Aiming at the problems of feedforward Convolutional Neural Network(CNN), such as small receptive field, insufficient context information acquirement and that only shallow features can be extracted by the feature extraction convolutional layer of the network, an improved remote sensing image fusion algorithm based on channel attention feedback network was proposed. Firstly, the detail features of PANchromatic(PAN) images and the spectral features of Low-resolution MultiSpectral(LMS) images were initially extracted through two convolutional layers. Secondly, the extracted features were combined with the deep features fed back from the network and inputted to the channel attention mechanism module to obtain the initially refined features. Thirdly, the deep features with stronger characterization capability were generated by feedback module. Finally, High-resolution MultiSpectral(HMS) images were obtained by putting the generated deep features into the reconstruction layer with deconvolution. Experimental results on three different satellite image datasets show that the proposed algorithm can well extract the detail features of PAN images and the spectral features of LMS images, and the HMS images recovered by this algorithm are clearer subjectively and better than the comparison algorithms objectively; at the same time, the Root Mean Square Error(RMSE) index of the proposed method is more than 50% lower than that the traditional methods, and more than 10% lower than that the feedforward convolutional network methods.

In order to more effectively combine the detail information of high resolution PANchromatic (PAN) image and the spectral information of low resolution MultiSpectral (MS) image, an improved panchromatic sharpening algorithm based on sparse representation was proposed. Firstly, the intensity channel of an MS image was down-sampled and then up-sampled to get its low-frequency components. Secondly, the MS image intensity channel minus low-frequency components to obtain its high-frequency components. Random sampling was performed in the acquired high and low frequency components to construct a dictionary. Thirdly, the PAN image was decomposed to get the high-frequency components by using the constructed overcomplete dictionary. Finally, the high-frequency components of the PAN image were injected into the MS image to obtain the desired high-resolution MS image. After a number of experiments, it was found that the proposed algorithm subjectively retains the spectral information and injects a large amount of spatial details. Compared with component substitution method, multiresolution analysis method and sparse representation method, the reconstructed high resolution MS image by the proposed algorithm is more clear, and the correlation coefficient and other objective evaluation indicators of the proposed algorithm are also better.

Among the dictionary-based Super-Resolution (SR) algorithms, the Anchored Neighborhood Regression (ANR) algorithm has been attracted widely attention due to its superior reconstruction speed and quality. However, the anchored neighborhood projections of ANR are unstable to cover varieties of mapping relationships. Aiming at the problem, an image SR algorithm based on adaptive anchored neighborhood regression was proposed, which adaptively calculated the neighborhood center based on the distribution of samples in order to pre-estimate the projection matrix based on more accurate neighborhood. Firstly, K-means clustering algorithm was used to cluster the training samples into different clusters with the image patches as centers. Then, the dictionary atoms were replaced with the cluster centers to calculate the corresponding neighborhoods. Finally, the neighborhoods were applied to pre-compute the projection matrix from LR space to HR space. Experimental results show that the average reconstruction performance of the proposed algorithm on Set14 is better than that of other state-of-the-art dictionary-based algorithms with 31.56 dB of Peak Signal-to-Noise Ratio (PSNR) and 0.8712 of Structural SIMilarity index (SSIM), and even is superior to the Super-Resolution Convolutional Neural Network (SRCNN) algorithm. At the same time, in terms of the subjective performance, the proposed algorithm produces sharp edges in reconstruction results with little artifacts.

Spectral interference in Red Green Blue-Near InfRared (RGB-NIR) images captured by single sensor results in colour distortion and detail information ambiguity of the reconstructed standard Red Green Blue (RBG) and Near InfRared (NIR) images. To resolve this problem, a demosaicing method based on deep learning was proposed. In this method, the grandient dppearance and dispersion problems were solved by introducing long jump connection and dense connection, the network was easier to be trained, and the fitting ability of the network was improved. Firstly, the low-level features such as pixel correlation and channel correlation of the mosaic image were extracted by the shallow feature extraction layer. Secondly, the obtained shallow feature graph was input into successive and multiple residual dense blocks to extract the high-level semantic features aiming at the demosaicing. Thirdly, to make full use of the low-level features and high-level features, the features extracted by multiple residual dense blocks were combined. Finally, the RGB-NIR image was reconstructed by the global long jump connection. Experiments were performed on the deep learning framework Tensorflow using three public data sets, the Common Image and Visual Representation Group (IVRG) dataset, the Outdoor Multi-Spectral Images with Vegetation (OMSIV) dataset, and the Forest dataset. The experimental results show that the proposed method is superior to the RGB-NIR image demosaicing methods based on multi-level adaptive residual interpolation, convolutional neural network and deep residual U-shaped network.

Because Support Vector Data Description (SVDD) fails in identifying the local geometric information, a new detection method, called One-class SVDD based on Local Patch (OCSVDD_LP), was proposed. First, the data was divided into many local patches. Then, each sample was reconstructed by using the corresponding local patch. Finally, the decision model was obtained through training on the reconstruction data with SVDD. The experimental results on the artificial data set demonstrate that OCSVDD_LP can not only capture the global geometric structure of the data set, but also uncover the local geometric information. Besides, the results on real-world data sets validate the effectiveness of the proposed method.