Aiming at the problems that the image super-resolution reconstruction model requires a large number of parameters to capture the statistical relationship between Low-Resolution (LR) images and High-Resolution (HR) images, and the use of network models optimized by L ₁ or L ₂ loss cannot effectively recover the high-frequency details of the images, an image super-resolution reconstruction method combining perceptual edge constraint and multi-scale fusion network was proposed. Based on the idea from coarse to fine, a two-stage network model was designed in this method. At the first stage, Convolutional Neural Network (CNN) was used to extract image features and upsample the image features to the HR size in order to obtain rough features. At second stage, multi-scale estimation was used to gradually approximate the low-dimensional statistical model to the high-dimensional statistical model. The rough features output at the first stage were used as the input to extract the multi-scale features of the image, and the features of different scales were gradually fused together through the attention fusion module in order to refine the features extracted at the first stage. At the same time, a class of richer convolutional features was introduced for edge detection and used as the perceptual edge constraint to optimize the network, so as to better recover the high-frequency details of the images. Experimental results on benchmark datasets such as Set5, Set14 and BSDS100 show that compared with the existing CNN-based super-resolution reconstruction methods, the proposed method not only reconstructs sharper edges and textures, but also achieves certain improvements in Peak Signal-to-Noise Ratio (PSNR) and Structural SIMilarity index (SSIM) when magnification factor is 3 and 4.

Concerning the problem that single face image is difficult to solve the large-pose profile face in face reconstruction, a face reconstruction method based on Multi-pose Feature Fusion Generative Adversarial Network (MFFGAN) was proposed. In this method, the relevant information between multiple profile faces with different poses was used for face reconstruction, and the adversarial mechanism was used to adjust network parameters. A new network was designed in the method, which consisted of a generator including multi-pose feature extraction, multi-pose feature fusion and frontal face synthesis, and a discriminator for adversarial training. In the multi-pose feature extraction module, multiple convolution layers were used to extract the multi-pose features of profile face images. In the multi-pose feature fusion module, the multi-pose features were fused into a fusion feature containing multi-pose face information. And, the fusion feature was added during the face reconstruction process in the frontal face synthesis module. Obtaining the relevant information and global structure by exploring the feature dependency between multi-pose profile face images can effectively improve the reconstruction results. Experimental results show that, compared with those of the state-of-the-art deep learning based face reconstruction methods, the contours of the frontal face recovered by the proposed method are clear, and the recognition rate of the frontal face recovered from two profile faces is increased by 1.9 percentage points on average; and the more profile faces are input, the higher the recognition rate of the recovered frontal face is, which indicates that the proposed method can effectively fuse multi-pose features to recover a clear frontal face.

Concerning the recovery problem of high-frequency information like texture details in image super-resolution reconstruction, an image super-resolution reconstruction method based on self-attention network was proposed. Two reconstruction stages were used to gradually restore the image accuracy from-coarse-to-fine. In the first stage, firstly, a Low-Resolution (LR) image was taken as the input through a Convolutional Neural Network (CNN), and a High-Resolution (HR) image was output with coarse precision; then, the coarse HR image was used as the input and a finer HR image was produced. In the second stage, the correlation of all positions between features was calculate by the self-attention module, and the global dependencies of features were captured to enhance texture details. Experimental results on the benchmark datasets show that, compared with the state-of-the-art deep neural networks based super-resolution algorithms, the proposed algorithm not only has the best visual effect, but also has the Peak Signal-to-Noise Ratio (PSNR) improved averagely by 0.1dB and 0.15dB on Set5 and BDSD100. It indicates that the network can enhance the global representation ability of features to reconstruct high quality images.

Concerning the problem that the extracted spatial-spectral features have the shortcoming of weak representation ability and high dimensionality, a HyperSpectral Image (HSI) classification method based on Spatial-Spectral Fusion Network (SSF-Net) was proposed. Firstly, a Two-channel Convolutional Neural Network (Two-CNN) was used to extract the features of spectral domain and spatial domain for HSI respectively. Secondly, Multimodal Compact Bilinear pooling (MCB) was employed to project the outer product of extracted multimodal features to the low dimensional space for producing jointly spatial-spectral features. The fusion network could not only analyze the complex relationship between elements in spectral and spatial eigenvectors, but also avoid directly computing the outer product of spectral and spatial vectors, resulting in high dimension and high computation times. The experimental results show that, compared with the state-of-the-art neural network based classification methods, the proposed algorithm can obtain higher pixel classification accuracy. It indicates that the spatial-spectral joint vector extracted by the proposed network has stronger representation ability.

To extract more effective features and speed up the convergence of model training, a super-resolution reconstruction algorithm based on parallel convolution neural network was proposed. The network consists of two different network structures, one is a simple residual network structure, which has a easier optimal residual mapping than the original one; the other is a convolutional neural network with nonlinear mapping, which can increase the non-linearity of the network. As the complexity of the parallel network structure, the convergence speed is the key issue. Aiming at this problem, the Local Response Normalization (LRN) layer was added to the convolution layers to simplify the model parameters and enhance the feature fitting ability, thus accelerating the convergence. Experimental results show that, compared with algorithms based on deep convolutional neural network, the proposed method accelerates the convergence, improves the visual quality, and increases Peak Signal-to-Noise Ratio (PSNR) at least 0.2 dB.

To circumvent the influence of pose variance on face recognition performance and considerable probability of losing the facial local detail information in the process of pose recovery, a multi-pose face reconstruction and recognition method based on multi-task learning was proposed, namely Multi-task Learning Stacked Auto-encoder (MtLSAE). Considering the correlation between pose recovery and retaining local detail information, multi-task learning mechanism was used and sparse auto-encoder with non-negativity constraints was introduced by MtLSAE to learn part features of the face when recovering frontal images using step-wise approach. And then the whole net framework was learned by sharing parameters between above two related tasks. Finally, Fisherface was used for dimensionality reduction and extracting discriminative features of reconstructed positive face image, and the nearest neighbor classifier was used for recognition. The experimental results demonstrate that MtLSAE achieves good pose reconstruction quality and makes facial local texture information clear; on the other hand, it also achieves higher recognition rate than some classical methods such as Local Gabor Binary Pattern(LGBP), View-Based Active Appearance (VAAM) and Stacked Progressive Auto-encoder (SPAE).

Focusing on the issue that remote sensing fusion image based on Contourlet transform has low spatial resolution, a remote sensing image fusion algorithm based on Modified Contourlet Transform (MCT) was proposed. Firstly, the multi-spectral image was decomposed into intensity component, hue component and saturation component by Intensity-Hue-Saturation (IHS) transform; secondly, Modified Contourlet decomposition was done between the intensity component and the panchromatic image after histogram matching to get low-pass subband coefficients and high-pass subbands coefficients; and then, the low-pass subband coefficients were fused by the averaging method, and the high-pass subbands coefficients were merged by Novel Sum-Modified-Laplacian (NSML). Finally, the fusion result was regarded as the intensity component of multi-spectral image, and remote sensing fusion image was obtained by inverse IHS transform. Compared with the algorithms based on Principal Components Analysis (PCA) and Shearlet, based on PCA and wavelet, based on NonSubsampled Contourlet Transform (NSCT), the average gradient that was used for evaluating image sharpness of the proposed method respectively increased by 7.3%, 6.9% and 3.9%. The experimental results show that, the proposed method enhances the frequency localization of Contourlet transform and the utilization of decomposition coefficients, and on the basis of keeping multi-spectral information, it improves the spatial resolution of remote sensing fusion image effectively.

To improve the accuracy of focusd regions in multifocus image fusion based on multiscale transform, a multifocus image fusion algorithm was proposed based on NonSubsampled Shearlet Transform (NSST) and focused regions detection. Firstly, the initial fused image was acquired by the fusion algorithm based on NSST. Secondly, the initial focusd regions were obtained through comparing the initial fused image and the source multifocus images. And then, the morphological opening and closing was used to correct the initial focusd regions. Finally, the fused image was acquired by the Improved Pulse Coupled Neural Network (IPCNN) in the corrected focusd regions. The experimental results show that, compared with the classic image fusion algorithms based on wavelet or Shearlet, and the current popular algorithms based on NSST and Pulse Coupled Neural Network (PCNN), objective evaluation criterions including Mutual Information (MI), spatial frequency and transferred edge information of the proposed method are improved obviously. The result illustrates that the proposed method can identify the focusd regions of source images more accurately and extract more sharpness information of source images to fusion image.