Hi! My name is Darren Ramsook & I’m a PhD Candidate at Trinity College Dublin (under the Sigmedia Group), working on perceptually optimized video restoration (a mixture of video processing and deep learning). I have a with a keen interest in image/video signal processing. This interest in signal processing is mainly due to the arts having a special place in my heart & also because I’m an avid musician. I have my MSc. Data Science, BSc. Electrical & Computer Engineering from the University of the West Indies. I combine my love for signal processing, software engineering and data science to help systems better analyze the world around us.
Projects that I've worked on:
Traditional metrics for evaluating video quality do not completely capture the nuances of the Human Visual System (HVS), however they are simple to use for quantitatively optimizing parameters in enhancement or restoration. Modern Full-Reference Perceptual Visual Quality Metrics (PVQMs) such as the video multi-method assessment fusion (VMAF) function are more robust than traditional metrics in terms of the HVS, but they are generally complex and non-differentiable. This lack of differentiability means that they cannot be readily used in optimization scenarios for enhancement or restoration. In this paper we look at the formulation of a perceptually motivated restoration framework for video. We deploy this process in the context of denoising by training a spatio-temporal denoiser deep convultional neural network (DCNN).
Video Restoration Human Visual System Perceptual Optimization Deep Convolutional Neural Networks Tensorflow/Keras
There has been limited use of perceptual visual quality metrics for automated image quality optimisation in applications like video enhancement. This is because many useful visual quality metrics for video e.g. VMAF are complex and non-differentiable. This work proposes the use of a CNN architecture to approximate the temporal behaviour of VMAF. Employing degradation generated with H.265 compression, our model achieves a 4.41% RMSE in predicting VMAF. This can now be deployed as a video based loss function in video enhancement and compression tasks.
Perceptual Optimization Video Processing Deep Convolutional Neural Networks Tensorflow/Keras
Research into two models for summarizing an instant message, through the use of text and the Emoji Unicode Characterset, for applications where screen space is limited (eg smartwatches). The first proposed model utilizes a Greedy N-gram token replacement method, while the second proposed model utilizes a Transformer network. The Transformer network was trained on a set of 532 manually labelled instant messages and 92554 automatically labelled dialogue sentences. Initial results showed the Greedy N-gram token replacement method produced higher quality results and this method was evluated using human participants. It was found that there is an increase in average time taken to read and interpret the summarized messages when compared to the unsummarized messages.
Summarization Abstractive Summarization Extractive Summarization Transformer
Developed and implemented a framework for determination of a single Figure of Merit (FoM) that can be used for high level monitoring while at the same time providing sufficiently valuable low level indicators to assist with the isolation and detection of problems. This was illustrated using data from a real cellular network.
Figure of Merit Performance Monitor Data Analytics Key Performance Indicator Telecommunications Infrastructure
This program utilizes the combined access log format of apache web servers to extract certain key features of actions and classifies users in one of three possible groups (Malicious, Unknown & Safe). This system allows for a headless operation once deployed. The system automatically performs pre-processing, classification and populates a list of Internet Protocol addresses whose access to the web server is prohibited. This system was made open source for both public use and development. This trained model resulted in an accuracy score of 94.5% on the test set.
Machine Learning Classification & Regression Trees (CART) Big Data SKLearn Python Software Development
A decision tree learning calculator for the Iterative Dichotomiser 3 (ID3) algorithm. By utilizing the ID3 Algorithm, the best feature to split on is decided. This program requires to additional libraries outside of the default libraries included with Python (math, csv). Therefore this needs to extra set-up configuration. Tested and working on Python 3.7.
Python Optimization Decision Tree ID3
This calculator was built based on the assumption that within a given Social Network, friends of a given person has an effect on the probability of that person clicking an advertisement. This calculator evaluates a scenario where advertisements are split up between two stages, with the second stage being affected by what has occured in the first stage. For more information about this check out this paper - On the Problem of Multi-Staged Impression Allocation in Online Social Networks, done by Inzamam Rahaman and Patrick Hosein.
Optimization Operations Research Big Data Algorithms Probablistic Modelling Python Tkinter
This cash register was designed, created and deployed for use in the Central Bank Mueseum of Trinidad & Tobago. It allows the user to scan items located in the mueseum. The user would then be presented with information about the scan as well as the prices across various time periods. It was created with the main user base being younger children.
Software Engineering Python UI/UX Design Interactive Electrical System Raspberry Pi
The created system was able to detect the heart-rate & glucose level of an individual, store the data, and give out an alarm if the heart rate drops below a certain threshold. A keypad & LCD display was used as means of system interaction.
Microcontroller PIC C Sensor Design Circuit Design Testing/Analysis