Research On The Application Of Convolution Restricted Boltzmann Machine In Remote Sensing Image Processing

As a basic application of image processing,image target extraction refers to recognition or interpretation the meaningful objects from a single image or sequence images.Image scene classification refers to recognition different kinds of scene images based on the scene features of the image.How to extract object and classify scenes from high-resolution remote sensing images has become a research hotspot in geological monitoring,geographic resource survey,urban planning and computer vision.In image segmentation tasks,shadow,occlusion and various interference in the image increase the difficulty of segmentation,that lead to unsatisfied results.Because the deep learning model has the ability to automatically propose features from massive images,use the deep learning method becomes a popular trend.Above all,this thesis studies the deep learning image segmentation theory and image scene classification theory,and explores the application of deep learning on remote sensing image segmentation and scene classification methods.The contents of this paper are as follows:(1)In order to solve the problem that segmentation results are easily affected by target occlusion and complex background in traditional segmentation methods,a CV segmentation model combined with Convolution Restricted Boltzmann Machine is proposed.By introducing the prior knowledge related to the object shape in the CV model segmentation method,an image segmentation algorithm combining CV model with deep learning shape priori method is proposed.The target shape priori information is modeled and generated using Convolution Restricted Boltzmann Machine.Then the energy function in CV model is constrained by the added prior shape term.Better segmentation results are obtained in remote sensing datasets Satellite-2000 and Vaihigen,whose target shapes and sizes are different using limited training data.Specifically,compared with CV model,the average global accuracy(Global ACC)of the proposed model on satellite-2000 and vaihigen data sets is increased to 97.670%and 91.769%,respectively.The average Intersection over Union(IOU)of satellite-2000 and vaihigen increased to 92.685%and 86.769%,respectively.(2)In order to solve the problem of small target segmentation in remote sensing images with complex background,for targets with a size of less than 2000 pixels,a CV small target segmentation model based on edge and Convolution Restricted Boltzmann Machine is proposed in this paper.With the model uses the shape modeling method of deep learning to extract the shape features of the target,and combined with the edge information extracted by Canny operator,the target shape with edge constraints is obtained by symbolic distance transformation,which is used as a priori information to re-feed into the model to get the segmentation results.The experimental results on the remote sensing data set Levir-oil?Levir-ship and Levir-airplane show that the small target segmentation model proposed in this paper can overcome the disadvantage of noise sensitivity in traditional small target segmentation methods,and can segment small targets quickly and accurately even when the training data is limited.Specifically,compared with the CV model,the average global ACC value of the proposed model on the test sets of Levir-oil drum?Levir-ship and levir-airplane increased to 98.654%,97.936%and 96.628%respectively.The average IOU values of Levir-oil drum?Levir-ship and levir-airplane were increased to 95.328%,94.140%and 92.425%,respectively.(3)In order to solve the problem of poor classification performance caused by lack of training samples and insufficient training of models in high-resolution remote sensing scene classification task,a remote sensing image scene classification model based on Convolution Restricted Boltzmann Machine is proposed in this paper.The model adopts pre-training fine-tuning method combined with softmax classifier to achieve the purpose of distinguishing image scene categories.Compared with the direct training model on the remote sensing data set WHU-RS19,the results show that the pre-training fine-tuning model can effectively improve the classification accuracy of scene recognition,and a higher recognition rate can be achieved even when the training samples of the data set are limited.