Deep Learning Building Recognition Methods Based On Multi-features

Buildings are important grounds in multispectral images,how to recognize the buildings accurately is the research hotspot at present.The deep learning technology is a branch of artificial intelligence.There are many methods for target recognition using deep learning technology,and they are effective.However,there are still some deficiencies in the building recognition of multispectral images.In this thesis,two kinds of deep neural network models and their recognition mechanism is studied,a building recognition method based on depth belief network and multi features is proposed.This method mainly studies the recognition method based on the mixture of artificial features and the features extracted by DBN.And a building recognition method based on the improved full convolutional network is proposed.This method mainly studies the feature extracted by neural network automatically.There are a lot of information such as texture,contour and color in buildings of multispectral images.Thus,utilizing these information and design suitable features are prerequisites for recognizing buildings.In the second chapter,this thesis first extracts the Gabor-HoG feature of buildings,which combines the texture features and edge features of the image and the extracted features are fused.Then inputting the fused features into the deep belief network to extract higher-level features.These features are input into the conditional random field to extract the context feature subsequently.Then the conditional random field model infers the probability of buildings target to realize the pixel-level recognition.It mainly solves the targets'context information and low-level features extraction that are ignored in the classical deep belief network recognition methods.Besides,in contrast with the recognition method based on DBN and CRF,the recognition method based on Gabor-HoG feature and DBN and a recognition method based on Gabor-HoG feature and CRF,it is concluded that this method is more effective for building recognition.The convolutional neural network can achieve automatic feature extraction and recognizing targets by training convolution kernels.In Chapter 3,this thesis proposes a model fully convolutional network based on multi-scale features and rotation expansion training sets for buildings recognition of multispectral images to recognize buildings.Compared with the recognition method based on depth belief network and multi features,this method extracted the features by the convolutional kernels which were trained by BP algorithm automatically,it need not design the feature mannualy,and the features extracted by this method are rich.This method firstly transforms the scales of buildings,then extracts multi-scale features by four sets of parallel convolutional kernels and fuses them to extract the higher-layer features serially.The extracted features classified by the Softmax classifier finally.By this method,the difference of the effect in the recognition caused by different scales in the classical fully convolutional neural network is resolved.In addition,it improves the accuracy influenced by the different orientation of the building.And by comparing to the FCN recognition method proposed by J Long,the improved recognition FCN method proposed by B Hariharan and a Faster R-CNN recognition method which proposed by Shaoqing Ren,it is concluded that this method works with better performance.