Muhammad Shafi

Eyes extraction from facial images using edge density

This paper proposes a novel method for eyes extraction in facial images using edge density information. The method is based on the observation that irrespective of skin colour, colour variations occur the most in the eye region. In the proposed method, edges are detected in the input facial image. Morphological dilation is applied twice and the holes are filled in the connected regions. This makes the high density edges regions appear as blobs. Certain shapes and geometrical rules are applied to these blobs to extract the eyes. The method was tested using images from the PICS facial images database. The accuracies of the initial blobs extraction and the final eyes extraction were 95% and 72% respectively.

LS estimator: Performance analysis for block-type and comb-type channel estimation in OFDM system

Channel Estimation in Orthogonal Frequency Division Multiplexing (OFDM) is imperative for cancellation of the impairments introduced by the fading channel. In this paper, the performance of Least Square (LS) estimator is studied for comb-type and block-type pilot insertion schemes over fast and slow fading Rayleigh channels respectively. The channel order increases proportionally with the distortions introduced from the environment. The effect of increase in channel order on the Bit Error Rate (BER) performance of LS estimator for pilot assisted channel estimation techniques is evaluated. Stability of LS estimator in terms of BER performance for block-type pilot arrangement is revealed by the Simulation results. However, increasing the channel distortion by increasing the channel order causes performance degradation of LS estimator for comb-type pilot insertion scheme.

DeepSIC: a deep model

Satellite Image Classification is the problem of classifying satellite images into their corresponding classes. Using supervised deep learning, we provide an efficient classification mechanism of earth observation imagery captured by satellite. Promising classification results were obtained through Convolution Neural Network. Two sets of experiments were carried out. In the first set, the traditional pre-deep learning era approach was followed: features were extracted first which were then given as input to Support Vector Machines (SVM). In the second set of experiments, the images were directly provided as input to Convolution Neural Networks and the output features were then used as input to the SVM. The proposed schemes were tested on satellite images datasets AID, UC-merced and WHU-RS, having varied classes and heterogeneous image dimensions. We show that our model, with Convolution Neural Network, achieves 92% accuracy on the AID dataset and outperforms the previously reported best accuracy of 89.64% (Xia et al. in CoRR, arXiv:1608.05167, 2016).

Job search website for illiterate users of Pakistan

Abstract Illiteracy is a major hurdle in socio-economic development of Pakistan. Research in the Information and Communication Technology (ICT) can help in tackling this issue by developing solutions tailored for illiterate population. The work presented within this paper proposes a job search website for illiterate population of KPK province of Pakistan. The designed website does not require any human assistance while searching a job. Moreover, the website uses culturally relevant graphical and audio content. The usability of the website is tested according to the ISO-9241-11 specification. Furthermore, we have studied the effect of the participants’ characteristics, i.e., their age, computer and mobile usage experience on the usability of the website. The results show that the participants found the website usable. Specifically, they could search suitable jobs with a minimal number of clicks and in less time as compared to the usual task completion rates reported throughout the literature. None of the participants’ characteristics were found to be affecting the usability of the website. These results suggest a promising potential of ICT solutions for providing services to illiterate population of Pakistan.

Face pose estimation from eyes and mouth

Face pose estimation plays an important role in human computer interaction, automatic human behaviour analysis, gaze estimation, virtual reality, pose independent face recognition, etc. Accuracy and speed are the most desirable features of a face pose estimation system. In this paper, a face pose estimation scheme based on the centres of the eyes and mouth is proposed. The proposed method is simple and is, therefore, very effective in terms of computation because it uses only three points, i.e., eyes and mouth centres. The use of only three points increases the pose estimation range and makes the method suitable for real time applications. Tests using the Pointing 04 database show that the proposed scheme is robust and fast.

PERFORMANCE ENHANCEMENT OF FACE RECOGNITION SYSTEM USING PRINCIPAL COMPONENT ANALYSIS MERGED WITH DISCRETE WAVELET TRANSFORMS

Performance of face recognition system can be enhanced by proposed technique titled as PCA merged with Discrete Wavelet Transform (DWT) instead of using the conventional PCA technique. In this technique to reduce the computational complexity of traditional PCA the size of the image is first reduced by taking the DWT of it. After applying the DWT the facial features of the image are extracted by calculating the Eigenface of the image with size already reduced by taking DWT. As a result of this process the size of database will reduce to one-fourth of the conventional PCA in which the facial features are extracted directly from the image by calculating the Eigenface. The size of the train database is reduced with the proposed technique which reduces the processing time of the face recognition without losing the accuracy. Performance of face recognition system is enhanced in terms of low processing time as shown by comparing the experimental results of conventional PCA and the proposed technique in this paper

Wi-Fi Range Extension with Cost Effective Simple Design of Multi-element Antenna

This paper presents a simple and low cost design of a high directional antenna for the frequency band of 2.4 to 2.4835 GHz. This range of frequency is in common use for Wi-Fi standard. Limited coverage and access has always been a bottleneck in setting up a wireless communication system. To overcome this issue, besides increasing the transmitter power, a suitable solution is to design an efficient and high range antenna. This paper presents the design of a Multi-Element Bi-Quad directional antenna which has a gain of 15dBi while retaining a minimum cost design. Copper has been used in the proposed design. The simulation was performed using Numerical Electromagnetics Code -2 (NEC-2) and the antenna was fabricated and tested using Nets tumbler. The various parameters e.g. Gain, Beam width, VSWR values for the fabricated antenna were found to be closely matched to the simulation results.