Hyperspectral Image Classification Based On Fourier Transform Channel Attention Network

Hyperspectral imaging system using optical system and hyperspectral technology,to acquire hyperspectral data.These data cubes contain both spatial and spectral information.Using remote sensing satellites to obtain geographic remote sensing hyperspectral images can be used for environmental monitoring,vegetation protection and resource exploration.Combining microscopy imaging system and hyperspectral technology to obtain micro hyperspectral images of biological tissues can provide a basis for medical aided diagnosis.Effectively using spatial and spectral features of hyperspectral images for hyperspectral image classification is the basis of application.Convolutional neural network has achieved good results in image processing and has been applied in hyperspectral image recently.However,the rich spectral bands of hyperspectral images may have redundancy,resulting the decline of the accuracy.Although some networks based on channel attention mechanism can pay attention to important channels,there is a problem of insufficient information representation in processing channel features.This paper studies and constructs a new channel attention mechanism to effectively extract the channel features of hyperspectral images,improve the classification accuracy of convolutional neural network,and apply it to the classification of remote sensing hyperspectral images and cancerous region recognition of cholangiocarcinoma micro hyperspectral images.The main work and innovations of this paper are as follows:1.Fourier Transform channel attention(FTCA)is proposed.FTCA first performs two-dimensional Fourier Transform on the input feature map to extract frequency features,then obtains the channel weight vector through two full connection layers,multiplies the channel weight with the corresponding channel feature,finally obtains the output fused with channel attention information.Compared with other channel attention mechanisms such as Squeeze-Excitation,our method can extract features more effectively and improve the classification result.2.According to the characteristics of unbalanced distribution among categories and high spectral dimension of remote sensing hyperspectral datasets,dense connection is used to strengthen the propagation and reuse of spatial features,and FTCA is used to strengthen channel feature extraction,build FTCA dense network.In this paper,experiments are carried out on three public remote sensing hyperspectral datasets.The Overall Accuracy on Salinas dataset is 96.30%,kappa coefficient is 0.9589,on Paviau dataset is 95.45% and 0.9410,and on Botswana dataset is 93.48% and 0.9382.The experimental results are better than the comparison method.3.In this paper,laboratory made hyperspectral imaging system is used to collect and produce cholangiocarcinoma hyperspectral image datasets.According to different shapes and sizes of the cancerous areas,Inception module is used to obtain features with different receptive fields,and Focal loss is introduced to solve the problem of data imbalance between the cancerous area and the normal area.The Inception-FTCA network is constructed for cancerous region identification.The experimental results of this method in the hyperspectral dataset of cholangiocarcinoma are higher than those of the comparative method.The Accuracy,Precision,Sensitivity,Specificity and Kappa coefficient are 0.9780,0.9654,0.9586,0.9852 and 0.9456 respectively.Using this method on hyperspectral images has higher results than that on color images,which provides a promising new technology for cancer auxiliary diagnosis.