaps层计算胶囊向量的模长来判...

HYPERSPECTRAL IMAGES CLASSIFICATION METHOD BASED ON IMPROVED CAPSULE NETWORK MODEL

Convolutional neural network usually cannot effectively extract spectral and spatial features of hyperspectral images and identify the spatial position between features.To overcome these problems,an improved neural network model based on capsule network is proposed.It reduced the dimension of the hyperspectral image block with 1×1 convolution kernel firstly;the primary features of dimensional reduction image were extracted by using two-channel convolutional neural networks and encapsulated into a capsule vector in the PrimaryCaps layer;the category of the center pixel of the image block was determined by the modulus length of the capsule vector in the DigitCaps layer.To express the feasibility and validity,the improved model was applied to classify two hyperspectral data sets,Indian Pines and Parvia University,and compared with other classified models.The experimental results show that the presented model has better generalization ability,which can effectively extract image features and identify spatial position between features,thereby the classification accuracy can be improved.