Thesis Supervisors

Prof. Dr. Guoying Zhao has supervised my Ph.D. in Technology at CMVS in University of Oulu, Finland. This thesis focused on analysis and recognition of 4-dimensional microexpressions using advanced deep learning.

Prof. Dr. Tareq Y. Al-Naffouri supervised my M.S. thesis in Electrical Engineering at KFUPM, Dhahran, Saudi Arabia. The thesis was related to compressive sensing based collaborative patch-based image denoising algorithms.

Prof. Dr. Mahmood Ashraf Khan supervised my B.S. final year project (FYP) at CUI, Islamabad, Pakistan. The project was about design and development of a real-time vehicle tracking system via GPS and GSM.

• 

  Real-time Face Mask Detection for COVID-19 Pandemic

  Real-time Face Mask Detection using Deep Learning - At University of Oulu, Finland

  To further help combat the corona virus (COVID-19), this project can be used for real-time face mask detection using webcam with an impressive accuracy. For convenience, the source codes are shared both in MATLAB and Python in my GitHub repo.

• 

  4D Facial Expression Recognition

  4D Facial Expression Recognition on BU-4DFE Dataset - At University of Oulu, Finland

  This project proposes a novel 4D Facial Expression Recognition (FER) method using Collaborative Cross-domain Dynamic Image Network (CCDN). Given a 4D data of face scans, we first compute its geometrical images, and then combine their correlated information in the proposed cross-domain image representations. The acquired set is then used to generate cross-domain dynamic images (CDI) via rank pooling that encapsulates facial deformations over time in terms of a single image. For the training phase, these CDIs are fed into an end-to-end deep learning model, and the resultant predictions collaborate over multi-views for performance gain in expression classification. Furthermore, we propose a 4D augmentation scheme that not only expands the training data scale but also introduces significant facial muscle movement patterns to improve the FER performance. Results from extensive experiments on the commonly used BU-4DFE dataset under widely adopted settings show that our proposed method outperforms the state-of-the-art 4D FER methods by achieving an accuracy of 96.5% indicating its effectiveness.

  
• 

  Face Recognition on Live Webcam

  AlexNet based Face Recognition on Self-Generated Live Webcam Dataset - At University of Oulu, Finland

  For this project, I worked (Matlab 2018b) on fine-tuning the final fully-connected layers of AlexNet (a 5 Convolutional and 3 Fully Connected Layers CNN trained on ImageNet of 1000 different classes with 227x227 color images). The final layers of AlexNet was retrained on 1000 sample images of 10 different subjects each (both male and female, and from different countries). The input image was first pre-processed to detect faces by Viola–Jones detector based on which the training images were cropped and fixed to 227x227, as AlexNet expects. After one hour of training (on a 64-bit Intel(R) Core(TM) i7-7700HQ CPU @ 2.80GHz (8 CPUs) processor machine with NVIDIA GeForce GTX 1050 graphic card, and 12GB RAM), the model was tested on live videos by using a webcam. This demo shows the performance of the AlexNet based facial recognition on a live webcam video using faces detected from the boundary-boxes of Viola–Jones detector.

• 

  Deep Learning on CIFAR10 Dataset

  Fine-Tuning ResNet on CIFAR10 Dataset - At University of Oulu, Finland

  For this project, I worked (Python 3.5, TensorFlow, Keras, Anaconda, Spyder) on fine-tuning the final layer of the pre-trained model ResNet-50 v3-fp32 using the CIFAR10 dataset (60,000 32x32 grayscale images in 10 classes, 6000 images/class) to get Training/Testing/Cross Entropy accuracy of 87.05%/86.75%/0.43, 38.92%/34.65%/1.48 and 16.40/15.84%/1.99 at 10,000. 1,000 and 100 iterations, respectively.

• 

  Deep Learning on MNIST Dataset

  Deep Learning Classification on Fashion MNIST Dataset - At University of Oulu, Finland

  For this project, I worked (Python 3.5, TensorFlow, Keras, Anaconda, Spyder) on importing the Fashion MNIST dataset (70,000 28x28 grayscale images in 10 categories), training, exploring and pre-processing the data, building and training a neural network model (Layers: 28*28 - 128 - 10, Parameters: 5 Epochs, traning size = 60,000) with training accuracy of 89.03% and testing accuracy of 87.06%. Performance parameters were later visualized by using the TensorBoard.

  
• 

  Emotions from Multi-modal Signals

  Multi-modal Expressions Recognition System - At University of Oulu, Finland

  For this project, I worked (Python 2.7 with NumPy, SciPy, Matplotlib, Scikit-learn, Skimage, Anaconda, Jupyter Notebook) on designing a multi-modal emotion recognition system to recognize happy versus sadness expressions (binary-class problem) by using the feature-level method to combine the facial expression features and audio features. Extracting the subspace for facial expression feature and audio features using Principal Component Analysis (PCA) and Canonical Correlation Analysis (CCA). Then fusing the features, and finally training a linear binary-class SVM. Accuracy of 98.21% and 92.31% was achieved.

• 

  Emotions from Speech

  Expressions Recognition from Speech Signals - At University of Oulu, Finland

  For this project, I worked (Python 2.7 with NumPy, SciPy, Matplotlib, Anaconda, Jupyter Notebook) on extracting prosodic correlates and cepstral features from speech recordings which are simulated emotional speech (i.e. acted) from ten speakers speaking five different pre-segmented sentences of roughly 2-3 seconds in two different emotional states (happy and sad) totaling 100 samples. The data was resampled, pre-emphasized and then visualized prior extracting the basic Prosodic and Mel-Frequency Cepstral Coefficients (MFCC features). SVM (polynomial kernel), Random Forest and Neural Network (three hidden layers of 50 units each with 20,000 iterations) classifiers were trained and evaluated to get accuracy of 86.21%, 88.31% and 90.35%, respectively.

  
• 

  Face Recognition

  Facial Expressions Recognition from eNTERFACE Dataset - At University of Oulu, Finland

  For this project, I worked (Python 2.7 with Skimage, NumPy, Matplotlib and DLib, Anaconda's Jupyter Notebook) on extracting a sub-set (100 samples) of eNTERFACE dataset (acted facial expressions from 10 actors with happy and sadness classes). Face preprocessing was performed to detect face and facial landmarks via DLib library, followed by face regsteration to a set of fixed landmarks from a standard model, and then extracting face from the registered image. LBP features were then extracted to train a binary-SVM classifier. Accuracy (reached 72.25%) and confusion matrix were computed to evaluate the performance.

  
• 

  Real-Time Fast Video Deraining and Desnowing

  Real-time and Fast Algorithm Designed - At Pukyong National University, South Korea

  Many computer vision applications assumed that the image/videos received are clean and can be used in its received form for real-time applications. However, many videos usually suffere from a number of degradations during the acquisition phase. One of the most common form of visual degradation is that of snow and rain which can hinder the visibility, for example, for a driver-less autonomous vehicle where clear visual input is extremely important. Therefore in this project, a video deraining and desnowing algorithm was designed to be able to remove the rain streaks from the input. Since the real-time applications asks for a faster processing, the algorithm was designed by reducing its computational desires to be able to used as a real-time fast desnowing algorithm. The designed algorithm mainly works on collecting the patches from neighbouring frames (temporal similarities) and implementing a weighted avergae approach over various channels thereby exploiting both the inter-frame as well as inter-channel correlation. The clean videos obtained as a result of this algorithm demonstrate it superiority.

  
• 

  Compressed Sensing based Image Denoising

  M.S. Thesis in Electrical Engineering - At King Fahd University of Petroleum and Minerals, Saudi Arabia

  In this thesis, we propose novel patch-based image denoising algorithms to take care of the undesired additive white Gaussian noise (AWGN) components introduced in an image. Our algorithms are broadly classified into grayscale and color image denoising. We propose to utilize normalized Euclidean distance and correlation based grouping of patches in our algorithms. Such grouping of patches is an effective way for collaborative support-agnostic sparse reconstruction to denoise a noisy patch. For the collaboration, we stack similarly structured patches via distance and correlation based intensity-invariant approach. The key idea is that since similar patches share the same support in the transformed domain, these supports can be used as probabilities of active taps to refine the sparse estimates. This ultimately yields a very good patch estimate and therefore increases the quality of recovered image by discarding the noise-causing components. For the smooth regions of an image, a specially developed post-processor is then applied to further increase the quality of the denoised image. Comparison results from extensive simulations against the existing state-of-the-art algorithms in terms of PNSR and SSIM over a wide range of scenarios demonstrate the superiority of our proposed algorithms.

  
• 

  GPS and GSM Based Vehicle Tracking System

  B.S. Final Year Project - At COMSATS University Islamabad, Pakistan

  It is a vehicle tracking system in which vehicles are tracked and controlled through a mobile phone just by sending an SMS. To trace the location of the car the owner of the vehicle simply sends an SMS to the tracking system installed in the vehicle. The tracking system in the vehicle contains a GPS receiver module and a GSM modem interfaced to a microcontroller. On receiving the SMS from the owner, the GSM modem acknowledges the controller, the controller takes location's longitude and latitude coordinates from the GPS receiver module, packs it in to an SMS and sends back to the owner through the GSM modem. The SMS is also sent to a GSM modem which is connected to a web server. When the web server receives the SMS containing the location coordinates of the vehicle, it shows the location of the vehicle in the form of an image taken from Google maps. The owner can also turn off the main ignition simply by sending an SMS in case of any vehicle theft.

• 

  RFID Based Remote Authentication System

  Embedded System Design Project - At COMSATS University Islamabad, Pakistan

  It's basically an embedded system that contains RFID Module (RFID tag reader) ID-20, GSM Module SIM900 and Microcontroller ATmega16 all mounted on a single board. The main theme of the project is to sense a tag by reading the code from the tag when it is moved near to the RFID reader. Then that code is authenticated by comparing the code with the list of codes in the database. If the code is authenticated, then specific function is performed and an SMS containing the info of the tag and the owner of the tag is sent to the administrator.

• 

  Line Tracking Robot

  ALPHATRON: A Quick and Accurate Robot - At COMSATS University Islamabad, Pakistan

  This project basically resulted with a deliverable of a line tracking robot. The crux of this project was to make a autonomous robot that could track a white line over a dark blue arena and could complete the arena by moving through some specified points as faster as possible. The robot's equipped with sensors that could distinguish between blue and white part of the arena. After taking part in a national level line tracking robotics competition ROBIAN'12, this robot got 2nd/Runner Up position out of 42 robots which participated in the event.

• 

  Client-Server Architecture on Linux OS

  Network/Socket Programming Project - At COMSATS University Islamabad, Pakistan

  This project was based on programming a client-server architecture through which a client can ask the server for services. The project, and hence the programming, was done in Linux OS. The concepts of network and socket programming were implemented to give this project an efficient as well as a professional interface. The programming language was C/C++. Login/Authentication feature was implemented in this project so whenever a client needed to ask for a service, it would first login to the system in order to ensure the authorization for the services. File transferring feature was also implemented through which files were sent from server to the client, upon request. Chat and a number of other features were implemented too. This project was also capable of handling multiple clients.

Gallery