<p dir="ltr">In high-performance building design, it is important to account for the dynamic influence of daylight on humans, as its non-visual effects significantly contribute to the regulation of various physiological and psychological functions. Furthermore, effective and controlled use of daylighting can lower energy consumption for electric lighting, while also minimizing internal lighting gains, excessive solar heat gains, and cooling energy demand. However, there are challenges in choosing appropriate metrics for modeling individual visual preferences and integrating them into control strategies, especially in smart control systems for high-performance buildings that demand self-tuning and personalized functionalities. Therefore, this Thesis aims to develop reliable features and a learning framework that reflect the occupant's visual preferences and can be incorporated into optimal daylighting control strategies using a low-cost high dynamic range imaging (HDRI) camera and deep learning techniques.</p><p dir="ltr">First, this Thesis presents a new method for classifying daylighting preferences based on deep learning models trained with pixel-wise similarity features extracted from pairs of luminance maps. A new composite luminance similarity index was developed, which utilizes the pixel-wise information from the entire luminance distribution and considers both the direction and magnitude of relative luminance change, instead of instantaneous metrics used in previous studies. The generated luminance and contrast similarity maps were directly used for training convolutional neural network (CNN) models to classify the occupant’s visual preferences. The results proved the superiority of the luminance similarity index map as a preference indicator variable. In contrast, common static lighting parameters could not estimate daylight preferences even when used in powerful computational models; they neglected visual information located in various parts of the visual scene and could not consider the change in perceived luminance distribution.</p><p dir="ltr">Second, this Thesis presents a novel method for inferring the relative degree of personal visual preference from pairs of luminance maps using convolutional autoencoder (CAE) and relative ranking concepts. There are practical challenges to utilizing trained CNN-based visual preference classification models for inferring the most preferable visual condition. Therefore, two-stage training was proposed starting from developing a CAE-based feature extraction module to make the model updatable from unseen luminance map characteristics and implementing the trained feature extractor to the visual preference inference model. To select the most preferable luminance distribution among the observed visual environments, the relative ranking concept was implemented in the CAE-based visual preference inference model in addition to binary classification layers. Then, the L2 norm and Euclidean distance were applied to determine the appropriate adjustment directions by analyzing the degree of difference between the captured luminance distribution and the inferred individual preferred luminance distribution. This analysis focused on the condensed latent pixels representing the window and background areas in each luminance distribution.</p><p dir="ltr">Finally, this Thesis expands the scope of utilizing low-cost HDRI sensors and deep learning techniques by demonstrating real-time monitoring of dynamic internal and solar heat gains in office spaces that are required for demand-driven control. For monitoring changes in occupancy, equipment, lighting, and window status in real-time, a convolutional neural network (CNN)-based multi-head classification model was developed and trained with High Dynamic Range (HDR) images, collected using a low-cost fisheye camera in offices. Then, to evaluate the impact of real-time monitoring of heat gains on energy demand, the open plan office space used for the experimental dataset collection was modeled using EnergyPlus software using (i) commonly assumed fixed schedules for occupancy, equipment, and lighting and (ii) real-time monitored dynamic schedules for internal and solar gain components under the same weather conditions. The results showed that the recommended fixed schedules may lead to significant errors in estimated internal and solar gains. The largest discrepancy was noted for occupancy and equipment usage, but other categories also showed both underestimation and overestimation of thermal load components.</p>

Cette thèse présente des contributions sur différents aspects du ''display retargeting'' dans le cadre de l'imagerie HDR (pour High Dynamic Range imaging en anglais). Bien que les contributions soient diverses, elles sont motivées par notre conviction que la préservation de l'intention artistique et la prise en compte de caractéristiques en termes de perception du système visuel humain sont essentielles pour un ''display retargeting'' esthétiquement et visuellement confortable. / This thesis presents various contributions in display retargeting under the vast field of High Dynamic Range (HDR) imaging. The motivation towards this work is the conviction that by preserving artistic intent and considering insights from human visual system leads to aesthetic, comfortable and efficient display retargeting.

In the past decade, camera manufacturers have offered smaller form factors, smaller pixel sizes (leading to higher resolution images), and faster processing chips to increase the performance of consumer cameras. However, these conventional approaches have failed to capitalize on the spatio-temporal redundancy inherent in images, nor have they adequately provided a solution for finding $3$D point correspondences for cameras sampling different bands of the visible spectrum. In this thesis, we pose the following question---given the repetitious nature of image patches, and appropriate camera architectures, can statistical models be used to increase temporal, structural, or spectral resolution? While many techniques have been suggested to tackle individual aspects of this question, the proposed solutions either require prohibitively expensive hardware modifications and/or require overly simplistic assumptions about the geometry of the scene. We propose a two-stage solution to facilitate image reconstruction; 1) design a linear camera system that optically encodes scene information and 2) recover full scene information using prior models learned from statistics of natural images. By leveraging the tendency of small regions to repeat throughout an image or video, we are able to learn prior models from patches pulled from exemplar images. The quality of this approach will be demonstrated for two application domains, using low-speed video cameras for high-speed video acquisition and multi-spectral fusion using an array of cameras. We also investigate a conventional approach for finding 3D correspondence that enables a generalized assorted array of cameras to operate in multiple modalities, including multi-spectral, high dynamic range, and polarization imaging of dynamic scenes.

<p>Imaging and computer vision systems deployed in real-world environments face the challenge of accommodating a wide range of lighting conditions. However, the cost, the demand for high resolution, and the miniaturization of imaging devices impose physical constraints on sensor design, limiting both the dynamic range and effective aperture size of each pixel. Consequently, conventional CMOS sensors fail to deliver satisfactory capture in high dynamic range scenes or under photon-limited conditions, thereby impacting the performance of downstream vision tasks. In this thesis, we address two key problems: 1) exploring the utilization of spatial multiplexing, specifically spatially varying exposure tiling, to extend sensor dynamic range and optimize scene capture, and 2) developing techniques to enhance the robustness of object detection systems under photon-limited conditions.</p> <p><br></p> <p>In addition to challenges imposed by natural environments, real-world vision systems are susceptible to adversarial attacks in the form of artificially added digital content. Therefore, this thesis presents a comprehensive pipeline for constructing a robust and scalable system to counter such attacks.</p>