The Research Of Information Extraction From VHR Image Based On Deep Learning Model With Attention Mechanism

With the development of artificial intelligence and deep learning.Semantic segmentation of images based on deep learning has become an important tools of remote sensing image extraction.Although the information extraction method of remote sensing based on deep learning semantic segmentation has achieved some success.In deep neural networks,the size of the receptive field can roughly indicate the degree of usage of contextual information.But in the deep neural networks,the practical receptive field of a CNN is much smaller than the theoretical receptive field,especially in high-level layers.This condition results in many networks being unable to fully integrate important global context information.which reduce the model's ability to perceive the scene and weakening the consistency of object classification in the scene.In addition,the fully convolutional network uses the the pooling layer to down-sample,which allows the network to see large contextual information,but the high-frequency details will be losed.This makes the boundary of the result blurred by semantic segmentation of the fully convolutional network.In order to solve the loss of high-frequency details caused by down-sampling,some scholars use the dilated convolutions to instead the pooling layer,but this method can not overcome the problem of "grid effect"(discontinuous feature information).Others use skip-connections to alleviate boundary ambiguity.Although the skip-connections connections can reintroduce high-frequency details after sampling,but it can also introduce a lot of redundant information into the network and cause over-segmentation problems.In addition,compared with the images from other perspectives,the objects in the remote sensing image are more difficult to distinguish,which also increases the difficulty of extracting information from the remote sensing image.Multi-source data fusion is an effective means to solve this problem.In the extraction of remote sensing information from multi-source data fusion,the method of simple stacked image and decision fusion do not perform well in practice.The method of feature-based fusion alternately trains multi-branch models with large computation,while the end-to-end model can not transmit information vertically and the feature can not be further fused horizontally.This study is aimed at the problem of deep learning image semantic segmentation model.Using attention mechanism can help humans to filter out effective information from massive information and suppress the characteristics of background and invalid information.Combining the attention mechanism with the deep learning model,the attention mechanism can help the deep learning model to better understand the external information,and solve the important global context information loss,over-segmentation,and multi-source existing in the current deep learning image semantic segmentation model.The problem of data features being difficult to deeply fuse.The specific research includes the following aspects:(1)Use the global attention mechanism to obtain effective information characteristics from global information.Combining the global attention mechanism with the deep learning image semantic segmentation model,a deep learning semantic segmentation model with global context awareness is designed to integrate the important global context information.And then to improve the accuracy of land use classification of remote seneing based on deep learning semantic segmentation model.(2)In order to solve the problem of reducing the classification information of high-level features,skipping-connection are used to restore high-frequency detail features in semantic segmentation of deep learning,which results in the multiple use of low-level features.In this paper,spatial attention mechanism is introduced into the upper sampling process of deep learning semantics segmentation model to guide high-level features to recover effective detail information,compress background and noise information,and then improve the problem of over-segmentation,improve the integrity and accuracy of information extraction from high-resolution remote sensing images.(3)In order to solve the problem that information can not be transmitted vertically and feature can not be further fused horizontally in multimodal data fusion.In this paper,the process of human vision in dealing with external information is combined with deep learning model.The process of human visual perception,firstly perceives various primary features of external information in parallel,then integrates various features perceived in the pre-attention stage to form semantic target features,so as to realize the process of comprehensive discrimination of semantic objectives.A deep learning network with pre-attention and feature integration is studied,which simulates the way human eyes process external information to extract and fuse features of multimodal data.So it is an effective method to realize the deep fusion of multi-source features.