Research On Target Extraction From High-resolution Remote Sensing Images Based On Convolution-deconvolution Deep Neural Network

Target extraction from high-resolution remote sensing images can timely obtain geographic updated data,providing an important basis for the construction of geospatial database.Currently,it has become a research hotspot in the fields of photogrammetry and remote sensing,geographic information system and computer vision.Deep learning technology has been widely used in remote sensing image recognition.The advantage of deep learning technology is that it can automatically extract more appropriate features.In general,features extracted automatically are more effective,but there is still room for improvement.First,the existing deep learning technology cannot directly process large-scale images,image must be cut into small pieces input network for segmentation extraction.In the general remote sensing target extraction data set,there is often a quantitative imbalance between the image containing the target and the full background image.As a result,the network training is unbalanced,and the cropped image of the full background area is prone to wrong segmentation during segmentation.Therefore,it is necessary to carry out scene recognition first.Secondly,for the deep convolutional neural network,the repeated use of pooling operation reduces the feature resolution,and the prediction result through up-sampling is rough,so it is difficult to accurately retain the target region edge,location and other details.In addition,the use of higher performance and deeper convolutional neural networks as the backbone of the feature extraction and category prediction module is beneficial to extract features,but it will increase the number of parameters and require more tag samples to train the network,while it is difficult to provide sufficient training samples for most practical applications.Thirdly,the deep convolutional neural network is difficult to bear the input of large-size images and extract small target objects to obtain high accuracy.In this paper,based on the existing research theories and taking into account the characteristics of remote sensing image size,small target and insignificant difference between target and background,the deep learning method is applied to the research of remote sensing image target extraction.The innovation points are as follows.(1)Extracting features from different input sources,such as multi-spectral remote sensing images,where one branch receives multi-spectral channel image input and the other branch receives panchromatic channel image input,or elevation information,etc.Although the information provided by data from different sources is redundant and complementary,weighted probability fusion is carried out at the end of the segmentation network,which can not only highlight the target but also effectively inhibit misclassification and improve the segmentation performance.The decision level fusion method is adopted,the weighted fusion of category probability graph is used at the end of the two branch networks to fully integrate the advantages of the two neural network branches,so that the network branches with better performance can play a greater role in the fusion.(2)A full resolution neural network model combining multi-source input information is proposed.The network uses a convolution-deconvolution network as the backbone network,and joins the full resolution network branch in the backbone network;A data exchange mechanism is established between the backbone network and the full resolution branches,which not only overcomes the problems of reduced feature resolution and loss of detailed information caused by repeated pooling operations,but also aggregates multi-scale features in convolution stage.The converged features are transferred to the corresponding layers in deconvolution stage,which enhances the feature fusion.(3)A remote sensing image target extraction method based on scene recognition task is proposed.The method firstly carries out scene recognition for the original remote sensing image and obtains the target scene image.Then the target scene image is sent into the segmentation module for target extraction.This method first identifies the scene,then segments the target in the scene where the target may exist,solves the problem of extracting specific target from large-scale high-resolution remote sensing image,and at the same time reduces the wrong segmentation of background region caused by the imbalance of positive and negative samples..