Semi-supervised Curvelet Compressed Convolution Networks For Remote Sensing Image Classification

This thesis focuses on remote sensing image classification based on semi-supervised Curvelet compression convolution network.In recent years,remote sensing image data has developed in both quantity and quality.The amount of labeled data used for training network only accounts for a very small portion of data to be classified.Therefore,one of the difficulties in remote sensing image classification is to extract more discriminant features using a small number of training data.Curvelet combined pyramid convolution network is proposed for extracting multi-scale,multi-directional and multi-resolution features.CRF-RF algorithm is proposed for features refinement and classification.Compressed convolution network with SVD is proposed for remote sensing image classification,which greatly reduces the parameter quantity and computational complexity when the accuracy is guaranteed.Above all,the main progress in this thesis are listed as follows:1.A CRF-RF and Curvelet convolutional neural network is proposed for remote sensing image classification.The Curvelet convolution model which is constructed by Curvelet Filtering and convolution neural network is used to extract multi-directional and multi-scale features.The CRF-RF algorithm is used as features refinement classifier,and the categorical algorithm is used to correct classification results.Compensating for the end-to-end convolutional neural networks requires a large amount of training data but extracted features are single.The model achieves effective results in multiple Pol SAR images.2.A semi-supervised pyramid convolution neural network is proposed for multi-spectral image classification.Superpixel segmentation algorithm combined with KNN algorithm is used for eliminating alternative pixel points.Combining the inverted pyramid model with the convolution pooling layer is used for features extraction task.The semi-supervised pyramid convolution neural network solves the problem that central pixel in the fixed input block of the convolution neural network can be mixed with different types of pixels in the surrounding,and also weakens the influence of information loss on classification result during training process.Compared with the single-scale deep convolution neural network,the proposed method can obtain more discriminant and robust features.The model achieves effective results in multi-spectral images.3.A multi-spectral images classification based on SVD lightweight convolution neural network is proposed.In this model,SVD is used for fusing high-resolution images with lowresolution images.We use deep separation convolution instead of standard convolution.The global average pooling layer is used to replace the first fully connected layer connected to the convolution.SVD decomposes the full connection layer parameters,and takes the obtained decomposition parameters as a spatial representation of the full connection layer.In the pursuit of making deeper,more complex,and more accurate convolutional neural networks,the problems of size and parameter quantity have been effectively solved.The model achieves effective results in multi-spectral images.