Lightweight Hyperspectral Image Classification Based On Attention Mechanism

In recent years,the hyperspectral image(HSI)classification method based on deep convolutional neural networks(DCNNs)has gradually become the mainstream method of current research,and remarkable classification results have been achieved.However,the significant increase in classification accuracy has caused a sharp proliferation in the complexity of the model,which makes it challenging for the model with state-of-the-art performance to be applied in the actual scene.To address the above problems,this paper proposed two more lightweight and efficient models to reduce the complexity and time consumption of the model.The details are as follows:This paper first visualized the feature map of the existing model and summarized the principle of extracting the spatial information of the HSIs from the current model.Then we presented a lighter and more efficient spatial attention module based on the feature map visualization experiment,which effectively alleviates the problem of a large number of redundant calculations in the existing attention module.Secondly,this paper utilized depthwise separable convolution to achieve feature extraction and further reduce the number of model parameters.Thirdly,due to the model’s lightweight,the classification accuracy of "difficult to classify samples" is significantly reduced.This paper applied multiclass focal-loss to adjust the distribution of optimizing resources on each sample dynamically.Finally,based on Mobile Net V3 and the attention mechanism as the core,this paper exploited a relatively lightweight hyperspectral image classification model.The experimental results demonstrate that in the case of using significantly few training sets.The proposed model can tremendously reduce the number of calculations and parameters while maintaining high accuracy.In order to ensure the feature extraction capability of the model,the residual structure used in the above model needs to expand the number of channels of the input feature map,which inevitably leads to an increase in the amount of parameters contained in the residual structure.Besides,the lightweight residual structure used in existing models generally reduces the number of parameters by reducing the dimensionality of the input feature map,which may cause the loss of information.Therefore,this paper adopted the idea of "divide and conquer" to propose a new lightweight residual structure.The residual structure divides the input feature map into two branches,only one feature map is further extracted,and the other feature map is only used as supplementary information.In addition,the existing residual structure generally performs feature fusion by directly superimposing the input feature map and the output feature map,ignoring the inconsistency of information between the two feature maps.Therefore,this paper introduced the self-attention mechanism in the residual structure so that the residual structure’s input feature map and output feature map can be adaptively fused.Experimental results show that the residual model can remarkably reduce the complexity and time consumption and maintain a good classification result,even surpassing the current mainstream classification model.