Ioannis Katsavounidis

Implementation of the AVS video decoder on a heterogeneous dual-core SIMD processor

Multi-core Application Specific Instruction Processors (ASIPs) are increasingly used in multimedia applications due to their high performance and programmability. Nonetheless, their efficient use requires extensive modifications to the initial code in order to exploit the features of the underlying architecture. In this paper, through the example of implementing Advance Video Coding (AVS) to a heterogeneous dual-core SIMD processor, we present a guide to developers who wish to perform task-level decomposition of any video decoder in a multi-core SIMD system. Through the process of mapping AVS video decoder to a dual-core SIMD processor we aim to explore the different forms of parallelism inherent in a video application and exploit to speed-up AVS decoding in order to achieve real time functionality. Simulation results showed that the extraction of parallelism at all levels of granularity, especially at the higher levels, can give a total speed-up of more than 195× compared to a software x86-based implementation, which enables realtime, 25fps decoding of D1 video.

A Multiscale Error Diffusion Technique for Digital Multitoning

Multitoning is the representation of digital pictures using a given set of available color intensities, which are also known as tones or quantization levels. It can be viewed as the generalization of halftoning, where only two such quantization levels are available. Its main application is for printing and, similar to halftoning, can be applied to both colored and grayscale images. In this paper, we present a method to produce multitones based on the multiscale error diffusion technique. Key characteristics of this technique are: 1) the use of an image quadtree; 2) the quantization order of the pixels being determined through “maximum intensity guidance” on the image quadtree; and 3) noncausal error diffusion. Special care has been given to the problem of banding, which is one of the inherent limitations in error diffusion when applied to multitoning. Banding is evident in areas of the image with values close to one of the available quantization levels; our approach is to apply a preprocessing step to alleviate part of the problem. Our results are evaluated both in terms of visual appearance and using a set of standard metrics, with the latter demonstrating the blue-noise characteristics and very low anisotropy of the proposed method.

A high performance and low power hardware architecture for the transform & quantization stages in H.264

In this work, we present a hardware architecture prototype for the various types of transforms and the accompanying quantization, supported in H.264 baseline profile video encoding standard. The proposed architecture achieves high performance and can satisfy Quad Full High Definition (QFHD) (3840·2160@150Hz) coding. The transforms are implemented using only add and shift operations, which reduces the computation overhead. A modification in the quantization equations representation is suggested to remove the absolute value and resign operation stages overhead. Additionally, a post-scale Hadamard transform computation is presented. The architecture can achieve a reduction of about 20% in power consumption, compared to existing implementations.

A Novel Low-Power Motion Estimation Design for H.264

The H.264 video coding standard can achieve considerably higher coding efficiency than previous video coding standards. The keys to this high coding efficiency are the two prediction modes (Intra & Inter) provided by H.264 which adopt many new features such as variable block size searching, motion vector prediction etc. However, these result in a considerably higher encoder complexity that adversely affects speed and power, which are both significant for the mobile multimedia applications targeted by the standard. Therefore, it is of high importance to design architectures that minimize the speed and power overhead of the prediction modes. In this paper we present a new algorithm, and the architecture that implements it, that can replace the standard sum of absolute differences (SAD) approach in the two main prediction modes, supports the variable block size motion estimation (VBSME) as it is defined in the standard and provide a power efficient hardware implementation without perceivable degradation in coding efficiency or video quality.

Power reduction in an H.264 encoder through algorithmic and logic transformations

The H.264 video coding standard can achieve considerably higher coding efficiency than previous video coding standards. The keys to this high coding efficiency are the two prediction modes (intra & inter) provided by H.264. Unfortunately, these result in a considerably higher encoder complexity that adversely affects speed and power, which are both significant for the mobile multimedia applications targeted by the standard. Therefore, it is of high importance to design architectures that minimize the speed and power overhead of the prediction modes. In this paper we present a new algorithm, and the logic transformations that enable it, that can replace the standard sum of absolute differences (SAD) approach in the two main prediction modes, and provide a power efficient hardware implementation without perceivable degradation in coding efficiency or video quality

Low-Complexity Hand Gesture Recognition System for Continuous Streams of Digits and Letters

In this paper, we propose a complete gesture recognition framework based on maximum cosine similarity and fast nearest neighbor (NN) techniques, which offers high-recognition accuracy and great computational advantages for three fundamental problems of gesture recognition: 1) isolated recognition; 2) gesture verification; and 3) gesture spotting on continuous data streams. To support our arguments, we provide a thorough evaluation on three large publicly available databases, examining various scenarios, such as noisy environments, limited number of training examples, and time delay in systems response. Our experimental results suggest that this simple NN-based approach is quite accurate for trajectory classification of digits and letters and could become a promising approach for implementations on low-power embedded systems.

Finger detection and hand posture recognition based on depth information

In this work, we propose a novel framework for automatic finger detection and hand posture recognition, based mainly on depth information. Our method locates apex-shaped structures in a hand contour and deals efficiently with the challenging problem of partially merged fingers. Hand posture recognition is achieved using Fourier Descriptors of the contour, while global information about the fingers helps reducing the size of the search space. Our experiments on a dataset obtained from a Kinect device confirm the high recognition accuracy of our approach.

Sparse representations for hand gesture recognition

Dynamic recognition of gestures from video sequences is a challenging task due to the high variability in the characteristics of each gesture with respect to different individuals. In this work, we propose a novel representation of gestures as linear combinations of the elements of an overcomplete dictionary, based on the emerging theory of sparse representations. We evaluate our approach on a publicly available gesture dataset of Palm Grafti Digits and compare it with other state-of-the-art methods, such as Hidden Markov Models, Dynamic Time Warping and the recently proposed distance metric termed Move-Split-Merge. Our experimental results suggest that the proposed recognition scheme offers high recognition accuracy in isolated gesture recognition and a satisfying robustness to noisy data, thus indicating that sparse representations can be successfully applied in the field of gesture recognition.

Topology control with coverage and lifetime optimization of wireless sensor networks with unequal energy distribution

Abstract In this work, we use an unequal energy distribution algorithm, with which we succeed to extend network lifetime. In the sequel, we, we provide the best effort coverage by arranging the nodes using topology control principles. We formulate an optimization problem and define an objective function that incorporates both network coverage and lifetime that should be both maximized. Such a complex problem can be solved in polynomial time through Simulated Annealing (SA). An updated topology is evaluated in each convergence point and a near optimal node placement together with near optimal unequal energy allocation charge scheme is achieved. Results reveal that the proposed near optimal topology leads to greater coverage and lifetime as compared to several random deployments. We complete this work by providing a reliability study of the results, which is carried out by analysing the survival function of the derived statistical propertied of the proposed objective function.

Initialization of dynamic time warping using tree-based fast Nearest Neighbor

Initializing LBKeogh Dynamic Time Warping search using the Euclidean Distance Nearest Neighbor.Employing a fast Nearest Neighbor algorithm (fastNN) to increase computational efficiency.Successful application on five gesture datasets.Requiring about 20% less search time than existing DTW implementations without any drop in recognition accuracy.Exploring system parameters: number of training examples, data nature, etc. An efficient way to perform Dynamic Time Warping (DTW) search is by using the LBKeogh lower bound, which can eliminate a large number of candidate vectors out of the search process. Although effective, LBKeogh begins the DTW search using the first candidate vector, which is typically arbitrarily chosen. In this work, we propose initializing the LBKeogh-based DTW search using the Euclidean Distance Nearest Neighbor, derived by a fast tree-based Nearest Neighbor technique. Our experimental results suggest that, on one hand, this simple NN-based approach is quite accurate for trajectory classification of digit and letter gesturing and can initialize the DTW search very efficiently, thus requiring about 20% less search time than existing DTW implementations without any drop in recognition accuracy.