Computational Imaging For Scene Understanding

Most cameras are inherently designed to mimic what is seen by the human eye: they have three channels of RGB and can achieve up to around 30 frames per second (FPS). However, some cameras are designed to capture other modalities: some may have the ability to capture spectra from near UV to near IR rather than RGB, polarimetry, different times of light travel, etc. Such modalities are as yet unknown, but they can also collect robust data of the scene they are capturing. This book will focus on the emerging computer vision techniques known as computational imaging. These include capturing, processing and analyzing such modalities for various applications of scene understanding.

The goal of this book is to address the use of several important machine learning techniques into computer vision applications. An innovative combination of computer vision and machine learning techniques has the promise of advancing the field of computer vision, which contributes to better understanding of complex real-world applications. The effective usage of machine learning technology in real-world computer vision problems requires understanding the domain of application, abstraction of a learning problem from a given computer vision task, and the selection of appropriate representations for the learnable (input) and learned (internal) entities of the system. In this book, we address all these important aspects from a new perspective: that the key element in the current computer revolution is the use of machine learning to capture the variations in visual appearance, rather than having the designer of the model accomplish this. As a bonus, models learned from large datasets are likely to be more robust and more realistic than the brittle all-design models.

Computational photography refers broadly to imaging techniques that enhance or extend the capabilities of digital photography. This new and rapidly developing research field has evolved from computer vision, image processing, computer graphics and applied optics—and numerous commercial products capitalizing on its principles have already appeared in diverse market applications, due to the gradual migration of computational algorithms from computers to imaging devices and software. Computational Photography: Methods and Applications provides a strong, fundamental understanding of theory and methods, and a foundation upon which to build solutions for many of today's most interesting and challenging computational imaging problems. Elucidating cutting-edge advances and applications in digital imaging, camera image processing, and computational photography, with a focus on related research challenges, this book: Describes single capture image fusion technology for consumer digital cameras Discusses the steps in a camera image processing pipeline, such as visual data compression, color correction and enhancement, denoising, demosaicking, super-resolution reconstruction, deblurring, and high dynamic range imaging Covers shadow detection for surveillance applications, camera-driven document rectification, bilateral filtering and its applications, and painterly rendering of digital images Presents machine-learning methods for automatic image colorization and digital face beautification Explores light field acquisition and processing, space-time light field rendering, and dynamic view synthesis with an array of cameras Because of the urgent challenges associated with emerging digital camera applications, image processing methods for computational photography are of paramount importance to research and development in the imaging community. Presenting the work of leading experts, and edited by a renowned authority in digital color imaging and camera image processing, this book considers the rapid developments in this area and addresses very particular research and application problems. It is ideal as a stand-alone professional reference for design and implementation of digital image and video processing tasks, and it can also be used to support graduate courses in computer vision, digital imaging, visual data processing, and computer graphics, among others.

This book presents a detailed analysis of spectral imaging, describing how it can be used for the purposes of material identification, object recognition and scene understanding. The opportunities and challenges of combining spatial and spectral information are explored in depth, as are a wide range of applications. Features: discusses spectral image acquisition by hyperspectral cameras, and the process of spectral image formation; examines models of surface reflectance, the recovery of photometric invariants, and the estimation of the illuminant power spectrum from spectral imagery; describes spectrum representations for the interpolation of reflectance and radiance values, and the classification of spectra; reviews the use of imaging spectroscopy for material identification; explores the recovery of reflection geometry from image reflectance; investigates spectro-polarimetric imagery, and the recovery of object shape and material properties using polarimetric images captured from a single view.

Conventionally, computer vision has imitated the human visual system. As we understand the world by seeing it with our eyes, a lot of research in computer vision focuses on deriving meaningful information from the digital images captured with conventional cameras. However, human eyes (and conventional cameras) can see only a fraction of the information that nature provides. Some useful scene information is exposed only when the scene is excited with external light. Active computational imaging enables to estimate various scene properties which cannot be captured with conventional cameras by using a controllable illumination source to help probe the scene actively. However, active computational imaging requires more cost and power consumption than conventional passive imaging due to additional light sources and often, specialized sensors. Although the cost and power constraints can be relaxed by lowering capture time or source power, it generally leads to lower signal-to-noise ratio (SNR) measurements. Moreover, scene property estimates by active imaging systems are prone to errors in non-ideal imaging conditions including defocus, multi-path and multi-camera interference, and ambient illumination. The goal of this thesis is to optimize three main parts of active imaging system, light source, sensor and computation to provide robust scene property estimates in various low SNR scenarios. Target scene properties in this thesis are 3D geometry and fluorescence lifetime which are widely used in many applications but challenging to estimate accurately. We show how these useful scene properties can be estimated robustly in challenging scenarios with various active imaging modalities. This thesis has four contributions. First, we propose a class of active 3D imaging systems which recover 3D geometry of piece-wise planar scenes (Blocks-World) in resource-limited conditions. Our approach, called Blocks-World Cameras, does not require acquisition of 3D point clouds which are generally memory intensive and are subject to errors in non-ideal imaging conditions. The Blocks-World Cameras based on a structured-light system project a single pattern with a sparse set of cross-shaped features. Dominant planar scenes are recovered using a novel geometric algorithm without explicit correspondence matching. Second, we propose a novel approach to mitigate multi-camera interference in active 3D imaging. Time-of-flight (ToF) cameras are a popular active 3D imaging modality, but multi-camera-interference emerges as an important issue when these cameras become ubiquitous. Our approach achieves high SNR by filtering out both AC and DC interference, robustness to ambient light by amplifying source peak power, and saturation-free 3D imaging by time-slotting. Third, we develop theory and algorithms to design temporal illumination patterns for high-performance active fluorescence lifetime imaging. Based on a novel surrogate objective function, we design high-SNR illumination patterns that achieve up to an order of magnitude shorter acquisition time as compared to existing ones. Lastly, we propose a general-purpose photon processing algorithm for active single-photon imaging, which is called CASPI (Collaborative photon processing for Active Single-Photon Imaging). CASPI is a technology-agnostic, application-agnostic, tuning-free and training-free photon processing pipeline, which enables to estimate scene properties reliably even under extreme lighting conditions. CASPI is versatile and can be integrated into a wide range of imaging applications including fluorescence microscopy, machine vision, and long-range 3D imaging. The performance benefits of all these active computational imaging approaches are demonstrated with thorough theoretical analysis, simulations and real experiments, across a wide range of challenging imaging scenarios in this thesis.

The introduction and recent advancements of computational photography have revolutionized the imaging industry. Computational photography is a combination of imaging techniques at the intersection of various fields such as optics, computer vision, and computer graphics. These methods enhance the capabilities of traditional digital photography by applying computational techniques both during and after the capturing process. This thesis targets two major subjects in this field: High Dynamic Range (HDR) image reconstruction and Light Field (LF) compressive capturing, compression, and real-time rendering. The first part of the thesis focuses on the HDR images that concurrently contain detailed information from the very dark shadows to the brightest areas in the scenes. One of the main contributions presented in this thesis is the development of a unified reconstruction algorithm for spatially variant exposures in a single image. This method is based on a camera noise model, and it simultaneously resamples, reconstructs, denoises, and demosaics the image while extending its dynamic range. Furthermore, the HDR reconstruction algorithm is extended to adapt to the local features of the image, as well as the noise statistics, to preserve the high-frequency edges during reconstruction. In the second part of this thesis, the research focus shifts to the acquisition, encoding, reconstruction, and rendering of light field images and videos in a real-time setting. Unlike traditional integral photography, a light field captures the information of the dynamic environment from all angles, all points in space, and all spectral wavelength and time. This thesis employs sparse representation to provide an end-to-end solution to the problem of encoding, real-time reconstruction, and rendering of high dimensional light field video data sets. These solutions are applied on various types of data sets, such as light fields captured with multi-camera systems or hand-held cameras equipped with micro-lens arrays, and spherical light fields. Finally, sparse representation of light fields was utilized for developing a single sensor light field video camera equipped with a color-coded mask. A new compressive sensing model is presented that is suitable for dynamic scenes with temporal coherency and is capable of reconstructing high-resolution light field videos.

Computational Photography combines plentiful computing, digital sensors, modern optics, actuators, probes, and smart lights to escape the limitations of traditional film cameras and enables novel imaging applications. This book provides a practical guide to topics in image capture and manipulation methods for generating compelling pictures for graphics, special effects, scene comprehension, and art. The computational techniques discussed cover topics in exploiting new ideas in manipulating optics, illumination, and sensors at time of capture. In addition, the authors describe sophisticated reconstruction procedures from direct and indirect pixel measurements that go well beyond the traditional digital darkroom experience.

A comprehensive and up-to-date textbook and reference for computational imaging, which combines vision, graphics, signal processing, and optics. Computational imaging involves the joint design of imaging hardware and computer algorithms to create novel imaging systems with unprecedented capabilities. In recent years such capabilities include cameras that operate at a trillion frames per second, microscopes that can see small viruses long thought to be optically irresolvable, and telescopes that capture images of black holes. This text offers a comprehensive and up-to-date introduction to this rapidly growing field, a convergence of vision, graphics, signal processing, and optics. It can be used as an instructional resource for computer imaging courses and as a reference for professionals. It covers the fundamentals of the field, current research and applications, and light transport techniques. The text first presents an imaging toolkit, including optics, image sensors, and illumination, and a computational toolkit, introducing modeling, mathematical tools, model-based inversion, data-driven inversion techniques, and hybrid inversion techniques. It then examines different modalities of light, focusing on the plenoptic function, which describes degrees of freedom of a light ray. Finally, the text outlines light transport techniques, describing imaging systems that obtain micron-scale 3D shape or optimize for noise-free imaging, optical computing, and non-line-of-sight imaging. Throughout, it discusses the use of computational imaging methods in a range of application areas, including smart phone photography, autonomous driving, and medical imaging. End-of-chapter exercises help put the material in context.

This book traces progress in photography since the first pinhole, or camera obscura, architecture. The authors describe innovations such as photogrammetry, and omnidirectional vision for robotic navigation. The text shows how new camera architectures create a need to master related projective geometries for calibration, binocular stereo, static or dynamic scene understanding. Written by leading researchers in the field, this book also explores applications of alternative camera architectures.

Acquiring 3D geometry of the scene is a key task in computer vision. Applications are numerous, from classical object reconstruction and scene understanding to the more recent visual SLAM and autonomous driving. Recent advances in computational imaging have enabled many new solutions to tackle the problem of 3D reconstruction. By modifying the camera's components, computational imaging optically encodes the scene, then decodes it with tailored algorithms. ☐ This dissertation focuses on exploring new computational imaging techniques, combined with recent advances in deep learning, to infer 3D geometry of the scene. In general, our approaches can be categorized into active and passive 3D sensing. ☐ For active illumination methods, we propose two solutions: first, we present a multi-flash (MF) system implemented on the mobile platform. Using the sequence of images captured by the MF system, we can extract the depth edges of the scene, and further estimate a depth map on a mobile device. Next, we show a portable immersive system that is capable of acquiring and displaying high fidelity 3D reconstructions using a set of RGB-D sensors. The system is based on structured light technique and is able to recover 3D geometry of the scene in real time. We have also developed a visualization system that allows users to dynamically visualize the event from new perspectives at arbitrary time instances in real time. ☐ For passive sensing methods, we focus on light field based depth estimation. For depth inference from a single light field, we present an algorithm that is tailored for barcode images. Our algorithm analyzes the statistics of raw light field images and conducts depth estimation with real time speed for fast refocusing and decoding. To mimic the human vision system, we investigate the dual light field input and propose a unified deep learning based framework to extract depth from both disparity cue and focus cue. To facilitate training, we have created a large dual focal stack database with ground truth disparity. While above solution focuses on fusing depth from focus and stereo, we also exploit combing depth from defocus and stereo, with an all-focus stereo pair and a defocused image of one of the stereo views as input. We have adopted the hourglass network architecture to extract depth from the image triplets. We have then studied and explored multiple neural network architectures to improve depth inference. We demonstrate that our deep learning based approaches preserve the strength of focus/defocus cue and disparity cue while effectively suppressing their weaknesses.