Anastasios Venetsanopoulos - Research Record

Research Record

Introduction. Professor Anastasios (Tas) Venetsanopoulos has a long and productive career in research, education and university administration. Over a period of four decades, he has established himself in the worldwide telecommunications and signal processing community as an outstanding researcher, scholar, professor and consultant. He has made contributions to telecommunications, signal and image processing, multimedia and biometrics research by authoring and co-authoring many journal papers and books. His pioneering and fundamental research contributions and the writing of numerous graduate level books have opened up new vistas in telecommunications, multidimensional filter theory and design, design of non-linear filters, multimedia neural networks, biometric applications, and WLAN positioning systems.

His work has been cited in over 10,000 research papers (Google Scholar) and 400 textbooks. He has been a mentor for over 160 graduate students and post doctoral fellows. He has motivated a generation of engineers in North America and around the world for careers in research and teaching in the areas of signal and image processing, telecommunications, multimedia, and biometrics.

Telecommunications. His early work dealt with the problem of optimum detection and signal design, for communication over purely random, general, linear, time-varying, very noisy undersea acoustic channels. The general undersea acoustic channel was modeled as a random, time-varying, linear filter, consisting of a set of randomly moving and correlated scatterers, distributed over a time-varying, random surface. His results contributed to the improvement of SONAR systems for undersea communications over fading dispersive channels and found later applications over ionospheric and tropospheric channels.

Later publications focused on the issue of image and video compression and made contributions to the area of progressive image transmission (PIT). PIT refers to the coding of still images at increasing levels of precision, the lower being used for rapid image identification. Through PIT, it is possible to expedite such activities as browsing through remote databases. He developed and tested a number of first and second generation morphological pyramidal techniques, which achieved compression ratios of around 100:1 for good quality, lossy, still image transmission. Professor Venetsanopoulos contributed to the area of vector quantization for lossy image compression and developed a number of hierarchical coding techniques for still images. Wavelet techniques for still image compression were also considered, as well as fractal-based techniques compressing and coding still images and video sequences. His latest contributions in telecommunications were in the area of mobility management, with the development of cost effective algorithms for mobile terminal location and determination and WLAN positioning systems. This area has attracted interest for its applications to emergency communications, location-sensitive browsing and resource allocation.

Signal and image processing. Dr. Venetsanopoulos is one of the first Canadian researchers to make contributions to the foundations of two-dimensional and multi-dimensional digital filtering, which are widely used in image and video processing. His early contributions in these areas received interest and a variety of techniques that led to efficient two-dimensional filter design were based on these early contributions. In the eighties, his interest was focused on the area of nonlinear filters. Nonlinear filters are more complex than linear filters but allow additional flexibility and speed in complex applications. In the area of nonlinear filters he contributed theoretical results, introducing new filter families. The “Nonlinear Order Statistics Filters” were introduced, a special case of which are linear median, order statistics, homomorphic, a-trimmed median, generalized mean, nonlinear mean and fuzzy nonlinear filters. New versions of polynomial filters, such as quadratic filters, were also studied. New morphological filters, which lead to various detection and recognition applications were designed. Finally, he conducted extensive research in the area of Adaptive filters. Adaptive Order Statistics filters, Adaptive LMS/RLS filters, Adaptive L-filters and Adaptive morphological filter algorithms were developed. These filters are extensively used in numerous biomedical applications, such as radiology, mammography, tomography, financial data processing and remote sensing applications among others.

In the nineties, he contributed to the area of color image processing and analysis, introducing a number of techniques for color image enhancement filtering and analysis. He introduced the so-called vector directional filter family, which operates along the direction of the color vectors. A new class of adaptive nonlinear filters was developed. Fuzzy membership functions based on different distance measures were adopted to determine the weights of new nonlinear, adaptive filters. The new filters encompass different classes of existing nonlinear filters as special cases. For the first time, the color image was treated as a vector field and edge information carried directly by the color vectors was exploited using vector order statistics.

Multimedia signal processing. In 1999 he became the Inaugural Chair of the Bell Canada Multimedia Systems Laboratory at the University of Toronto. Since then he contributed to the area of multimedia data mining and information retrieval by addressing two key technical challenges: a) the problem of similarity determination within the visual data domain, b) interactive learning of user intentions and automatic adjustment of system parameters for improved retrieval accuracy. He has developed still image and video retrieval systems that utilize color content queries. The system implements a new vector-based approach to image retrieval using an angular-based similarity measure. The developed scheme addresses the drawbacks of the histogram techniques, it is flexible, and outperforms established retrieval systems. He has also developed an interactive learning algorithm for resolving ambiguities arising due to the mismatch between machine-representation of images and human context-dependent interpretation of visual content. His proposed solution exploited feedback from the users during the retrieval sessions to adapt their query intentions and improve the accuracy of the retrieved results.

Biometric research. For thousands of years, humans have used visually perceived body characteristics such as face and gait to recognize each other. This remarkable ability of human visual system has led Professor Venetsanopoulos to build automated systems to recognize individuals from digitally captured facial images and gait sequences. Face and gait recognition belong to the field of biometrics, a very active area of research in the computer vision and pattern recognition society, mainly motivated by government and security-related applications. Face and gait are two typical physiological and behavioral biometrics respectively. Venetsanopoulos has contributed to both areas and his research was cited extensively. There are two general approaches: the model-based approach and the appearance-based approach. Appearance-based face recognition approach processes a 2-D facial image as 2-D holistic patterns. The whole face region is the raw input to a recognition system and each face image is commonly represented by a high-dimensional vector consisting of the pixel intensity values in the image. Thus, face recognition is transformed to a multivariate statistical pattern recognition problem. Although the embedding is high-dimensional, the natural constraints of the face data indicate that the face vectors lie in a lower-dimensional subspace (manifold). The popular subspace learning is such a method to identify, represent, and parameterize this subspace with some optimality criteria. Similar to appearance-based face recognition, appearance-based gait recognition approach considers gait as a holistic pattern and uses a full-body representation of a human subject as silhouettes or contours. Gait video sequences are naturally three-dimensional objects, formally named tensor objects, and they are very difficult to deal with using traditional vector-based learning algorithms. In order to deal with these tensor objects effectively, Venetsanopoulos and his research team developed a framework of multilinear subspace learning, so that computation and memory demands are reduced, natural structure and correlation in the original data are preserved, and more compact and useful features can be obtained. Model-based gait recognition approach considers a human subject as an articulated object, represented by various body poses. Professor Venetsanopoulos proposed a full-body layered deformable model (LDM) inspired by the manually labeled body-part-level silhouettes. The LDM has a layered structure to model self-occlusion between body parts and it is deformable, so simple limb deformation is taken into consideration. In addition, it also models shoulder swing. The LDM parameters can be recovered from automatically extracted silhouettes and then used for recognition.