Multi-scale Deep Network Based On Visual Attention For Optical Remote Sensing Image Classification

Optical remote sensing image processing has always been a significant research topic in the fields of urban planning,natural disaster prevention,and surface conditions and changes.And it also has a very important value in the military field.With the development of aerospace technology and the advancement of hardware systems,the optical remote sensing images also have higher resolution and more complex content.These high-quality optical remote sensing images have richer content,such as more texture information,more complex objects,and finer shapes.Moreover,remote sensing images are different from natural images in data sources,target structure and image background.Due to the rich content and large amount of current optical remote sensing image data,it is difficult for traditional machine learning methods to extract effective feature information.In addition,traditional machine learning methods require a lot of calculations for feature extraction,which makes it difficult to apply in actual scenarios with large amounts of data.In recent years,deep learning methods have been widely used in computer vision and natural language processing.The deep learning method is based on convolutional networks.Under the current conditions of massive data and super-powered computers,the deep learning method can automatically and accurately extract the features of the image data.At present,deep learning methods have achieved excellent results in many image processing fields such as image classification,target detection and semantic segmentation.Based on the current theory and combining the characteristics of optical remote sensing image data,this thesis introduces a multi-scale method in the convolutional neural network model,and makes full use of the attention mechanism to fuse and enhance the features in the model,further enhancing the deep learning method in optical remote sensing image classification task.The main research work of this thesis is as follows:(1)Aiming at the characteristics of target size changes in optical remote sensing images,a multi-scale deep learning model based on dilated convolution is proposed.Most of the traditional multi-scale methods are implemented by down-sampling and filtering,which are difficult to directly combine with deep learning methods.Dilated convolution can increase the receptive field of the convolution kernel without increasing the amounts of parameters,so that the model can flexibly deal with targets of different scales.The multiscale deep learning model introduces multiple sets of convolutions with different receptive fields in parallel to extract shallow features from optical remote sensing images,and then sends the extracted multi-scale features to the deep classification module to obtain the classification results.The experimental results show that this method has a good effect in optical remote sensing image scene classification tasks.(2)Aiming at the problem of feature fusion methods in convolutional neural networks,an adaptive feature fusion method based on channel attention mechanism is proposed.Currently,more feature fusion methods of direct addition or splicing along the channel are used.These methods are simple to operate,but often do not make full use of the information in the features.The adaptive feature fusion method in this thesis uses the channel attention mechanism to generate different weights for each channel of the two feature maps,and then adds the two feature maps with weights to obtain the fused feature map.Experimental results show that the adaptive method can more flexibly fuse two feature maps,and improve the distinguishability and robustness of model features,further improving the accuracy of optical remote sensing image scene classification.(3)Aiming at the global feature acquisition method in the current attention mechanism,a method of compressing the high-order feature to obtain the global feature is proposed and applied to the channel attention mechanism and the spatial attention mechanism.The current global feature acquisition methods in the attention mechanism all directly adopt the pooling method,which is simple and fast,but does not make full use of global information,and does not consider the relationship between features.The high-order feature compression method uses the covariance matrix to calculate the relationship between different features,and then generates the final compression result according to the covariance matrix.This method can make full use of high-order features,and improve the nonlinearity of the model,and strengthen the feature expression ability of the model.