Lightweight Hyperspectral Recognition Network Based On Attention Mechanism

Hyperspectral data has a narrow sensitive frequency band and complex and diverse spatial structure.Conventional machine learning methods for ground object classification are often limited by their insufficient ability to express sensitive information and spatial structure in the spectral domain.In recent years,convolutional neural network has integrated spatial information on the basis of spectral information,which is helpful for the recognition of hyperspectral images.However,the development of convolutional neural network also exposed many problems: in order to obtain higher recognition accuracy,the model is often designed in depth,resulting in the growth of network parameters and the reduction of network learning efficiency,which has high requirements for hardware equipment and hardware resources.However,hyperspectral remote sensing data has many wavebands and consumes a large amount of computing resources for information processing.Therefore,how to quickly and accurately identify hyperspectral images with deep learning model has become a key issue that needs urgent consideration.In order to improve the feature extraction and hyperspectral image recognition functions of the traditional convolutional neural network,a lightweight convolutional neural network architecture based on deep separable convolution was proposed in combination with the idea of lightweight network.The lightweight of the convolutional neural network is realized by using the multi-scale improved network with deep convolution,point-by-point convolution and channel mix-wash method.Split over a channel to fusion way to improve the light in the network information interaction between channel characteristics graph,and USES the public data sets through experiment contrast a variety of network structure analysis model,the experimental results show that the use of lightweight network to ensure network identification precision of the reduced network model and quantity at the same time,processing speed of network also has a further increase.Aiming at improving the feature extraction and expression ability of traditional convolutional neural network,and combining the idea of attention mechanism,a convolutional neural network architecture based on attention mechanism is proposed.By comparing improved networks with added channel attention,spatial attention,and mixed attention,experiments using open datasets were conducted to compare various structures and analyze the models.The experimental results show that the attention mechanism is helpful for the network to extract and express the image features.While reducing model parameters,lightweight network will lose part of feature extraction ability.By combining the attention mechanism to make up for the feature expression ability of the network,an attention mechanism network feature extraction scheme based on the salience of the target is proposed.Taking the significance of the target as the attention factor in the learning process of the network,the channel attention scheme,spatial attention scheme and mixed attention scheme were designed to enhance the representation of the network feature map and increase the retention of useful information.On the basis of the proposed lightweight module,the improved attention structure was compared with the existing attention structure,and it was found that the proposed attention structure could be applied to the task of hyperspectral remote sensing image recognition while maintaining the network lightness.