Search

Colour and texture are two perceptual stimuli that determine, to a great extent, the appearance of objects, materials and scenes. The ability to process texture and colour is a fundamental skill in humans as well as in animals, therefore, reproducing such capacity in artificial (`intelligent’) systems has attracted considerable research attention since the early 70s. Whereas the main approach to the problem was essentially theory-driven (`hand-crafted’) up to not long ago, in recent years the focus has moved towards data-driven solutions (deep learning). In this overview we retrace the key ideas and methods that have accompanied the evolution of colour and texture analysis over the last five decades, from the `early years’ to convolutional networks. Specifically, we review geometric, differential, statistical and rank-based approaches. Advantages and disadvantages of traditional methods vs. deep learning are also critically discussed, including a perspective on which traditional methods have already been subsumed by deep learning or would be feasible to integrate in a data-driven approach.

As training Deep Neural Networks (DNNs) becomes more expensive, the interest in protecting the ownership of the models with watermarking techniques increases. Uchida et al. proposed a digital watermarking algorithm that embeds the secret message into the model coefficients. However, despite its appeal, in this paper, we show that its efficacy can be compromised by the optimization algorithm being used. In particular, we found through a theoretical analysis that, as opposed to Stochastic Gradient Descent (SGD), the update direction given by Adam optimization strongly depends on the sign of a combination of columns of the projection matrix used for watermarking. Consequently, as observed in the empirical results, this makes the coefficients move in unison giving rise to heavily spiked weight distributions that can be easily detected by adversaries. As a way to solve this problem, we propose a new method called Block-Orthonormal Projections (BOP) that allows one to combine watermarking with Adam optimization with a minor impact on the detectability of the watermark and an increased robustness.

The aim of this paper is to describe a general framework for texture analysis which we refer to as the HEP (histograms of equivalent patterns). The HEP, of which we give a clear and unambiguous mathematical definition, is based on partitioning the feature space associated to image patches of predefined shape and size. This task is approached by defining, a priori, suitable local or global functions of the pixels' intensities. In a comprehensive survey we show that diverse texture descriptors, such as co-occurrence matrices, gray-level differences and local binary patterns, can be seen all to be examples of the HEP. In the experimental part we comparatively evaluate a comprehensive set of these descriptors on an extensive texture classification task. Within the class of HEP schemes, improved local ternary patterns (ILTP) and completed local binary patterns (CLBP) emerge as the best of parametric and non-parametric methods, respectively. The results also show the following patterns: (1) higher effectiveness of multi-level discretization in comparison with binarization; (2) higher accuracy of parametric methods when compared to non-parametric ones; (3) a general trend of increasing performance with increasing dimensionality; and (4) better performance of point-to-average thresholding against point-to-point thresholding. Ministero dell’Istruzione, dell’Università e della Ricerca (Italy) and Mondial Marmi S.r.l. (Italy) | Ref. 39554 Ministerio de Ciencia en Innovación | Ref. TRA2011- 29454-C03-01 Ministerio de Economía y Competitividad | Ref. CTM2010-16573

Multimedia contents play an important role in our society. They serve as a means of communication and can be used not only as an entertainment, but also to inform or even to disseminate knowledge. The increasing relevance of multimedia contents, such as digital images, audio, or video sequences, has been tied to the development of editing software tools enabling their adjustment and enhancement, but ultimately allowing an unskilled person to easily manipúlate them. As a consequence, their credibility as a source of information has been questioned and an important concern has arisen regarding their authenticity. With the aim of recovering trust on multimedia objects, this thesis presents new techniques to detect and localize forgeries, but likewise to infer information about the processing history undergone by a multimedia content. The design of the proposed approaches is based on the theoretical analysis of characteristic traces or footprints that emerge from the application of certain processing to multimedia contents. In this thesis the derived research work encompassing multimedia forensics is divided in two parts. The first part tackles the study of the resampling operation applied when a geometric transformation is performed to adapt a forged content to a genuine scene. The modeling of the resampling operation is addressed from different perspectives, establishing connections between this problem and other similar ones arising in distinct fields, and finally taking advantage of concepts from cyclostationarity theory, set-membership theory, or linear algebra, among others. We design different strategies for resampling factor estimation to characterize the particular transformation applied, providing estimates of the scaling factor or the rotation angle. The case of resampling detection is also considered to unveil the presence of resampling traces. The second part of the thesis is focused on the forensic analysis of video compressed sequences. We start exposing the presence of a new footprint stemming from the double compression of video streams. By exploiting this feature, the detection of double encoding and the estimation of part of the processing history of a double compressed video are further investigated. Then, being capable of extracting information from the first compression, we move to the localization of intra-frame forgeries by applying a subsequent double quantization analysis.

Gabor filtering is a widely adopted technique for texture analysis. The design of a Gabor filter bank is a complex task. In texture classification, in particular, Gabor filters show a strong dependence on a certain number of parameters, the values of which may significantly affect the outcome of the classification procedures. Many different approaches to Gabor filter design, based on mathematical and physiological consideration, are documented in literature. However, the effect of each parameter, as well as the effects of their interaction, remain unclear. The overall aim of this work is to investigate the effects of Gabor filter parameters on texture classification. An extensive experimental campaign has been conducted. The outcomes of the experimental activity show a significant dependence of the percentage of correct classification on the smoothing parameter of the Gabor filters. On the contrary, the correlation between the number of frequencies and orientations used to define a filter bank and the percentage of correct classification appeared to be poor. Consiglio Nazionale delle Ricerche (Italy) | Ref. 1402 Xunta de Galicia | Ref. PGIDIT04REM303003PR

La visión artificial es una disciplina transversal ampliamente estudiada en las últimas décadas. Los sistemas de visión artificial han sido implantados con éxito en procesos industriales tales como automatización, inspección o control de calidad, en sistemas de vigilancia o de seguridad, en procesos médicos, etc. Sus objetivos y ventajas van desde la mejora en la velocidad de los procesos a la automatización de tareas que no pueden hacerse de otro modo. Sin embargo, a día de hoy, los sistemas de adquisición y las condiciones de entorno siguen siendo un factor muy importante a la hora de marcar las limitaciones de uso del sistema. Es frecuente encontrar procesos en los que no es posible interferir, por lo que no se pueden obtener las condiciones de iluminación, espacio, ausencia de obstáculos, etc., necesarias para la implantación de un sistema de visión. En la presente Tesis Doctoral abordaremos el problema de analizar y desarrollar técnicas de visión artificial robustas capaces de funcionar en entornos no controlados. Para ello nos centraremos en varios campos de aplicación, en los que propondremos varias contribuciones con el objetivo de mejorar el rendimiento de los sistemas en entornos realistas: adquisición de imagen, lectura de texto y reconocimiento facial. En primer lugar, nos centraremos en el problema de la adquisición de imágenes de la mayor calidad posible para las posteriores etapas de procesado. Este es un aspecto fundamental para poder mejorar el rendimiento de los sistemas. Propondremos aquí un mecanismo de evaluación de calidad de imagen orientado al ajuste inicial de sistemas de adquisición para visión artificial. Este es un método que se engloba dentro de las técnicas de medida de calidad relativa de imagen. En segundo lugar, abordaremos un escenario de lectura automática de texto, orientado a la automatización en la lectura de instrumentación digital en procesos de calibración. Propondremos aquí varias contribuciones. En primer lugar presentaremos un método para la localización automática de texto en pantallas de dispositivos. A continuación pasaremos a presentar un método en el ámbito del reconocimiento de caracteres, en donde propondremos una serie de características blandas para la clasificación. Finalmente, propondremos un sistema de lectura completo (end-to-end) con el objetivo de automatizar la lectura de pantallas de dispositivos digitales, alcanzando unas prestaciones adecuadas para su uso en escenarios realistas. En tercer lugar, nos centraremos en el problema del reconocimiento facial en dispositivos móviles. El reconocimiento facial es una disciplina dentro de la visión artificial en la que se han realizado grandes avances en los últimos tiempos. Con la aparición de smartphones y tablets en los últimos cinco años, la aplicación del reconocimiento facial al control de acceso y seguridad de la información en movilidad ha cobrado un enorme interés. Estos entornos de movilidad son escenarios no controlados especialmente complejos, tanto por sus condiciones cambiantes como por sus restricciones computacionales. Propondremos aquí varias contribuciones. En primer lugar se propondrá un método de localización precisa de ojos para aplicación a reconocimiento facial en condiciones realistas. En el mismo ámbito, también se propondrán un marco de evaluación exigente que permita la comparación justa de algoritmos de localización de ojos en condiciones no controladas. A continuación, se propondrá un método de verificación de vida colaborativo para su uso como medida contra ataques de suplantación de identidad en sistemas de acceso móvil mediante reconocimiento facial. Finalmente, se realizará una evaluación de rendimiento de algoritmos de procesado en dispositivo móvil para su uso en reconocimiento facial y se propondrá un esquema de reconocimiento completo y robusto, capaz de funcionar íntegramente en el dispositivo. Finalmente, revisaremos la transferencia tecnológica alcanzada con las investigaciones realizadas y las contribuciones propuestas, mostrando así la versatilidad y aplicabilidad de los trabajos realizados en escenarios reales que van desde la automatización industrial a la seguridad en movilidad. Computer vision is a transversal discipline extensively studied in the last decades. Computer vision systems have been successfully introduced in industrial processes, such as automation, inspection or quality control, in surveillance or security processes, in medical applications, etc. Its goals and advantages range from the improvement in speed to the automation of tasks that cannot be done otherwise. However, nowadays the acquisition systems and the environment conditions are still a very important factor to set the limitations of the system use. It is common to find processes in which it is not possible to interfere, so the lighting and space conditions, absence of obstacles and other necessary conditions to introduce a computer vision system cannot be achieved. In this Thesis the problem of achieving robust computer vision techniques in uncontrolled environments will be addressed. To do so, we will focus in several application fields in which we will propose different contributions aiming to improve the performance of the systems in realistic environments. Firstly, we will focus on the acquisition stage and the problem of capturing the best image possible, for being used in the following stages. This is an essential aspect for us to be able to improve the efficiency of the systems. We will propose an image quality assessment method intended to the initial adjustment of acquisition systems. Secondly, we will address an automatic text reading scenario, focused on the automation of digital instrumentation reading in calibration processes. We will propose here several contributions. Firstly we will present a method for automatic localization of text in digital screens. Next, we will present a method in the field of character recognition, in which a series of soft features for classification will be proposed. And finally, we will propose an end-to-end reading system with the aim of automating the reading in digital devices, achieving good performance suitable for its use in real scenarios. Thirdly, we will focus on the problem of face recognition in mobile devices. Facial recognition is a computer vision discipline in which big advances have been made in the last years. The growth of the smartphone and tablet market in the last five years encouraged the interest in using face recognition for information access control and security in these devices. These mobile environments are especially complex uncontrolled scenarios, not only due to the changing conditions, but also because of the computing restrictions. Several contributions will be proposed in this area. Firstly, a precise eye localization method for facial recognition under realistic conditions will be suggested. In the same sphere, a challenging benchmark will be designed. This will allow the fair comparison between algorithms for eye localization under realistic conditions. Next, a collaborative liveness detection method for its use against identity theft attacks will be proposed. Finally, we will make a performance evaluation of facial recognition processing algorithms in mobile devices. We will propose a whole and robust recognition scheme, able to work completely embedded in the device. Finally we will review the technology transfer achieved through the research and the proposed contributions in this Thesis. We will show the versatility and applicability of the developed work in realistic scenarios, covering from industrial automation to other scenarios like mobile security.

Carrier removal is a key step in the Fourier transform method (FTM) of fringe pattern analysis because it can give rise to significant errors in the recovered phase. The existing methods of fringe carrier removal in the FTM are reviewed and a comparison of three different methods, translation of the sidelobe to the frequency origin, least-squares fit and subtraction of the phase of the undeformed carrier fringes, are presented. Computer-generated fringe patterns are used to determine the differences between the original and the retrieved phase distributions. Several figures of merit are proposed to assess the performance of the mentioned methods. Experimental fringe patterns, obtained by double-pulsed-subtraction TV holography, are analyzed by the methods considered here, and the retrieved phase distributions are also compared.