Page 1

Bushra Ferdousi Portfolio Curriculum Vitae Bushra Ferdousi

144 Halsey Drive, Apt 03, West Lafayette, Indiana-47906, USA E-mail: Cell: +1(765)-409-8607 Website: Education Purdue University West Lafayette,IN, USA Masters in Computer Graphics and Technology, Starting from Fall 2018 • Advisor: Dr. David M Whittinghill United International University, Dhaka MSc.Engg., Computer Science and Engineering, November 2015 • CGPA: 3.833 out of 4.00 • Thesis Title: Cough Detection Using Speech Analysis. • Advisor: Dr. Mohammad Nurul Huda Ahsanullah University of Science and Technology, Dhaka BSc.Engg., Computer Science and Engineering, February 2014 • CGPA: 3.647 out of 4.00 B.A.F. Shaheen School and College, Dhaka Higher Secondary Certificate (H.S.C.), 2009 • GPA: 4.63 out of 5.00 Secondary School Certificate (S.S.C.), 2007 • GPA: 5.00 out of 5.00 Research Interests Professional Experience

Analytics and Modeling, Computer Graphics, Human-Computer Interaction, Machine learning, Information retrieval, Data Mining and Big Data Analytics

Junior Software Engineer NNS Solution October 2013 - March 2014 Cartoonist Unmad Satire Magazine February 2013 - present


Bushra Ferdousi Portfolio Curriculum Vitae Research Experience

Working with Prof. Dr. Mohammad Nurul Huda June 2014 - present • Implementing and analyzing machine learning algorithms for Cough Detection using speech signal.

Publications Conference • Bushra Ferdousi, S M Ferdous, Khondaker Abdullah Al Mamun and Mohammad Nurul Huda, Cough Detection Using Speech Analysis, 2015 18th International Conference on Computer and Information Technology (ICCIT), Dhaka, 2015, pp. 60-64. doi: 10.1109/ICCITechn.2015.7488043 Certifications • The Data Scientists Toolbox, Johns Hopkins University, Coursera, August 2017 (earned 100%) • R Programming , Johns Hopkins University, Coursera, October 2017 (earned 100%) Technical Skills

Programming: R, C, C++, Java, C#, Prolog, Assembly Language PL/SQL, OpenGL. Web Programming: HTML, PHP, CSS, WPF Application. Modelling: UML, E-R Diagram, DFD. Database Management System: Microsoft SQL Server 2008, Oracle , MySQL. Simulation: MATLAB. Statistical Tool: MATLAB. Version Controlling Tool: GitHub Game Engine: Unreal Engine Animation Software: Blender, Photoshop Applications: TEX, LATEX, BibTEX, Microsoft Office, and other common applications for Windows.

Professional Software Projects

EBL(Eastern Bank Limited) Dispatch Management System: October2013 This was an individual project. The aim of the project was to automate the dispatch management unit of Eastern Bank Ltd., a leading private bank in Bangladesh. The aim of the project was to create an automated dispatch notification system using e-mail. When an item entered or realised in dispatch unit the system generates automated mail to inform the authority with all details of that item. It includes different sub system like keeping track on


Bushra Ferdousi Portfolio Curriculum Vitae particular item, including vendor information, maintaining admin and staff access ability on software and so on. The project was implemented by C# .NET Framework.

Undergrad Software Projects

Result Processing System of Ahsanullah University of Science and Technology: October2012 This was a team project. The aim of the project was to automate the result processing system of my university. The result processing system includes different sub system such as the submission of the results (from the instructors), Gradesheet preparation, Annoncement of the results (to the students) and post processing (such as cumulative results preparation etc.). The project was implemented by C# using WPF. Inventory Management System: April 2012 This was also a database software implemented in .NET platform using C#. The software was about managing inventory where the user can add, delete or update certain products. The system was some more capabilities other than storage such as it generated warning to the user if some products count become less than a threshold.

Undergrad Hardware Projects

Arithmetic Logic Unit A system designed to perform arithmetic and logic operations. Modified Booth algorithm A system designed for multiplication of two n bits signed binary number. 4-bit Digital Computer Designed and developed a digital computer capable of executing 28 instructions from the Intel 80x86 processor family. Its main features include shared bus, micro programmed control unit and separate data and instruction RAM.

References Dr. David M Whittinghill Associate Professor, Department of Computer Graphics Technology Purdue University, West Lafayette, Indiana 47907-2021, USA E-mail: Dr. Mohammad Nurul Huda Professor, Department of Computer Science and Engineering United International University, Dhaka-1209, Bangladesh E-mail: Dr. Mohammad Shafiul Alam Associate Professor, Department of Computer Science and Engineering Ahsanullah University of Science and Technology, Dhaka-1208, Bangladesh E-mail:,


Masters Thesis Paper



Cough Detection using Speech Analysis Bushra Ferdousi

S M Ferdous

United International University Ahsanullah University of Science and Technology Dhaka, Bangladesh Dhaka, Bangladesh Email: Email: Abstract—Common cold is a common disease now-a-days. Due to common cold patient faces cough, sore throat, sneezing and runny nose problem. Most of the time patients’ speech sounds different due to cough. In this paper, analyzing speech recording of cough and normal state of a person, we have derived two sets of representative features. These features are used for classifying normal and cough state of the patient. The classiďŹ cation algorithms we have used are Support Vector Machine, Bayesian ClassiďŹ er and Neural Network. On the generated real life dataset, we have applied the features and classiďŹ ers. We have listed the performance statistics of the exhaustive experiment. The performance measures reveal that the classiďŹ ers with the second feature set provide very good accuracy (greater than 70% for all the classiďŹ ers). Among the three classiďŹ ers Bayesian provides the best accuracy (86.31%).


Khondaker Abdullah-Al-Mamun and Mohammad Nurul Huda United International University Dhaka, Bangladesh

[17], authors applied Empirical Mode Decomposition (EMD) and Hidden Markov Model (HMM) to detect cough sound. Contrary to the previous works, our goal in this research to analyze normal speech of a cough affected (or cold affected) patient. The speech do not include explicit coughing. II.


A. Data Ten representative words are selected as input. We record the speech denoting the words from a person with cough. When the person get cured, the same speech in same environment is recorded from him. The speech as saved as wave ďŹ le format. We have used ten different bengali words as data. First we record those ten words from a person in cough state, then we collect those same words from same person in non cough state. These ten words are listed in Fig. 1.


A cold is a mild viral infection of upper respiratory tract such as nose, throat, sinuses and upper airways. It may be caused by many different viruses. The symptom of the common cold is cough, sore throat, sneezing and runny nose. A cough is a reex action to clear the airways of mucus and irritants such as dust or smoke. There are over 200 viruses known to cause cough. Chronic cough is a cough that persists over time. Chronic cough is not a disease by itself, but rather a symptom of an underlying condition. Chronic cough is a common problem and the reason for many doctor visits. Approximately 1 billion cold patient in the USA annually, Children get 6-10 yearly, adults 2-4 yearly. But surprisingly cough detection has not got enough attention in the literature.

Fig. 1.

Ten Bengali words with English pronunciation

Due to cough or cold Nasal Congestion happens commonly. Nasal Congestion is the blockage of the nasal passages usually due to membranes lining the nose becoming swollen from inamed blood vessels. It is also known as nasal blockage, nasal obstruction, blocked nose, stuffy nose, or plugged nose. It can interfere with hearing and speech. In this research we are interested to detect patient with cough from their speech. We are motivated by the fact that the speech of a cough affected person is signiďŹ cantly different from the speech of the person when he cures [8]. Our goal is to exploit this and develop a machine learning based algorithm to detect cough patient from their speech.

Silence is the presence of sounds of very low intensity. Silence in speech can give us wrong result on our work, for that reason we remove beginning and end silence from our speech using Audacity software.1 Figs. 2-3 illustrated the output after removing silence from one of the speech.

Surprisingly cough detection has not got enough attention in the literature. Most of the previous works are on detecting cough sound. For example, a tree classiďŹ cation based algorithms was proposed in [9]. In [7], the author presented validation of a cough analysis tool. They developed a software using speech analysis and machine learning that can assess the improvement of TB patients undergoing treatment. In

Fig. 2.

B. Removing Silence

Speech with beginning and end silence


ÂŞ*&&& 60


representative features. We further subdivided the features into to categories namely time domain features and frequency domain features.

Fig. 3.

1) Time domain features: Time domain features are calculated directly from the sampled frame value. The first 3 features are from time domain. They are listed below

Speech without beginning and end silence


1) The Zero-Crossing Rate (ZCR) is the rate of sign-changes along a signal, i.e., the rate at which the signal changes from positive to negative or back [4]. Zero crossing rates are used for Voice activity detection(VAD), i.e., finding whether a segment of speech is voiced or unvoiced. zcr is defined formally as:


Two different feature sets are used in our work. In this section we will describe elaborately those feature sets.

Algorithm 1 featureExtraction(cough data, noncough data) T −1 count = 0 1  zcr = Π(St St−1 < 0) (1) win = 100/1000 T − 1 t=1 shif t = 25/1000 for F eatures ← 1tosize(cough data) do where S is a signal of length T and the indicator function while length(cough data(i) ≥ shif t) do Π(A) is 1 if its argument A is true and 0 otherwise. F eatures ← extractF eature(F rame, f s, win, shif t) 2) The Energy [6], Es of an audio frame, S is defined as, cough data(i) ← shif t 1 count + + |St |2 (2) Es = F eatures(i, end) ← 1 T end while where T is the length of the frame. end for 3) The Energy entropy [1] of a given audio frame calculates for F eatures ← 1tosize(noncough data) do the entropy of the sub frame energies. By subdividing while length(noncough data(i) ≥ shif t) do each frame into a set of sub-frames, their respective F eatures ← extractF eature(F rame, f s, win, shif t) short-term energies are calculated and treating them as noncough data(i) ← shif t probabilities, we can calculate their entropy. The total count + + energy of a frame, S denoted by TE is F eatures(i, end) ← 0  end while Te = (S 2 ) end for We then sub-divided the frame into n sub-frames denoted by s1 , s2 , . . . sn . The normalized sub-frame energy is, The detailed feature extraction outline is presented in n 2 Algorithm 1. Given the cough and non cough sample of a s se = i=1 i person, the algorithm first sets an window size of 100 ms with Te shift of 25 ms. for each audio frame of 100 ms, the algorithm So the entropy of the sub-frame energies is, computes two sets of 35 features and labelled them as 1 (cough  sample) or 0 (non cough sample). E=− se log(se ) (3) A. First features set

2) Frequency domain features:

The first feature set captures the top 35 intensities in frequency domain. Given an audio frame sample data, the first feature set (henceforth labelled as Featureset1) is calculated as follows.

1) The Spectral centroid is a measure used in digital signal processing to characterise a spectrum. It indicates where the “center of mass” of the spectrum is. Perceptually, it has a robust connection with the impression of “brightness” of a sound [10]. It is calculated as the weighted mean of the frequencies present in the signal, determined using a Fourier transform, with their magnitudes as the weights[12]: N −1 n=0 f (n)x(n) (4) C=  N −1 n=0 x(n)

1) Firstly we apply Hamming Window [11] to generate windowed frame sample. 2) Secondly, Fast Fourier Transform (FFT) [3] is applied on the windowed samples to change the domain from time to frequency of the signal. We recorded the amplitudes from real and imaginary part of the output of FFT. 3) Finally we calculate the decibel (dB) values to get intensities. The top 35 intensities are recorded as features for the sample.

where x(n) represents the weighted frequency value, or magnitude, of bin number n, and f (n) represents the center frequency of that bin. Bin the range of values that is, divide the entire range of values into a series of small intervals and then count how many values fall into each interval.

B. Second features set Given an audio frame sample, the second feature set (henceforth labelled as Featureset2) calculates 35 61


The Spectral Spread is calculated as, N −1 f (n)(C − x(n))2 S = n=0 N −1 n=0 x(n)

b) Map the powers of the spectrum obtained above onto the mel scale, using hamming overlapping windows. c) Take the logs of the powers at each of the mel frequencies. d) Take the discrete cosine transform of the list of mel log powers, as if it were a signal. e) The MFCC are the amplitudes of the resulting spectrum. We have taken the first 13 coefficients. 6) If a signal is periodic with frequency f, the only frequencies composing the signal are integer multiples of f, i.e., f, 2f, 3f, 4f, etc. These frequencies are called harmonics. The first harmonic is f, the second harmonic is 2f, the third harmonic is 3f, and so forth. The first harmonic (i.e., f) is also given a special name, the fundamental frequency. 7) In [15]author introduce chroma vectors, that represents musical tonality. These features are shown to outperform other commonly used features in multiple conditions and corpora.


C and S constitute the 4th and 5th features. 2) The Spectral Entropy [16] is similar to Energy Entropy in time domain. The only difference is this time we are working on fourier transformed signal rather than the original signal. The total spectral energy of a frame, F denoted by TE is  Te = (F 2 ) We then sub-divided the frame into n sub-frames denoted by f1 , f2 , . . . fn . The normalized sub-frame spectral energy is, n f2 fe = i=1 i Te So the spectral entropy of the sub-frame energies is,  Ef = − fe log(fe ) (6)

Table I summarizes the 35 features used in Featureset2.

3) The Spectral Flux [1] is a measure of how quickly the power spectrum of a signal is changing, calculated by comparing the power spectrum for one frame against the power spectrum from the previous frame [5]. More precisely, it is usually calculated as the 2-norm (also known as the Euclidean distance) between the two normalised spectra. Let the FFT of current frame is fc and the FFT of previous frame fp , the the spectral flux is,  (7) F lux = (fc − fp )2 ;


4) The Spectral Rolloff Point is defined as the N th percentile of the power spectral distribution, where N is usually 90%. The rolloff point is the frequency below which the N % of the magnitude distribution is concentrated [14]. Mathematically, The the rollof, R is, R = {n : En = 0.9 ∗ Tf }


Feature No. 1 2 3 4 5 6 7 8 9 - 21 22 23 24 - 35




Description Zero Crossing Rate Energy Energy Entropy Spectral Centroid Spectral Spread Spectral Entropy Spectral Flux Spectral Rolloff Point 13 Mel-frequency cepstral coefficients Harmonic Ratio Fundamental Frequency Chroma Vector


The training algorithms are Support Vector Machine (SVM), Naive Bayesian Classifier (Bayesian) and Neural Network (Neural Net). Algorithm 2 presents the generic steps. Firstly we shuffle the dataset and choose the first 80% as training data and the rest are for testing. Then we call the trainTest procedure parameterized by the training and testing set.


where En is the nth spectral energy and Tf is the total spectral energy. This measure is useful in distinguishing voiced speech from unvoiced: unvoiced speech has a high proportion of energy contained in the high-frequency range of the spectrum, where most of the energy for voiced speech and music is contained in lower bands. 5) Mel-frequency cepstral coefficients (MFCC) are coefficients that collectively make up an MFC. They are derived from a type of cepstral representation of the audio clip (a nonlinear “spectrum-of-a-spectrum”). The difference between the cepstrum and the mel-frequency cepstrum is that in the MFC, the frequency bands are equally spaced on the mel scale, which approximates the human auditory system’s response more closely than the linearly-spaced frequency bands used in the normal cepstrum. This frequency warping can allow for better representation of sound. MFCCs are commonly derived as follows:[2][13] a) Take the Fourier transform of (a windowed excerpt of) a signal.

Algorithm 2 applyClassifier(F eatureset, Classlabel) shuf f led index ← indexes af ter shuf f ling shuf f led f eatureset ← F eatureset(shuf f led index) shuf f led classlabel ← Classlabel(shuf f led index) F eature tr ← 80% of shuf f led f eatureset Classlabel tr ← 80% of shuf f led classlable F eature tst ← 20% of shuf f led f eatureset Classlabel tst ← 20% of shuf f led classlable

Classif ier = trainT est(D trF eature, D trClass, D tstF eature, D tstClass)

Procedure 3 summarizes the steps of training and testing. At first using training set we train the dataset. Then testing set is used to collect the performance measures of training. V.


In this experiment we used a personal computer and the configuration of this computer are given below: 62


Algorithm 3 trainTest(D trF eature, D trClass, D tstF eature, D tstClass) It can be easily verified from the Table II that Featureset2 provides much better performance than bayesianClassif ier(D trF eature, D trClass) Featureset1. SVM with Featureset1 is slightly better svmClassif ier(D trF eature, D trClass) than the other two classifiers. On the other hand for nnClassif ier(D trF eature, D trClass) stat bayesian = test bayesian(D tstF eature, D tstClass) Featureset2, Bayesian classifier excels in accuracy and also in sensitivity. The Neural Network is less accurate than stat svm = test svm(D tstF eature, D tstClass) Bayesian but the specificity is better. The SVM reports the stat nn = test nn(D tstF eature, D tstClass) worst accuracy out of the three classifiers for Featureset2. •

Processor: Intet(R) @2.50GHz 2.50Ghz

Main Memory: 4 GB

L2-Cache: 3 MB

Operating System: Windows 8.1 64 bit



The single run shows that Featureset2 is much better in all of the performance criteria than Featureset1. It also revealed that Bayesian is better to use then the other classifiers. To investigate the performance further we randomly selects 20 sets of training and testing data. Each set contains 821 training samples with 206 testing samples. For both of the feature set we trained the three classifiers on training samples and measured the performance on the testing samples. Table III shows the average performance statistics of the three classifiers using both of the feature sets. The columns are same as Table II with the exception that now the columns represent average of 20 runs of the algorithms. Additionally, the last column reports the standard deviation of accuracy of 20 runs. The table basically draws the same conclusion as observed during the single run. Out of two feature set, the second one is much better than the first one. Numerically the average improvement in accuracy across the classifiers for the Featureset2 is more than 20%.


A. Performance Measures To compare the classifier algorithms with each others, we need some representative performance measures. The performance measures those are used in the experiments are explained below. 1) Accuracy is the percentage of samples correctly classified from the Testing data set. 2) True Positive count is the number of samples from the test set those are both labelled and predicted as 1 (cough sample). 3) False Positive count is the number of samples from the test set those are labelled as 0 (non cough sample) but predicted as 1 (cough sample). 4) False Negative count is the number of samples from the test set those are labelled as 1 (cough sample) but predicted as 0 (non cough sample). 5) True Negative count is the number of samples from the test set those are both labelled and predicted as 0 (non cough sample). 6) Sensitivity(also called the true positive rate) measures the proportion of positives which are correctly identified as such. Mathematically Sensitivity is CorrectlyClassif iedP ositiveSamples/T rueP ositiveSamples. 7) Specificity (also called the true negative rate) measures the proportion of negatives which are correctly identified as such. Mathematically Specificity is CorrectlyClassif iedN egativeSamples/T rueN egativeSamples. VI.

Bayesian classifier provides better performance than the other two classifiers for Featureset2. Numerically the average improvement in accuracy for Bayesian is approximately 14% more than SVM and 6% more than Neural Network. For SVM, the sensitivity is much lower than the other two classifiers but the specificity is the best amongst the three. Analyzing all these data, we can conclude that Bayesian is the best classifier for the dataset. VII.


In this paper we have developed two representative features for cough detection data using speech analysis. We have also applied three classifiers algorithms on the feature sets. From the analysis we can conclude that the Featureset2 is much more efficient than the other. Among the three classifiers, Bayesian is better in accuracy when using all the features. On the other hand SVM provides comparable accuracy with much fewer features. In this paper, we have applied hamming window for cough detection problem. A good future research could be applying other windows such as hanning window, rectangular window, triangular window. We have also limited our study on SVM, Neural Network and Bayesian. In future it might be interesting to develop other algorithm Decision trees, Nearest neighbors (kNN), Markov chain etc. to solve this problem.


For the classification task, we have developed two sets of features namely Featureset1 and Featureset2. Both of these sets contain 35 features. We randomly chose 80% of the data as training set and rest of the data as testing set and run the classifier algorithms. The number of training sample was 821 and the number of test sample was 206 with 110 samples labeled as 1 (cough sample) and the rest 96 are labelled with 0 (non-cough sample). The performance measure of single run is listed in Table II. The third to ninth column of each of the tables represents the accuracy, True Positive sample counts, False Positive sample counts, False Negative sample counts, True Negative sample counts, Sensitivity and Specificity respectively.

R EFERENCES [1] Introduction to audio analysis. In T. Giannakopoulos and A. Pikrakis, editors, Introduction to Audio Analysis. Academic Press, Oxford, 2014. [2] K. Aizawa, Y. Nakamura, and S. Satoh. Advances in Multimedia Information Processing - PCM 2004: 5th Pacific Rim Conference on Multimedia, Tokyo, Japan, 63




Featureset1 Featureset2




[5] [6]


[8] [9]


TP count

FP count

FN count

TN count



57.28 55.83 51.94 78.64 89.32 80.10

52 12 57 54 84 75

44 7 60 2 10 20

44 84 39 42 12 21

66 103 50 108 100 90

54.17 12.50 59.38 56.25 87.50 78.13

60.00 93.64 45.45 98.18 90.91 81.82




accuracy SVM Bayesian Neural Net SVM Bayesian Neural Net

SVM Bayesian Neural Net SVM Bayesian Neural Net


TP count

FP count

FN count

TN count




62.06 53.81 58.01 72.14 86.31 80.66

67.75 12.15 61.85 51.25 86.05 80.95

44.95 6.35 47.4 7.7 13.3 19.85

33.2 88.8 39.1 49.7 14.9 20

60.1 98.7 57.65 97.35 91.75 85.2

67.14 11.97 61.66 52.08 85.22 80.27

57.27 93.94 55.15 93.39 87.37 81.08

2.34 3.26 5.40 5.41 2.94 4.55

November 30 - December 3, 2004, Proceedings, ... Part I (Lecture Notes in Computer Science). Springer-Verlag New York, Inc., Secaucus, NJ, USA, 2005. E. O. Brigham. The Fast Fourier Transform and Its Applications. Prentice-Hall, Inc., Upper Saddle River, NJ, USA, 1988. C. H. Chen, L. F. Pau, and P. S. P. Wang, editors. Handbook of Pattern Recognition &Amp; Computer Vision. World Scientific Publishing Co., Inc., River Edge, NJ, USA, 1993. D. Giannoulis and M. Massberg. Papers digital dynamic range compressor design a tutorial and analysis, 2013. M. Jalil, F. Butt, and A. Malik. Short-time energy, magnitude, zero crossing rate and autocorrelation measurement for discriminating voiced and unvoiced segments of speech signals. In Technological Advances in Electrical, Electronics and Computer Engineering (TAEECE), 2013 International Conference on, pages 208–212, May 2013. S. Larson, G. Comina, R. H. Gilman, B. H. Tracey, M. Bravard, and J. W. Lpez. Validation of an automated cough detection algorithm for tracking recovery of pulmonary tuberculosis patients. PLoS ONE, 7, 10 2012. C. Lightfoot. Some effects of the common cold on speech. Archives of Otolaryngology, 51(4):500–513, 1950. J. Martinek, M. Tatar, and M. Javorka. Distinction between voluntary cough sound and speech in volunteers by spectral and complexity analysis. J Physiol Pharmacol, 59(6), Dec. 2008. S. Mcadams. Perspectives on the contribution of timbre

[11] [12] [13]



[16] [17]

to musical structure. Comput. Music J., 23(3):85–102, Sept. 1999. A. V. Oppenheim and R. W. Schafer. Discrete-Time Signal Processing. Prentice Hall Press, Upper Saddle River, NJ, USA, 3rd edition, 2009. G. Peeters. A large set of audio features for sound description (similarity and classification) in the CUIDADO project. Tech. rep., IRCAM, 2004. M. Sahidullah and G. Saha. Design, analysis and experimental evaluation of block based transformation in mfcc computation for speaker recognition. Speech Commun., 54(4):543–565, May 2012. E. Scheirer and M. Slaney. Construction and evaluation of a robust multifeature speech/music discriminator. In Proceedings of the 1997 IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP ’97)-Volume 2 - Volume 2, ICASSP ’97, pages 1331–, Washington, DC, USA, 1997. IEEE Computer Society. G. Sell and P. Clark. Music tonality features for speech/music discrimination. In Acoustics, Speech and Signal Processing (ICASSP), 2014 IEEE International Conference on, pages 2489–2493, May 2014. A. M. Toh, R. Togneri, and S. Nordholm. Spectral entropy as speech features for speech recognition. Proceedings of PEECS, 1, 2005. B. Tracey, G. Comina, S. Larson, M. Bravard, J. Lopez, and R. Gilman. Cough detection algorithm for monitoring patient recovery from pulmonary tuberculosis. In Engineering in Medicine and Biology Society, EMBC, 2011 Annual International Conference of the IEEE, pages 6017–6020, Aug 2011.



Undergrad Project


Bushra Ferdousi Portfolio Undergrad Project Project Name

Result Processing System of Ahsanullah University of Science and Technology

Project Objective

The aim of the project was to automate the result processing system of my university. The result processing system includes different sub system such as the submission of the results (from the instructors), Gradesheet preparation, Annoncement of the results (to the students) and post processing (such as cumulative results preparation etc.). The project was implemented by C# using Windows Presentation Foundation (WPF).

Data Flow Diagram of Result Processing System

Entity Complete ERD of Result Processing System without attributes. Relationship Diagram (ERD) of Result Processing System


Bushra Ferdousi Portfolio Undergrad Project Segmented ERD of Result Processing System with attributes

Output Report of Result Processing System

The Software Result Processing System has mainly two types of CrystalReportsViewers Reports in the Report Form, which could be accessed through the administration form. The Reports are: â&#x20AC;˘ Admin can Report Student Transcript â&#x20AC;˘ Admin can Report Student Marksheet Report of Student Marksheet

Report of Student Transcript


3D Modeling


Bushra Ferdousi Portfolio 3D modeling 3D Modeling


Here is my 3D model, designed by unreal engine . I have designed two rooms (living room and study room). The rooms are connected by a door. The video of the 3d model is here.

Front view

Front upper view

Backward view

Backward upper view

Inside view of study room

Inside view of living room

I also design some animation in unreal engine. In my one design I used Maximo Vampire Character as an actor. I also have used side scroller 2D project to make a simple game. Fallowing links are the video of my designed animation Maximo Vampire Character.


Samples of Cartoons and Caricatures


Bushra Ferdousi Portfolio Samples of Cartoons and Caricatures Logo Design


Here is some of my published and unpublished logo design.

AUST CSE Alumni Association Logo

ArdhoChondro, a Bangla Band Logo

Purdue, Bangladesh Student Association logo

VERTIGO 42 , Purdue Band Logo Edited By:Tridib K. Saha

When I get free time I love to draw caricature. I have drawn a number of caricature of famous persons from Bangladesh USA. Here are some of them.

Ahsan Habib Cartoonist, Editor of Unmad magazine

Humayun Ahmed Bangladeshi Writer

Trevor Noah South African comedian

Syed Rashed Imam Tanmoy Cartoonist, Assistant Editor of Unmad


Bushra Ferdousi Portfolio Samples of Cartoons and Caricatures Live Caricature

I drew live caricature of people ranging from different ages during Cartoon Fest 2014 in Bangladesh and Purdue Spring Fest 2017. Here are glimpse of some of my live caricatures.

Cartoon Fest 2014 in Bangladesh

Spring Fest 2017, at Purdue University

Spring Fest 2017, at Purdue University

Spring Fest 2017, at Purdue University


Bushra Ferdousi Portfolio Samples of Cartoons and Caricatures Exhibitions

Some of my cartoons were exhibited in several well-known exhibitions in Bangladesh.

Theme: Against Corruption Cartoon Fest 2013 Organized by: Transparency International Bangladesh

Theme: Communication Media Cartoon Fest 2014 Organized by: Bangladesh Cartoonist Association

Theme: Global Warming Cartoon Fest 2016 Organized by: Unmad Magazine

Theme: Facebook Cartoon Fest 2017 Organized by: Bangladesh Cartoonist Association


Bushra Ferdousi Portfolio Samples of Cartoons and Caricatures Cartoons for Magazine

My several cartoons have been published in Unmad Magazine and Purdue, Bangladesh Student Association yearly magazine 2017. This cartoon was published in Unmad Magazine series 301 Theme name: Who are they? Script: Kamrul Islam, Draw: Bushra Ferdousi Idea: In professional life, intimacy is often driven by advantage. We see people in important position are circled by lots of so called â&#x20AC;?well-wishersâ&#x20AC;?. In this cartoon, we tried to categorized who are those people.

1st Panel: Middle aged with clear glass is a film producer. New actor, Sponsor agent, New TV channel agent,Lame Director, Actors Manager always stay beside him. 2nd Panel : In this panel, the second person from the left is a politicians. Flatterer, Businessmen, Lobbyist, gunmen always surround the politician. 3rd Panel : 2nd person from the left side is a doctor. He always stay among Medical Representative, Patient, Herbal Doctor, Insurance Company Representative and Journalist. 4th Panel : Drama producer, local guardian, lover boy and ex-lover stay besides a beautiful girl. 5th Panel : Here Unmad magazine (A satire magazine) is represented as a person (the second person from the left side). He stays among street salesman, newspaper seller, furious fan, cartoonist and cartoon idea consultant.


Bushra Ferdousi Portfolio Samples of Cartoons and Caricatures

This cartoon was published in Purdue, BDSA anual magazine 2017 Theme name: Fill in the gap. Script: S M Ferdous and Bushra , Draw: Bushra Ferdousi Idea: It shows some ways, how a grad student can get their home furniture.

1st way: A grad student can get their furniture from bin, which is a very common in USA. During Moving out, lots of people left useful things near the bin for collection. 2nd way: A grad student can collect furniture from seniors, who may already got those things from their seniors. 3rd way: If a grad student ready to spend money(!), can easily buy furniture from Amazon, Best buy, Walmart and Ikea. 4th way: If a grad student is lucky he can even get stuffs from her advisor. 5th way: There are several Facebook pages (e.g., Free For Sale for Purdue), where a grad student can buy their stuffs in cheap price, even sometimes they can get free stuffs from these pages.



Portfolio of Bushra Ferdousi  
Portfolio of Bushra Ferdousi