Research On Hyperspectral Object Classification Based On Multi-scale Residual Network With Attention Mechanis

In recent years,with the rapid development and mature application of convolutional neural networks(CNN)in the field of image processing,it has also been introduced into the field of Hyperspectral Image(HSI)surface feature classification and become the mainstream surface feature classification method.Compared with traditional classification algorithms,CNN algorithm has the characteristics of automatic extraction of image features and self-learning update of parameters,and does not require complex pre-processing of the original data.However,the current mainstream classification algorithms usually have high computational complexity and large network parameters.In order to achieve the high precision classification of HSI and reduce the computational complexity and network parameters,three improved network models are proposed in this paper.(1)A high precision surface feature classification network model using multi-scale feature fusion strategy and residual connection.In order to fully extract HSI spectral-spatial features,we propose an end-to-end multi-scale feature fusion identity(MFFI)module.This module combines three methods of 3D multi-scale convolution,feature fusion and residual connection to realize the joint extraction of HSI multi-scale spectral spatial features.Due to the end-to-end characteristics of the module,the MFFI network with the ability to extract deep features can be obtained by stacking multiple MFFI modules.(2)A lightweight object classification model combining spectral,spatial attention and staggered attention mechanism,which ensures the network’s high accuracy classification performance while reducing the network’s computational complexity and parameters.First,we construct an additional hybrid branch for the extraction of spectral-spatial joint features,and take this feature as the supplementary information of pure spectral and spatial features.Then,spectral and spatial attention modules are introduced to enhance the corresponding domain features.Finally,the interleave-attention mechanism is designed to enhance the interaction between different domain features,and realize the full use of attention mechanism.(3)A lightweight dual channel residual architecture based on convolution optimization is proposed,which is only combined with spectral attention mechanism to ensure the high accuracy classification results and further reduce the network complexity.First,to further reduce the number of network parameters,we propose a convolution optimized dual channel residual module for joint spectral-spatial feature extraction.In this module,two size optimized 3D convolutions are used to extract spectral and spatial features respectively;At the same time,the residual connection based on convolution is also proposed to improve the network nonlinearity,so as to improve the generalization ability of the network.Finally,the improved spectral attention mechanism is also introduced to emphasize the spectral band information with high discrimination,so as to further improve the network performance.The classification results on Salinas(SA),Indian Pines(IP)and University of Pavia(PU)datasets show that the above network can achieve the high precision classification performance of HSI,while significantly reducing the computational complexity and parameters of the network.