Nadeem A. Khan

Selection of variable block sizes in H.264

The paper aims at selecting an efficient variable block size mode in H.264 video coding standard for better compression performance. This standard allows video frames to be partitioned into variable block sizes such that blocks containing highly detailed motion are represented using small block sizes and the rest using large block sizes. New techniques for intelligent selection of the variable block sizes have been developed to reduce the computational complexity without sacrificing the quality of the coder. The proposed schemes are based on the motion vector cost and previous frame information. An improvement in the encoding time with negligible impact on the subjective and the quantitative performance has been achieved. A comparison of the proposed techniques for various test sequences is also provided.

A variable block size motion estimation algorithm for real-time H.264 video encoding

This paper presents an efficient variable block size motion estimation algorithm for use in real-time H.264 video encoder implementation. In this recursive motion estimation algorithm, results of variable block size modes and motion vectors previously obtained for neighboring macroblocks are used in determining the best mode and motion vectors for encoding the current macroblock. Considering only a limited number of well chosen candidates helps reduce the computational complexity drastically. An additional fine search stage to refine the initially selected motion vector enhances the motion estimator accuracy and SNR performance to a value close to that of full search algorithm. The proposed methods result in over 80% reduction in the encoding time over full search reference implementation and around 55% improvement in the encoding time over the fast motion estimation algorithm (FME) of the reference implementation. The average SNR and compression performance do not show significant difference from the reference implementation. Results based on a number of video sequences are presented to demonstrate the advantage of using the proposed motion estimation technique.

Developments in Distributed Video Coding

Distributed Video Coding (DVC) is a new coding paradigm based on two major information theory results namely the Slepian-Wolf (1973) and Wyner-Ziv(1976) theorems. Slepian-wolf coding is the lossless source coding and Wyner-Ziv coding is the lossy coding with receiver side information. Distributed Video Coding allows the implementation of lightweight encoders and complex decoders. This is the dual to conventional video coding such as MPEG or AVC/H.264 where the encoder has to carry most of the computational burden. There are various practical solutions proposed to address the two areas where DVC is applicable. These areas are low complexity video coding and robust video coding. These solutions are compared based on techniques used for compression, rate control, decoding and motion estimation.

Pain Intensity Evaluation through Facial Action Units

In this work we present a system that enables automatic estimation of Pain from image sequences with frontal views of faces. The system uses facial characteristic points to characterize different Action Units (AU) of pain and is able to operate in cluttered and dynamic scenes. Geometric features are computed using 22 facial characteristic points. We use k-NN classifier for classifying AU. Only action units relevant to pain are classified. Validation studies are done on UNBC McMaster Shoulder Pain Archive Database [8]. We also classify action unit intensities for evaluating pain intensity on a 16 point scale. Our system is simpler in design compared to the already reported works in literature. Our system reports AU intensities on a standard scale and also reports pain intensity to assess pain. We have achieved more than 84% accuracy for AU intensity levels and 87.4% area under ROC curve for pain assessment as compared to 84% of state-of-the-art scheme.

Performance improvement in motion estimation of Dirac wavelet based video codec

This paper presents an efficient motion estimation algorithm for enhancing the performance of Dirac wavelet based video encoder. The proposed scheme is based on a 3D Recursive Search algorithm that uses the previously calculated motion vectors from the neighboring blocks to calculate the motion vector for the current block. The computational complexity is reduced drastically with an improvement in the encoding time. The results show that the proposed algorithm gives approximately 40% to 60% reduction in terms of number of SAD calculations compared with the reference codec implementation. The average PSNR performance and compression efficiency remains close to reference codec implementation. Results based on a number of video sequences are presented to demonstrate clearly the benefits of the proposed improvements in Diracs motion estimation module.

Multilevel optimization of speech coding algorithms for modern DSP architectures

Real-time implementation of speech coding algorithm requires the DSP code to be highly optimized and the underlying hardware to be fast. This paper presents the results obtained from multi-level optimization of ITU-T G.729A low complexity speech codec on an embedded DSP platform. The implementation platform used in this work is based on analog devices Blackfin BF533 fixed-point media processor. Several high-level and low-level optimization techniques have been concurrently employed in this work to improve the run-time performance of the codec. Improvement in performance has been evaluated through the reduction in MIPS count for each optimization step. Results show that over twenty full-duplex channels of G.729A can be supported on this DSP platform in real-time.

Efficient scheme for motion estimation and block size mode selection in H.264

The paper presents efficient algorithms for enhancing the computational performance of an H.264 video encoder. We employ previously calculated motion vectors and variable block size modes from neighboring blocks to select the best modes and motion vectors for the current macroblock. The scheme results in 40% to 70% improvement in the encoding time over a reference implementation without degradation in the subjective quality. The average SNR and the compression performance do not show variation from the reference codec implementation. Results based on a number of video sequences are presented that clearly demonstrate the benefit of the proposed motion estimation and block size mode selection algorithm. The proposed scheme has been shown to provide the least computational load with improvement in the encoding time. The reduction in memory requirements and calculations makes this scheme useful for VLSI implementation and realtime applications.

Computer Assisted Analysis System of Electroencephalogram for Diagnosing Epilepsy

Automation of Electroencephalogram (EEG) analysis can significantly help the neurologist during the diagnosis of epilepsy. During last few years lot of work has been done in the field of computer assisted analysis to detect an epileptic activity in an EEG. Still there is a significant amount of need to make these computer assisted EEG analysis systems more convenient and informative for a neurologist. After briefly discussing some of the existing work we have suggested an approach which can make these systems more helpful, detailed and precise for the neurologist. In our proposed approach we have handled each epoch of each channel for each type of epileptic pattern exclusive to each other. In our approach feature extraction starts with an application of multilevel Discrete Wavelet Transform (DWT) on each 1 sec non-overlapping epochs. Then we apply Principal Component Analysis (PCA) to reduce the effect of redundant and noisy data. Afterwards we apply Support Vector Machine (SVM) to classify these epochs as Epileptic or not. In our system a user can mark any mistakes he encounters. The concept behind the inclusion of the retraining is that, if there is more than one example with same attributes but different labels, the classifier is going to get trained to the one with most population. These corrective marking will be saved as examples. On retraining the classifier will improve its classification, hence it will tries to adapt the user. In the end we have discussed the results we have acquired till now. Due to limitation in the available data we are only able to report the classification performance for generalised absence seizure. The reported accuracy is resulted on very versatile dataset of 21 patients from Punjab Institute of Mental Health (PIMH) and 21 patients from Children Hospital Boston (CHB) which have different number of channel and sampling frequency. This usage of the data proves the robustness of our algorithm.

A spatio-temporal recursive search based prediction scheme for efficient multi-frame and bidirectional motion estimation

In this paper a new multi-frame/bidirectional motion estimation algorithm based on the concept of recursive-search in spatial and temporal space is proposed that effectively minimizes the number of candidate motion vectors in the prediction set. The algorithm is extremely computationally light compared to full search and outperforms existing fast approaches. The algorithm can operate effectively on reference frames that may be chosen to be consecutive or temporally separated. The average Peak Signal-to-Noise Ratio (PSNR) performance and compression efficiency is virtually not compromised and very close to full search. Results based on a number of video sequences and on different GOP (Group of Pictures) structures are presented to clearly demonstrate the benefits of the proposed motion estimation technique.

A discriminative spectral-temporal feature set for motor imagery classification

This paper presents a novel technique for motor imagery event classification. Extraction of discriminative feature is a key to accurate classification. To realize this objective we have explored the use of nonnegative matrix factorization (NNMF) for sparse representation of our input signal and determining the discriminative basis vector. We extract both spectral as well as temporal features from this representation to construct our features set. Band power has been shown to be a powerful discriminative feature of the spectral domain for motor imagery classes. Time Domain Parameter (TDP) taken as a temporal feature measures power of EEG using first few derivatives. Our approach is novel in proposing a fusion of both these features. We have used Hierarchical Alternating Least Square (HALS) as a convergence solution to minimize error function of NNMF as it converges more rapidly as compared to other methods. The proposed feature set has been tested using LDA and SVM classifiers technique for classification of 4-class motor imagery signals. We have compared our approach with others presented in literature using the Dataset 2a of BCI competition IV and has shown that our approach achieves the highest reported mean kappa value of 0.62 with the SVM classifier.