Image Processing Based On Deep Nonlinear Feature Representation Learning

As one of the most important information transmission media,image contains abundant visual perception and has been widely used in everyday life.Under the background of the big data,with various ways to acquire images,there are increasing requirements for a variety of image applications,from image beautification to face recognition,from autopilot to remote sensing,and so on.To meet these demands,many researchers focus on proposing machine learning algorithms and building artificial intelligence systems to extract the rich information in images,where the image representation learning is very important.Compared with shallow and linear models,deep nonlinear ones have more powerful expressive ability,that are more likely to reveal the complex nonlinear relationships between the original image space and the corresponding feature space.Therefore,this paper aims to study the image processing based on deep nonlinear feature representation methods.We start our research based on kernel methods,demonstrating the effectiveness of nonlinear models in data representation.To realize efficient representation learning under the background of big data,we investigate variational probabilistic generative frameworks parameterized by neural networks.In supervised image super-resolution task,such kind of framework achieves promising performance in data generation.Inspired by this phenomenon,we further apply this kind of model to unsupervised hyperspectral image super-resolution.The main contents of this dissertation are summarized as follows:1.A novel discriminative nonlinear dictionary learning approach is proposed for image object recognition.The objective for the representation learning contains two terms.The reconstruction term construct a structured kernel dictionary to extract the sparse representation of the image,while the discriminative term further nonlinearly maps the sparse representation to be a discriminative feature via kernel dictionary learning.Different from traditional methods that focus on the specific form of every discriminative feature,our method pays more attention on the correlations between different features.We use inner-product to describe such correlation,which matches the classification mechanism of support vector machine.A structured kernel KSVD algorithm is proposed to train the model.Experimental results demonstrate that the proposed approach outperforms many state-of-the-art dictionary learning approaches for face,scene and synthetic aperture radar vehicle target recognition.2.A general variational probabilistic generative framework is proposed for single image super-resolution,which tries to extract the shared latent representation of high-resolution(HR)and low-resolution(LR)patches in a supervised manner.We use probabilistic generative networks with a conditional prior to model the joint full likelihood of a pair of LR and HR patches which are generated from a shared latent representation.An inference model is applied to infer the stochastic distribution of the latent representation.By jointly optimizing the generative and inference models,a regression process to approximate the distribution of the HR patch is implied during the learning phase,which provides an efficient forward mapping to accomplish the super-resolution task.We likewise show how three existing coding-based and regression-based methods are “reinvented” under our framework,further demonstrating the advantages of our model in efficiency and robustness.3.A nonlinear variational probabilistic generative model is proposed to realize the unsupervised hyperpsectral image(HSI)super-resolution via fusing an LR-HSI and an HR multispectral image(MSI).We model the nonlinear spectral mixture process as a probabilistic generative model,and express the spectral response function as the generative process from the HR-HSI to the HR-MSI.For efficient inference,we construct two recognition models.We extend the probabilistic model as a deep neural network,which thus can be optimized via stochastic gradient updates.Compared with linear-mixture based fusion models,the proposed model has more powerful expressive ability,showing better fusion performance.Compared with multi-step methods,the proposed model jointly updates all the parameters,achieving close interactions between images to be fused.Our model is able to realize learning based on extra LR-HSI and HR-MSI in advance in an unsupervised manner,and processes the images at the test phase in real time.4.To better utilize the spatial correlations and local spectral structures,and better describe the spatial correlations between the LR and HR image pair,we further improve the variational probabilistic framework in Section 4,and employ convolutional neural networks to express the framework,called Fusion Net.Besides,for fast adaptation to different observation scenes,we give a meta-learning explanation to the fusion problem,and combine the Fusion Net with meta-learning in a synergistic manner.To the best of our knowledge,this is the first work that employs convolutional networks to solve such unsupervised fusion task,and the first work that combines meta-learning with the fusion task.