Special Sessions

This session is a tribute to the memory of Vicent Caselles, who passed away in August 2013, leaving abundant and remarkable contributions to image processing. Vicent Caselles was one of the most gifted mathematicians of his time and had a tremendous interest for practical applications. He made some of the most elegant contributions to the area of image processing, showing that great mathematical ideas can lead to great algorithms for real and practical problems; he was a vanguard in the community of applied math for image processing. This session focuses on recent research contributions related to areas where the pioneering work of Vicent Caselles had an essential impact, focusing specifically on studies close to the work of his last years, such as scale-space problems, as well as color and 3-D analysis which often involve both variational methods and morphological operations. The special session—reflecting Prof. Caselles personality and scientific career—will be composed of works by a combination of senior and junior researchers. The contributors will present recent results that have been significantly influenced by the legacy of Prof. Caselles. The session will thus be at the highest possible scientific level while at the same time honoring and celebrating the life of one of the most brilliant contributors in the area.

The dimensionality of image and video samples is huge. Fortunately, natural imagery tends to lie in low-dimensional manifolds embedded in this high-dimensional space. Discovering the fundamental dimensions (features) of the image subspace and describing its structure are key for image processing and understanding. Human vision is closely tied to these statistical regularities as well. For example, knowledge of this structure is fundamental for an efficient resource allocation in image coding; noise breaks the structure of the manifold resulting in samples in unusual places so that learning/enforcing the right features is central in image restoration and regression; characterization of clusters depending on the nature of the images is key in image recognition and classification; and, finally, structure of these curved manifold subspaces is important to propose meaningful distance measures between images, which is the challenge central to image retrieval and image-quality assessment. As a consequence, a manifold-oriented approximation to image processing cuts to the heart of nonlinear feature extraction, while manifold regularization encompasses the inclusion of prior knowledge, the learning and encoding of different types of invariance/redundancy, and the characterization of spatial-temporal-spectral texture. Feature learning is interdisciplinary in nature: on the one hand, automatic discovery of relevant directions in image space (machine learning) may explain the behavior of the human visual system (neuroscience), and—interestingly for the engineering community—perceptually relevant features may unveil unexpected statistical meaning, and vice versa. This special session focuses on raising key scientific questions that need to be addressed in the context of image processing: the study of the intrinsic dimensionality of images and how to discover low-rank subspaces; dictionary learning for different criteria such as error minimization, information maximization, or classification performance; the domain-adaptation issue; how to evaluate the quality of features or their perceptual relevance; and the effect of invariance or certain symmetries on the appropriate image representation, or the metric-learning issue.

Plenoptics is an emerging field that unifies a number or research disciplines ranging from signal processing, computer vision, and computer graphics to human visual perception, physics, and optics. In short, plenoptics is the theory of the light field and the processing thereof. While the areas of plenoptic capture, processing, and reconstruction were founded by pioneering work ranging back over the last century, rapid development in the technology of optics, electrical engineering, and computing in recent years has made practical applications of early plenoptic theory possible. Consequently, there is increasing interest in deploying plenoptics for next-generation 3D capture, representation, and display with applications such as computational photography, accurate recognition in security systems, personalized 3D televisions, ultra-realistic telepresence, and industrial optical inspections. This special session will be devoted to recent advances in the plenoptics field.

This special session will address the problem of photon-limited image reconstruction. Optical imaging systems, such as CCD sensors, acquire an image by integrating at each pixel the number of incoming photons during the exposure time. This process is known to be subject to counting noise, which corrupts the signal of interest as much as the expected number of photons is low. Hence, reconstruction of the signal becomes problematic under low-light conditions with short exposure time (typical examples include night vision, astronomy, and microscopy). Surprisingly, while such behavior is inherent to optical systems, most reconstruction algorithms ignore it. A possible reason for this is that handling Poisson-type degradation usually lead to very difficult optimization and statistical problems. In particular, Poisson models are far more intricate than commonly considered Gaussian degradation models. Moreover, one might need to redesign entire algorithms to treat such noise models directly, explaining why focus has typically fallen on transformations to render data to be approximately Gaussian (e.g., the Anscombe transform). Nevertheless, direct methods have recently emerged that have proved their ability to deal with such Poisson models; these methods have shown clear benefit, especially in low-light conditions. This special session aims to disseminate the latest solutions developed for the problems of noise estimation, denoising, and deconvolution that rely on suitable models of photon-limited corruptions.

Optical sensing has come a long way—from grayscale to multispectral to hyperspectral images. The advances in imaging hardware over recent decades have enabled availability of imagery with high spatial, spectral, and temporal resolution for a variety of applications. Hyperspectral imagery, also called imaging spectroscopy, consists of acquiring images of a given area using a large number (typically a few hundreds) of narrow and contiguous spectral bands, covering a wide range of the electromagnetic spectrum from the visible to the infrared domain. Compared to standard color imagery (with 3 spectral bands covering the red, green, and blue), or compared to standard multispectral imagery (typically up to a dozen spectral bands in the visible and near-infrared domains), hyperspectral data provides a fine description of the chemical components in the sensed materials, thereby permitting their detection, discrimination, and characterization. In standard gray-level imagery, two pixels having the same intensity, or luminance, can actually correspond to different colors. In a similar way, two pixels having the same color (or perceived as identical by a human observer) in standard color imagery can indeed correspond to different materials. However, these materials can be distinguished thanks to the spectral diversity provided by hyperspectral sensors. Such advances have created unique challenges for researchers in the remote-sensing community working on algorithms for representation, exploitation, and analysis of such data. At the same time, availability of hyperspectral-imaging capabilities for a wide variety of applications has grown substantially in the past decade, owing primarily to lower hardware costs of imaging systems operating in the visible, very near-infrared, and short-wave infrared regions of the electromagnetic spectrum. This special session will be devoted to recent advances in image-processing techniques designed specifically for the unique characteristics of hyperspectral imagery, particularly in the following important and emerging topical areas: image analysis via classification, change detection, segmentation, spectral unmixing for interpretation of images containing sub-pixel objects, effective utilization of sparse representation and convex-optimization approaches for hyperspectral image analysis, compressive sensing of hyperspectral imagery, and sensor fusion.

Compactly representing visual content through local features (such as SIFT or SURF) is of fundamental importance to enabling critical content-based image retrieval (CBIR), such as the search through large image repositories or sensor-based scene analysis. Compact descriptors are essential to efficiently browse the huge catalog of visual content made available through online social-sharing platforms using content-based retrieval approaches such as the query-by-example paradigm. At the same time, concise and computationally efficient feature-based representations have led to the rise of a broad spectrum of distributed services—especially in the field of visual sensor networks—which blend together several diverse disciplines, from computer vision to human-computer interaction and advanced sensing. Examples of these application scenarios include augmented reality using visual awareness; indoor localization with lightweight infrastructure; gesture-based human-computer interaction; advanced surveillance; and remote sensing. In the last few years, the interest in efficient visual search has led to a few commercial recognition/retrieval services, such as Google Goggles and Kooaba, and the need for succinct—yet accurate—visual descriptors has motivated MPEG to initiate standardization activities addressing this need. The latest standardization effort is represented by the Compact Descriptors for Visual Search (CDVS) initiative, which targets the efficient coding of visual descriptors extracted from static images to support visual search from mobile devices. However, 2D images represent only a portion of the wide variety of media and types of content available nowadays, which include dynamic 2D and 3D video, as well as “texture+depth” and high-dynamic-range content. Therefore, CDVS can be considered as merely a first step towards the development and standardization of content-based multimedia-search solutions. This special session looks to gather and discuss efficient techniques to compactly represent and process dynamic visual content. The goal is to provide solutions that will meet the emerging requirements of content indexing and browsing beyond the world of 2D static images considered so far, e.g., by enabling cross-representation queries, new feature-based services, and efficient coding of descriptors for dynamic and 3D content.

Optimization is at the heart of many problems in modern image processing. It turns out that many of the objective functionals involved in these problems have a structure that can be exploited to design fast, effective, and provably convergent splitting algorithms. This special session will cover a wide spectrum of recent advances in optimization that solve challenging inverse problems, considering contributions from both theoretical as well as application perspectives.

In spite of well-established stereoscopic video technologies, the future of 3D multimedia systems is currently being challenged due to its unconvincing uptake by both content producers and consumers. In order to address this issue, this special session will identify and address the emerging challenges of quality of experience (QoE) in 3D multimedia systems, encompassing research from a wide range of related fields: display technology, multimedia communications, perceptual and cognitive psychology, signal processing, and human-computer interaction, among others. Thus, this special session will have an inherent interdisciplinary nature which bridges image- and video-processing research within the human-factors community, presenting cutting-edge results as well as demonstrating novel research avenues for 3D media processing under the constraints of QoE management. This special session will be jointly organized by two European Cooperation in Science and Technology (COST) actions: Qualinet—Network on Quality of Experience in Multimedia Systems and Services; and 3D-ConTourNet—3D Content Creation, Coding, and Transmission over Future Media Networks.

Astronomy is a very rich and active field wherein signal and image processing is assuming an ever-growing role. This is especially the case for recent and upcoming projects wherein advanced image-processing techniques is becoming crucial: inverse problems; compressed sensing; inference in optical, IR, or radio interferometry in projects like ALMA, Lofar, SKA or the VLT; blind source separation; parametric estimation; statistical analysis in Planck or the future Euclid space missions; high-dimensional and big-data analysis; and machine learning. This special session is intended to cover a large range of astronomical fields, observation techniques (optical, IR, radio interferometry, hyperspectral, etc.), as well as advanced image-processing topics (Bayesian inference, sparsity, compressed sensing, inverse problems, parametric estimation, etc.). Toward that purpose, this session will give voice to speakers from the applied-mathematics, signal/image-processing, and astro-statistics communities, emphasizing, in particular, cutting-edge data-processing methods over a wide range of data-analysis issues in astronomy.

Materials science is evolving from materials discovered in Nature by chance to designed materials that repair themselves, adapt to their environment, capture and store energy or information, or help elaborate new devices. Materials are now designed from scratch with initial blueprints, starting from atoms and molecules. This evolution, at the confluence of science, technology, and engineering, is driven by the synergy of materials science and physics, mechanics, chemistry, biology, and engineering, with image processing taking part in this challenge. Indeed, the possibility of designing, analyzing, and modeling materials from images (or generally two- or three-dimensional modalities) has yielded important contributions to the materials-science field. The appearance of materials changes significantly with imaging techniques, depending on the scale of analysis, imaging settings, physical properties, as well as preparation of materials, and understanding these aspects is critical for material analysis and modeling. This special session will target relevant problems in material characterization that can be addressed with classical or advanced methods of signal and image processing with a focus on techniques that employ methods such as restoration, segmentation, mathematical morphology, texture analysis, multiscale- and directional-feature extraction, color and multispectral processing, and stochastic models.

More and more of our interactions with others are online in social networks. To date, however, online interactions have been a poor substitute for real human interactions. Consequently, 3D virtual environments which allow safe and enjoyable collaborative interaction while bringing together realistic interpersonal communication and interaction with 3D media creation are crucial. Such environments will permit users to meet, socialize, and share experiences using equipment that they already have at home. This special session will to bring together researchers and engineers working on the next-generation of immersive and realistic 3D virtual environments in which virtual humans can interact remotely. The primary objective is to present and discuss key research issues related to the generation of immersive and 3D cooperative virtual worlds as well as the reconstruction, simulation, and animation of virtual humans.

Three-dimensional (3D) structural studies of biological matter, from proteins to whole cells, are of a great importance for fully understanding the function of macromolecular complexes and organelles within cells. The 3D structure of a cellular component is tightly related to its function within a cell, and knowledge of both the structure and the function is necessary, for instance, to design drugs whose targets are specific proteins. The quality of the 3D structure is usually expressed in terms of resolution, which measures the level of detail contained in the structure. High resolution means the possibility of interpreting the structure at a high level of detail, which is mandatory for understanding its function. Several complementary techniques have been developed for determining the 3D structure of biological specimens. In the case of molecules that can either auto-assemble or be coerced to assemble in 3D crystals, X-ray diffraction is traditionally used to determine their atomic structure. However, a large number of macromolecules diffract poorly or cannot be crystallized because of their flexibility. For large macromolecular complexes, cellular components, or small entire cells, 3D transmission electron microscopy (TEM) is used. This special session will address recently developed image-processing methods that solve several currently important TEM image-processing problems. The perspectives of the field at the intersection between biology, mathematics, informatics, and physics will also be discussed.

This special session will provide a forum for dissemination of the most recent results and ideas in the field of face recognition and morpho-functional sign analysis, a field with potential for deep impact in the health, security, social, and entertainment domains. It is expected that the session will attract researchers working in the emerging interface of imaging, image processing, psychology, and medicine. In the recent years, the field of facial-sign analysis has seen a number of significant advances, both in imaging technology as well as software analysis tools. The research in this field has so far been able to successfully solve a number of problems including face recognition and, to some extent, interpretation of facial articulations. The next big challenges for facial-sign analysis include the development of techniques leading to machine-based recognition of human emotion and intentions as well as the inference of health signs. The latter is particularly interesting since, if successful, it would enable fast, inexpensive, and non-intrusive means of monitoring a person’s well-being status. This special session is inspired by a recent project funded by the FP7 addressing exactly such issues. The objective of the special session is to bring researchers working in this area to present the most current advances in facial-sign analysis and interpretation.

This special session will focus on modern image-processing techniques relevant to synthetic-aperture-radar (SAR) images, focusing on both fundamentals of the involved processing techniques as well as applications in the SAR area. A major goal is also to extend mutual interaction between the SAR and image-processing communities by creating a forum for fruitful discussions.

Easy electronic access over digital distribution channels creates difficulties for copyright protection of multimedia digital. The topic of this special session—3D data security—includes various aspects such as the watermarking of 3D objects, enforcement of various aspects of digital rights management (DRM) in 3D data, copyright protection, fingerprinting, graphical-model authentication, cryptography, and security in general of 3D data and objects. These topics have a large number of applications which go beyond DRM and include the organization of virtual-object databases, digital cinematography, virtual reality, 3D computer graphics, and computer-aided design (CAD) in the context of 3D printing. The security in general of 3D models has recently received a renewed interest due to the advances in 3D printing wherein copyright infringement of design plans used in 3D manufacturing is of particular interest. Additionally, security and privacy considerations in applications using 3D models and data representations have become increasingly important in recent years. The special session will have a clear interdisciplinary aspect, involving the latest research results from 3D signal processing, computer graphics, 3D data compression, cryptography, data security, and protection.

Deployment of 3D multimedia services and applications has recently gained momentum with the proliferation of supporting high-speed computing and communication technologies. With the advent of the “anytime-anywhere-with-any-device” concept, users nowadays expect the provision of 3D multimedia with seamless continuity and synchronicity of services across ubiquitous yet heterogeneous networking environments. Widespread use of 3D content supported by the recent advances in 3D content-representation and rendering schemes leads to increased load on networking resources due to the amplified data volume. Furthermore, considering an increasingly heterogeneous ecosystem of networks and terminal devices of different capabilities, adaptive processing and distribution of 3D multimedia are essential. Optimization of resources can be achieved by joint consideration of robust 3D multimedia coding and transport mechanisms, with the deployment of cross-layer architectures, overseeing the quality of experience (QoE) of media consumers, wherein QoE is a combination of both network dynamics and source-content attributes. A number of delivery methods ranging from unicast to multicast and broadcast, as well as a broad range of networking environments, from wireless to broadcast networks, IP-based networks, and other Future Media Internet architectures, are all anticipated to facilitate the effective delivery of large volume of 3D multimedia content. To increase effectiveness, it will also be necessary to provide hybrid solutions using two or more delivery technologies at once. At the receiving end, 3D content rendering inherently requires adapting to user context for enhanced experience with multi-speaker audio systems and extended displays in addition to the primary visualization medium where available. This special session will report recent research efforts that cover a variety of related topics including: 3D multimedia coding, processing, and transport techniques; fixed and mobile 3D multimedia applications and systems; error-resiliency methods for robust 3D multimedia distribution; cross-layer architectures for efficient 3D multimedia delivery; hybrid distribution networks and technologies for 3D multimedia delivery; rendering techniques for 3D multimedia content, including advanced displays and spatial audio listening environments; QoE modeling for emerging 3D multimedia services and applications; and 3D multimedia adaptation strategies across the end-to-end distribution chain.

Advances in video compression, which have enabled a proliferation of video through bandwidth-constrained channels, have been critical in the rapid growth of video. As the push for higher coding efficiency, higher resolutions, and more sophisticated multimedia applications continues, the required computations per pixel as well as the pixel-processing rate grow exponentially. This poses significant power and performance challenges for battery-operated devices such as smartphones and tablets, as well as for real-time applications such as video conferencing. Thus, implementation challenges must be taken into consideration when developing signal-processing algorithms for next-generation video-coding systems. The latest video coding standard, High Efficiency Video Coding (HEVC), was finalized in January 2013. Joint design of algorithms and architectures was used in the development of HEVC to overcome various implementation challenges, while still enabling improvement in coding efficiency compared to H.264/AVC. This special session will focus on implementation-friendly video-coding algorithms developed for HEVC and its fidelity-range extensions with particular focus on both software multi-core and hardware implementations.

The purpose of behavior imaging is to capture, model, and analyze social and communicative behaviors during different situations: human to human as well as human to computer, virtual agent, or robot. Long-term goals are challenging with several open issues. This special session will target several of these challenges by bringing together researchers active in different areas of behavior imaging, including human social-behavior modeling and analysis; facial, gesture, and speech analysis for behavior and interaction modeling; human-computer or human-human interaction and their mutual influence during interaction; affective computing; activity analysis; as well as datasets and annotation development for benchmarking. The objective of this special session is to discuss which methodologies and problems may take advantage of new advanced techniques in the field of image processing, computer vision, machine learning, and pattern recognition.

Roads are important man-made infrastructures playing a crucial role for the mobility of people, goods, and merchandise. Road maintenance is essential to ensure correct pavement performance and to preserve its structural integrity; hence, periodic road surveys are needed to evaluate pavement surface condition. Given the size of a typical road network, there is growing demand for high-speed image-acquisition systems coupled with automatic image-analysis tools which are able to overcome limitations of traditional visual pavement-surface inspections by humans. However, high-speed image-acquisition systems originate large amounts of image data that need to be efficiently and accurately processed in order to obtain a reliable assessment of road condition. This special session will overview key progress in this field in terms of imaging-acquisition devices and the subsequent image-processing and analysis.

This special session will address the ways and means whereby privacy-respecting compliance can be best aided through socio-ethically informed design of image-processing and pattern-recognition algorithms and techniques that enhance privacy protection as an inherent property of the multimedia (audio-visual) content-analysis process. The aim is to provide insights on resolution of problems of automatic decision making in privacy-aware systems by examining the achievable accuracy of different state-of-the-art signal-processing and pattern-recognition techniques, by presenting new approaches to enhancing the privacy-protection capability of such systems, and by achieving a privacy-risks impact assessment for the level of privacy protection that can be reliably provided by each proposed privacy filter in each context. Of particular interest will be methodologically-guided approaches for privacy-by-co-design, accountability by design, and social-impact assessment aided by (ir)reversible audio-video analytics techniques to support audio and video privacy. This special session will also include presentation and discussion of approaches to stakeholder engagement, requirements engineering, dynamic requirements prioritization, usability evaluation and adaptation, as well as social-impact analysis and self-audit of privacy-risk mitigation design and deployment approaches, in particular, audio-visual privacy-filtering solutions and their deployment and scalability.