Kambiz Nayebi

Differential of the mutual information

In this letter, we compute the variation of the mutual information, resulting from a small variation in its argument. Although the result can be applied in many problems, we consider only one example: the result is used for deriving a new method for blind source separation in linear mixtures. The experimental results emphasize the performance of the resulting algorithm.

Separating Convolutive Mixtures by Mutual Information Minimization

Blind Source Separation (BSS) is a basic problem in signal processing. In this paper, we present a new method for separating convolutive mixtures based on the minimization of the output mutual information. We also introduce the concept of joint score function, and derive its relationship with marginal score function and independence. The new approach for minimizing the mutual information is very efficient, although limited by multivariate distribution estimations.

ECG beat classification based on a cross-distance analysis

This paper presents a multi-stage algorithm for QRS complex classification into normal and abnormal categories using an unsupervised sequential beat clustering and a cross-distance analysis algorithm. After the sequential beat clustering, a search algorithm based on relative similarity of created classes is used to detect the main normal class. Then other classes are labeled based on a distance measurement from the main normal class. Evaluated results on the MIT-BIH ECG database exhibits an error rate less than 1% for normal and abnormal discrimination and 0.2% for clustering of 15 types of arrhythmia existing on the MIT-BIH database.

A new segmentation technique for multi font Farsi/Arabic texts

Segmentation is a very important stage of Farsi/Arabic character recognition systems. A new segmentation algorithm - for multi font Farsi/Arabic texts - based on the conditional labeling of the up contour and down contour is presented. A pre-processing technique is used to adjust the local base line for each subword. This algorithm uses an adaptive base line for each subword to improve the segmentation results. This segmentation algorithm, in addition to up and down contours, takes advantage of their curvatures also. The algorithm was tested on a data set of printed Farsi texts, containing 22236 characters, in 18 different fonts. 97% of characters were correctly segmented.

A new method to improve multi font farsi/arabic character segmentation results: using extra classes of some character combinations

A new segmentation algorithm for multifont Farsi/Arabic texts based on conditional labeling of up and down contours was presented in [1]. A preprocessing technique was used to adjust the local base line for each subword. Adaptive base line, up and down contours and their curvatures were used to improve the segmentation results. The algorithm segments 97% of 22236 characters in 18 fonts correctly. However, finding the best way to receive high performance in the multifont case is challengeable. Different characteristics of each font are the reason. Here we propose an idea to consider some extra classes in the recognition stage. The extra classes will be some parts of characters or the combination of 2 or more characters causing most of errors in segmentation stage. These extra classes will be determined statistically. We have used a learn document of 4820 characters for 4 fonts. Segmentation result improves from 96.7% to 99.64%.

Blind Source Separation by Adaptive Estimation of Score Function Difference

In this paper, an adaptive algorithm for blind source separation in linear instantaneous mixtures is proposed, and it is shown to be the optimum version of the EASI algorithm. The algorithm is based on minimization of mutual information of outputs. This minimization is done using adaptive estimation of a recently proposed non-parametric “gradient” for mutual information.

A minimization-projection (MP) approach for blind separating convolutive mixtures

A new algorithm for blind source separation in convolutive mixtures, based on minimizing the mutual information of the outputs, is proposed. This minimization is done using a recently proposed minimization-projection (MP) approach for minimizing mutual information in a parametric model. Since the minimization step of the MP approach is proved to have no local minimum, it is expected that this new algorithm has good convergence behaviour.

Multirate structures for arbitrary rate error control coding

We present the most general form for error control coding using finite field multirate filters. This method shows how different types of codes can easily be generated by multirate filters and filter banks. In all previous works, codes and syndromes were generated using prefilters. We present simple multirate structures for encoding and generating syndromes. We show that all kinds of arbitrary rate K/L, circulant linear codes can be generated by these structures. Then we claim that a similar simple structure exists for syndrome generation in all presented cases.

Real-time implementation and optimization of ITU's G.728 on TMS320C64X DSP

The objective of this paper is to illustrate the details of optimization and real-time implementation of ITUs G.728 on C64x DSPs. First using pseudo codes provided by CCITTs published documents, we implemented the algorithm in C language. This implementation was performed for G and H (12.8 and 9.6 kbit/s) annexes. Next we optimized the written codes for implementation on DSP. At first stage, using different techniques based on DSPs hardware characteristics, we rewrote and changed the most time-consuming parts of our codes in order to reduce their execution time. At second stage, we balanced the computational load of G.728 coder algorithm by splitting the Durbins recursion for synthesis filter between different input speech vectors. In each stage, we verified the correctness of our implementation by testing our codes against testing vectors provided by ITU. Applying the above mentioned methods enabled us to optimize the C codes into 22.7 MIPS in worst case. At the end we also implemented the optimized codes in real-time on a DSK6416

A new efficient PNLMS based algorithm for adaptive line echo cancellation

Echo cancellers with long impulse responses are usually used in networks with long echoes. Since long adaptive filters with sparse target filters usually converge very slowly, adaptation algorithms with fast convergence rate are needed. PNLMS is one of the algorithms that is designed for fast convergence on sparse impulse responses. But there are some disadvantages to this algorithm. The computational complexity of this algorithm is prohibitive, especially for long echo tails. Another disadvantage is that its convergence rate slows down significantly after the adaptation of large taps. In this paper we propose new algorithms to solve both these problems. In another part of this paper, we present a modified PNLMS algorithm with better computational and convergence characteristics. Proposed algorithms show a better performance than PNLMS both in convergence rate and computational complexity.