Sharif Digital Repository

In the last two decades, novel view synthesis has become increasingly important in the movie and game industries. The new generated views correspond to virtual viewpoints, where no real camera has taken snapshot. Depending on the amount of available information from the scene, different approaches have been proposed to create novel views. If the depth information is available, then by knowing camera relative motion, novel views can be constructed. The conventional cameras only record the intensity of the captured scene, and discard the depth information. In the recent concept of light field cameras, by locating an array of micro-lenses between the camera sensor and the main lens, one can... 

The conventional cameras and displays do not have the ability to record and display the full brightness of the world around us. These deficiencies have led to the development of methods known as High Dynamic Range Imaging. Most of the work done in this field falls into two groups: compression of the light range and its expansion. In compression, the brightness range is intended to display content with a high dynamic range in simple displays. But the goal of expanding the brightness range is to reconstruct HDR content from a low dynamic range one. In addition to the above classification, this field is also categorized based on the type of content (image or video). In this study, the... 

Nowadays in movie and photography industry, scene reconstruction is of high importance. Capturing data containing the depth information of the scene can help us achieve this goal more easily. Ordinary Cameras do not record depth information of the scene so they can’t ease our task. In 2005, some Stanford Researchers, inspired by Adelson’s work, were able to build the first camera that captures depth information. They called it the light filed camera. In these cameras an array of microlenses is inserted below the main lens. By doing this the camera is able to capture the rays’ intensities and their directions in space too. With this kind of information, we are able to obtain the depth of a... 

In this project, we are going to face the problem of learning a graph on images with big sizes which has many applications in image processing like image segmentation. There are two main obstacles to doing so. One is the fact that the task of graph learning becomes immensely time-consuming when we have a large number of nodes. The other is the memory shortage since most computers do not have rams with big capacities, which is a requirement when the number of nodes is too high. We are going to suggest two suboptimal solutions for finding a sparse graph on images with respect to smoothness criteria on the Laplacian matrix of the acquired graph and reduce the time needed for this task. Then we... 

In the past few years, the available knowledge in graph-based processing has made significant progress, and as a result, powerful tools have been created. In this regard, graph learning with the assumption of data smoothness on the final result can be considered a successful example. Briefly, in graph learning, to describe the relationship between the problem components, a graph is learned using the available data whose nodes represent the problem components, and its edges represent how much these components are connected. The usefulness of this method lies in the possibility of using the obtained graph as the input to currently known methods of classification and achieving better results... 

Hyperspectral images are utilized in various fields due to their rich spectral information, but they often suffer from blur induced by the optical system, which is typically variable across the spectrum. Accurate estimation of this blur is essential for correct information retrieval. This research presents a novel optimization framework for the accurate and efficient estimation of Point Spread Function (PSF) in hyperspectral images. The core innovation of this method is the introduction of a "spectral coupling" constraint, which models the smooth variations of the PSF among adjacent bands, significantly enhancing the stability of the estimation. By transforming the problem into the Fourier... 

Compressed sensing theory with the goal of sparse recovery using as few measurements as possible, has found much attention in the past decade. A fundamental assumption in compresses sensing is the linear nature of the measurements, that establishes a matrix equation between the unknowns and unknowns. However, in some applications, the physics of the problem may impose some type of nonlinear relationship between the unknowns and the measurements. Some solutions concerned with this type of problems,approximate the nonlinear relationship using a second order Taylor expansion. Using this approximation besides some modifications in the way the second order relationship is presented, a more simple... 

The design of deterministic sensing matrices is an important issue in compressive sensing in sparse signal processing. Various designs using finite field structures, combinatorics, and coding theory have been presented. The contribution of this thesis is designing many codes with large minimum distance using algebraic curves. Here, we initially design a algebraic-geometric code over a maximal curve in a Galoi field Fpm Afterwards, we map the code to the field Fp using trace map. This code has a large minimum distance. Using this code, we design a matrix with low coherence. One of the main issues in presented designs is that the number of matrix rows is considered to specific integers near... 

Sampling and recovery of a signal is one of the crucial issues in communication systems. In conventional methods, proper recovery is achieved by sampling the signal at the Nyquist rate, which is twice the signal bandwidth. There have been attempts on reducing the required sampling rate, all of which end in rates equal to a factor of the signal bandwidth. Assuming the sparse nature of the signal in hand, which is a reasonable assumption in many real world scenarios, the theory of Compressed Sensing suggests a sampling rate much less than the Nyquist rate. Designing suitable sensing matrices and efficient recovery of the signal from its samples are the two major challenges of Compressed... 

In this thesis, we propose a new compressed-sensing-based algorithm for unmixing of hyperspectral data, and show that the reconstruction quality could be significantly improved. In addition, we illustrate that by utilizing this approach, we can achieve an approximate estimation of the Point Spread Function (PSF) of the hyperspectral images. In this way, we first assume that the PSF belongs to a specific family of functions, such as Gaussian, then, by sweeping the parameters of the assumed PSF, we obtain the abundance coefficient matrix of the reconstructed image. Now, by choosing the sparsest coefficient matrix as the best one, we estimate the corresponding PSF. Then, we further investigate... 

Video stabilization is one of the main quality enhance- ment techniques for the data captured by a moving/shaking video recorder. In this work, we focus on video stabilization in smartphones equipped with inertial measurement units (IMU). Specifically, we present a fully differentiable dynamic computational graph to compensate for simple pinhole camera model deficiencies, using the suggested model and estimated angular rotation changes inferred from IMU data, We try to improve and stabilize shaky videos. Further using qualitative and quantitative approaches we show our method’s supremacy against current state of the art stabilization methods. This work accompanied with software and necessary... 

Hyperspectral unmixing (HU) is a method used to estimate the fractional abundances corresponding to endmembers in each of the mixed pixels in the hyperspectral remote sensing image. In recent times, deep learning has been recognized as an effective technique for hyperspectral image classification. In this thesis, an end-to-end HU method is proposed based on the convolutional neural network (CNN) and multi-layer perceptron (MLP). which consists of two steps: the first stage extracts features from the input data along with the inverse learning of the spectral library matrix in the hyperspectral image where columns represent the pure spectral of endmembers and The second stage is to estimate... 

In this thesis, we generalize the concepts of kernels, weak stationarity and white noise from undirected to directed graphs (digraphs) based on the Jordan decomposition of the shift operator. We characterize two types of kernels (type-I and type-II) and their corresponding localization operators for digraphs. We analytically study the interplay of these types of kernels with the concept of stationarity, specially the filtering properties. We also generalize graph Wiener filters and the related optimization framework to digraphs. For the special case of Gaussian processes, we show that the Wiener filtering again coincides with the MAP ‌estimator. We further investigate the linear minimum... 

white noise or an innovation process through an all-pole filter. Applications of these processes include speech processing, RADAR signals and stock market data modeling. There exists an extensive research material on the AR processes with Gaussian innovation, however studies about the non-Gaussian case have been much more limited, while in many applications the asymptotic behavior of the signal is non-Gaussian. Non-Gaussian processes have an advantage over Gaussian ones in being capable of modeling sparsity. Assuming an appropriate non-Gaussian innovation one can suggest a more realistic description of sparse signals and predict their behavior or estimate their unknown values successfully.... 

Channel estimation is one of the fundamental challenges in every communication system and different algorithms have been proposed to deal with it. Obviously, type of a communication channel is an important factor in choosing the appropriate method for channel estimation. Sparse channels are one kind of them that occur in many real-world applications such as wireless communication systems. In addition, emergence of a new means in signal processing to deal with sparse signals, known as Compressed Sensing(CS), paved the way for their extensive usage in many applications including sparse channel estimation.On the other hand, one of the most fundamental problems in sparse signal recovery using CS... 

Brain states analysis during consciousness is emerging research in brain-computer interface(BCI). Emotion recognition can be applied to learn brain states and stages of neural activities. Therefore, emotion recognition is crucial to the analysis of brain functioning. Electrical signals such as electroencephalogram (EEG), electrocardiogram (ECG) and functional magnetic resonance imaging(fMRI) are frequently used in emotion recognition researches. Convenience in recording, non-invasive nature and high temporal resolution are the factors that have made EEG popular in brain researches. EEG can be used to identify brain region activity solely or the connectivity of various regions in time in the... 

In this thesis, we deal with the problem of direction of arrival (DOA) using single-snapshot samples taken from multiple source signals by an array of sensors. The array we are looking for in this work is a uniform linear array. There are many ways to solve this problem, including MUSIC and Prony. But one of the few methods that are able to estimate correctly with single-snapshot samples with incomplete uniform linear arrays in compressed sensing base algorithms. In the first step, we can estimate the data of a incomplete uniform linear array by arranging them in the form of low-rank Hankel matrices and using leverage scores and minimizing the weight nuclear norm. If the weight matrices are... 

In competitive markets where a product or service is provided by multiple providers, as the telecom market, keeping active users is expected to be less expensive than attracting new users. In this regard, first of all, churning should be predicted for active users, and secondly, proper recommendations should be provided to prevent churning. In this thesis, by modeling customer churn as a response to the recommendations, we study the churn prediction and prevention problem as a recommender system. This model enables us to select the best offer for each user to prevent it from churning.Modeling customer churn in a recommender system introduces new challenges, including delay in observing... 

High dynamic range (HDR) images provide more realistic experience in displaying real-world scenes than conventional low dynamic range (LDR) images by providing much more detailed luminance information; However, most imaging content is still available in low dynamic range. Inverse tonemapping is known as the problem of inferring an HDR image from a single-exposure LDR image in which the lost data caused by saturation of bright parts and quantization must be reconstructed.To address this problem, in this thesis, two fully-automatic architectures based on convolutional neural networks, are proposed. Both these architectures utilize a number of convolutional auto-encoders as... 

Falling is a prevalent event among elderly people, which sometimes leads to their death. Automatic detection of fall can significantly reduce the resulting damages.Fallings can be detected using various modalities, among which we choose RGB videos captured by CCTV cameras because of its advantages. Due to the great advances in deep learning-based image/video classification methods, we focused on using these methods for fall detection. One of the main challenges in using deep learning methods is lack of enough training data. Unlike other activities, there are not enough falling samples available which is due to its unconscious nature. Moreover, simulating falling by actors can endanger their...