Contourlet Residual Networks And Ant Colony Optimization Algorithm For Multi-spectral Image Classification

With the rapid development of remote sensing imaging technology,the resolution of multispectral images is getting higher and higher,and the spectral information contained in them is also becoming more and more abundant.However,the classification of multi-spectral images is usually faced with four problems,same object with different spectra,the different object with same spectra,polymorphic object and the small number of labeled samples caused by the difficulty of labeling.Based on the above problems and the advantages of the deep convolutional neural network in image feature extraction,this thesis proposes a multi-spectral image classification based on Contourlet residual network and ant colony optimization algorithm.The main progress in this thesis is as follows:1.A multi-spectral image classification based on the ant colony optimization(ACO)algorithm and deep residual network.Aiming at the problem of the small number of labeled samples in the multi-spectral images,a training sample selection strategy based on the ACO algorithm is proposed to select diverse samples,it can obtain a good generalization performance.At the same time,in order to solve the problems of the same object with different spectra and different object with same spectra,this method also constructs a deep residual network based on the residual block to extract features with more discriminating,thereby improving the classification accuracy of the model.Experiments on four multispectral images in Vancouver and Xi’an have proved that the deep residual network can achieve better classification accuracy results.At the same time,the training sample selection strategy based on the ACO algorithm is better than random sampling and cluster sampling.2.A multi-spectral image classification based on weighted deformable convolution network(WDCNet).Aiming at the polymorphic object in the multi-spectral image,a deep convolutional neural network based on weighted deformable convolution module is proposed,and the network is named WDCNet.Compared with the traditional convolution module,the weighted deformable convolution module can improve the flexibility of the neural network in extracting features,thereby improving the feature representation ability of the model.Experiments on four multi-spectral images in Vancouver and Xi’an have proved that the classification performance of WDCNet is better than that of residual network and deformable convolution network,and the sample selected based on ACO algorithm is used as the training sample of WDCNet,the classification accuracy of the model can be further improved.3.A multi-spectral image classification based on nonsubsampled contourlet transform(NSCT)and WDCNet.Aiming at the problem that WDCNet is difficult to classify edge points with complex information and similar shapes,the weighted loss function based on NSCT is proposed as the objective function of WDCNet.Firstly,the non-downsampled contourlet transform is used for edge detection,and then the weighted loss function based the edge detection result is designed,which increases the proportion of the loss of edge points,thus improving the classification accuracy of edge points.Experiments on four multi-spectral images in Vancouver and Xi’an show that the classification accuracy of edge detection using NSCT is better than that of Canny operator and Sobel operator.The model trained by weighted loss function is better than the traditional cross entropy loss function,and it improves the generalization performance of the model.