Abstract: Medical imaging feeds research in many computer vision and image processing fields: image filtering, segmentation, shape recovery, registration, retrieval and pattern matching. Because of their low contrast changes and large variety of artifacts and noise, medical imaging processing techniques relying on an analysis of the geometry of image level sets rather than on intensity values result in more robust treatment. From the starting point of treatment of intravascular images, this PhD thesis ad- dresses the design of differential image operators based on geometric principles for a robust shape modelling and restoration. Among all fields applying shape recovery, we approach filtering and segmentation of image objects. For a successful use in real images, the segmentation process should go through three stages: noise removing, shape modelling and shape recovery. This PhD addresses all three topics, but for the sake of algorithms as automated as possible, techniques for image processing will be designed to satisfy three main principles: a) convergence of the iterative schemes to non-trivial states avoiding image degeneration to a constant image and representing smooth models of the originals; b) smooth asymptotic behav- ior ensuring stabilization of the iterative process; c) fixed parameter values ensuring equal (domain free) performance of the algorithms whatever initial images/shapes. Our geometric approach to the generic equations that model the different processes approached enables defining techniques satisfying all the former requirements. First, we introduce a new curvature-based geometric flow for image filtering achieving a good compromise between noise removing and resemblance to original images. Sec- ond, we describe a new family of diffusion operators that restrict their scope to image level curves and serve to restore smooth closed models from unconnected sets of points. Finally, we design a regularization of snake (distance) maps that ensures its smooth convergence towards any closed shape. Experiments show that performance of the techniques proposed overpasses that of state-of-the-art algorithms.

Abstract: More realistic and accurate models of the scene illumination and object reflectance can greatly improve the quality of many computer vision and computer graphics tasks. Using such model, a more profound knowledge about the interaction of light with object surfaces can be established which proves crucial to a variety of computer vision applications. In the current work, we investigate the various existing approaches to illumination and reflectance modeling and form an analysis on their shortcomings in capturing the complexity of real-world scenes. Based on this analysis we propose improvements to different aspects of reflectance and illumination estimation in order to more realistically model the real-world scenes in the presence of complex lighting phenomena (i.e, multiple illuminants, interreflections and shadows). Moreover, we captured our own multi-illuminant dataset which consists of complex scenes and illumination conditions both outdoor and in laboratory conditions. In addition we investigate the use of synthetic data to facilitate the construction of datasets and improve the process of obtaining ground-truth information.

Abstract: Nowadays, digital images are used in many areas in everyday life, but they tend to be big. This increases amount of information leads us to the problem of image data storage. For example, it is common to have a representation a color pixel as a 24-bit number, where the channels red, green, and blue employ 8 bits each. In consequence, this kind of color pixel can specify one of 224 ¼ 16:78 million colors. Therefore, an image at a resolution of 512 £ 512 that allocates 24 bits per pixel, occupies 786,432 bytes. That is why image compression is important. An important feature of image compression is that it can be lossy or lossless. A compressed image is acceptable provided these losses of image information are not perceived by the eye. It is possible to assume that a portion of this information is redundant. Lossless Image Compression is defined as to mathematically decode the same image which was encoded. In Lossy Image Compression needs to identify two features inside the image: the redundancy and the irrelevancy of information. Thus, lossy compression modifies the image data in such a way when they are encoded and decoded, the recovered image is similar enough to the original one. How similar is the recovered image in comparison to the original image is defined prior to the compression process, and it depends on the implementation to be performed. In lossy compression, current image compression schemes remove information considered irrelevant by using mathematical criteria. One of the problems of these schemes is that although the numerical quality of the compressed image is low, it shows a high visual image quality, e.g. it does not show a lot of visible artifacts. It is because these mathematical criteria, used to remove information, do not take into account if the viewed information is perceived by the Human Visual System. Therefore, the aim of an image compression scheme designed to obtain images that do not show artifacts although their numerical quality can be low, is to eliminate the information that is not visible by the Human Visual System. Hence, this Ph.D. thesis proposes to exploit the visual redundancy existing in an image by reducing those features that can be unperceivable for the Human Visual System. First, we define an image quality assessment, which is highly correlated with the psychophysical experiments performed by human observers. The proposed CwPSNR metrics weights the well-known PSNR by using a particular perceptual low level model of the Human Visual System, e.g. the Chromatic Induction Wavelet Model (CIWaM). Second, we propose an image compression algorithm (called Hi-SET), which exploits the high correlation and self-similarity of pixels in a given area or neighborhood by means of a fractal function. Hi-SET possesses the main features that modern image compressors have, that is, it is an embedded coder, which allows a progressive transmission. Third, we propose a perceptual quantizer (½SQ), which is a modification of the uniform scalar quantizer. The ½SQ is applied to a pixel set in a certain Wavelet sub-band, that is, a global quantization. Unlike this, the proposed modification allows to perform a local pixel-by-pixel forward and inverse quantization, introducing into this process a perceptual distortion which depends on the surround spatial information of the pixel. Combining ½SQ method with the Hi-SET image compressor, we define a perceptual image compressor, called ©SET. Finally, a coding method for Region of Interest areas is presented, ½GBbBShift, which perceptually weights pixels into these areas and maintains only the more important perceivable features in the rest of the image. Results presented in this report show that CwPSNR is the best-ranked image quality method when it is applied to the most common image compression distortions such as JPEG and JPEG2000. CwPSNR shows the best correlation with the judgement of human observers, which is based on the results of psychophysical experiments obtained for relevant image quality databases such as TID2008, LIVE, CSIQ and IVC. Furthermore, Hi-SET coder obtains better results both for compression ratios and perceptual image quality than the JPEG2000 coder and other coders that use a Hilbert Fractal for image compression. Hence, when the proposed perceptual quantization is introduced to Hi-SET coder, our compressor improves its numerical and perceptual e±ciency. When ½GBbBShift method applied to Hi-SET is compared against MaxShift method applied to the JPEG2000 standard and Hi-SET, the images coded by our ROI method get the best results when the overall image quality is estimated. Both the proposed perceptual quantization and the ½GBbBShift method are generalized algorithms that can be applied to other Wavelet based image compression algorithms such as JPEG2000, SPIHT or SPECK.

Abstract: One of the fundamental problems of computer vision is to represent images with compact semantically relevant embeddings. These embeddings could then be used in a wide variety of applications, such as image retrieval, object detection, and video search. The main objective of this thesis is to study image embeddings from two aspects: color embeddings and deep embeddings.
In the first part of the thesis we start from hand-crafted color embeddings. We propose a method to order the additional color names according to their complementary nature with the basic eleven color names. This allows us to compute color name representations with high discriminative power of arbitrary length. Psychophysical experiments confirm that our proposed method outperforms baseline approaches. Secondly, we learn deep color embeddings from weakly labeled data by adding an attention strategy. The attention branch is able to correctly identify the relevant regions for each class. The advantage of our approach is that it can learn color names for specific domains for which no pixel-wise labels exists.
In the second part of the thesis, we focus on deep embeddings. Firstly, we address the problem of compressing large embedding networks into small networks, while maintaining similar performance. We propose to distillate the metrics from a teacher network to a student network. Two new losses are introduced to model the communication of a deep teacher network to a small student network: one based on an absolute teacher, where the student aims to produce the same embeddings as the teacher, and one based on a relative teacher, where the distances between pairs of data points is communicated from the teacher to the student. In addition, various aspects of distillation have been investigated for embeddings, including hint and attention layers, semi-supervised learning and cross quality distillation. Finally, another aspect of deep metric learning, namely lifelong learning, is studied. We observed some drift occurs during training of new tasks for metric learning. A method to estimate the semantic drift based on the drift which is experienced by data of the current task during its training is introduced. Having this estimation, previous tasks can be compensated for this drift, thereby improving their performance. Furthermore, we show that embedding networks suffer significantly less from catastrophic forgetting compared to classification networks when learning new tasks.

Abstract: Over the last few years, the number of consumer computer vision applications has increased dramatically. Today, computer vision solutions can be found in video game consoles, smartphone applications, driving assistance – just to name a few. Ideally, we require the performance of those applications, particularly those that are safety critical to remain constant under any external environment factors, such as changes in illumination or weather conditions. However, this is not always possible or very difficult to obtain by only using visible imagery, due to the inherent limitations of the images from that spectral band. For that reason, the use of images from different or multiple spectral bands is becoming more appealing.
The aforementioned possible advantages of using images from multiples spectral bands on various vision applications make multi-spectral image processing a relevant topic for research and development. Like in visible image processing, multi-spectral image processing needs tools and algorithms to handle information from various spectral bands. Furthermore, traditional tools such as local feature detection, which is the basis of many vision tasks such as visual odometry, image registration, or structure from motion, must be adjusted or reformulated to operate under new conditions. Traditional feature detection, description, and matching methods tend to underperform in multi-spectral settings, in comparison to mono-spectral settings, due to the natural differences between each spectral band.
The work in this thesis is focused on the local feature description problem when cross-spectral images are considered. In this context, this dissertation has three main contributions. Firstly, the work starts by proposing the usage of a combination of frequency and spatial information, in a multi-scale scheme, as feature description. Evaluations of this proposal, based on classical hand-made feature descriptors, and comparisons with state of the art cross-spectral approaches help to find and understand limitations of such strategy. Secondly, different convolutional neural network (CNN) based architectures are evaluated when used to describe cross-spectral image patches. Results showed that CNN-based methods, designed to work with visible monocular images, could be successfully applied to the description of images from two different spectral bands, with just minor modifications. In this framework, a novel CNN-based network model, specifically intended to describe image patches from two different spectral bands, is proposed. This network, referred to as Q-Net, outperforms state of the art in the cross-spectral domain, including both previous hand-made solutions as well as L2 CNN-based architectures. The third contribution of this dissertation is in the cross-spectral feature description application domain. The multispectral odometry problem is tackled showing a real application of cross-spectral descriptors
In addition to the three main contributions mentioned above, in this dissertation, two different multi-spectral datasets are generated and shared with the community to be used as benchmarks for further studies.

Abstract: Pedestrian detection continues to be an extremely challenging problem in real scenarios, in which situations like illumination changes, noisy images, unexpected objects, uncontrolled scenarios and variant appearance of objects occur constantly. All these problems force the development of more robust detectors for relevant applications like vision-based autonomous vehicles, intelligent surveillance, and pedestrian tracking for behavior analysis. Most reliable vision-based pedestrian detectors base their decision on features extracted using a single sensor capturing complementary features, e.g., appearance, and texture. These features usually are extracted from the current frame, ignoring temporal information, or including it in a post process step e.g., tracking or temporal coherence. Taking into account these issues we formulate the following question: can we generate more robust pedestrian detectors by introducing new information sources in the feature extraction step?
In order to answer this question we develop different approaches for introducing new information sources to well-known pedestrian detectors. We start by the inclusion of temporal information following the Stacked Sequential Learning (SSL) paradigm which suggests that information extracted from the neighboring samples in a sequence can improve the accuracy of a base classifier.
We then focus on the inclusion of complementary information from different sensors like 3D point clouds (LIDAR – depth), far infrared images (FIR), or disparity maps (stereo pair cameras). For this end we develop a multi-modal framework in which information from different sensors is used for increasing detection accuracy (by increasing information redundancy). Finally we propose a multi-view pedestrian detector, this multi-view approach splits the detection problem in n sub-problems.
Each sub-problem will detect objects in a given specific view reducing in that way the variability problem faced when a single detectors is used for the whole problem. We show that these approaches obtain competitive results with other state-of-the-art methods but instead of design new features, we reuse existing ones boosting their performance.

Abstract: Put succinctly, the bag-of-words based image representation is the most successful approach for object and scene recognition. Within the bag-of-words framework the optimal fusion of multiple cues, such as shape, texture and color, still remains an active research domain. There exist two main approaches to combine color and shape information within the bag-of-words framework. The first approach called, early fusion, fuses color and shape at the feature level as a result of which a joint colorshape vocabulary is produced. The second approach, called late fusion, concatenates histogram representation of both color and shape, obtained independently. In the first part of this thesis, we analyze the theoretical implications of both early and late feature fusion. We demonstrate that both these approaches are suboptimal for a subset of object categories. Consequently, we propose a novel method for recognizing object categories when using multiple cues by separately processing the shape and color cues and combining them by modulating the shape features by category specific color attention. Color is used to compute bottom-up and top-down attention maps. Subsequently, the color attention maps are used to modulate the weights of the shape features. Shape features are given more weight in regions with higher attention and vice versa. The approach is tested on several benchmark object recognition data sets and the results clearly demonstrate the effectiveness of our proposed method. In the second part of the thesis, we investigate the problem of obtaining compact spatial pyramid representations for object and scene recognition. Spatial pyramids have been successfully applied to incorporate spatial information into bag-of-words based image representation. However, a major drawback of spatial pyramids is that it leads to high dimensional image representations. We present a novel framework for obtaining compact pyramid representation. The approach reduces the size of a high dimensional pyramid representation upto an order of magnitude without any significant reduction in accuracy. Moreover, we also investigate the optimal combination of multiple features such as color and shape within the context of our compact pyramid representation. Finally, we describe a novel technique to build discriminative visual words from multiple cues learned independently from training images. To this end, we use an information theoretic vocabulary compression technique to find discriminative combinations of visual cues and the resulting visual vocabulary is compact, has the cue binding property, and supports individual weighting of cues in the final image representation. The approach is tested on standard object recognition data sets. The results obtained clearly demonstrate the effectiveness of our approach.

Abstract: Recent advances in thermal infrared imaging (LWIR) has allowed its use in applications beyond of the military domain. Nowadays, this new family of sensors is included in different technical and scientific applications. They offer features that facilitate tasks, such as detection of pedestrians, hot spots, differences in temperature, among others, which can significantly improve the performance of a system where the persons are expected to play the principal role. For instance, video surveillance applications, monitoring, and pedestrian detection.
During this dissertation the next question is stated: Could a couple of sensors measuring different bands of the electromagnetic spectrum, as the visible and thermal infrared, be used to extract depth information? Although it is a complex question, we shows that a system of these characteristics is possible as well as their advantages, drawbacks, and potential opportunities.
The matching and fusion of data coming from different sensors, as the emissions registered at visible and infrared bands, represents a special challenge, because it has been showed that theses signals are weak correlated. Therefore, many traditional techniques of image processing and computer vision are not helpful, requiring adjustments for their correct performance in every modality.
In this research an experimental study that compares different cost functions and matching approaches is performed, in order to build a multimodal stereovision system. Furthermore, the common problems in infrared/visible stereo, specially in the outdoor scenes are identified. Our framework summarizes the architecture of a generic stereo algorithm, at different levels: computational, functional, and structural, which can be extended toward high-level fusion (semantic) and high-order (prior).The proposed framework is intended to explore novel multimodal stereo matching approaches, going from sparse to dense representations (both disparity and depth maps). Moreover, context information is added in form of priors and assumptions. Finally, this dissertation shows a promissory way toward the integration of multiple sensors for recovering three-dimensional information.