Research Outline (During Master's):
Measurement Framework for Continuous 1-D Signal (ECG)
Measurement Framework for 2-D Signal (non-aspect ratio preserved)
Measurement Framework for 2-D Signal (aspect ratio preserved)
- A deterministic sensing based framework for segmented ECGs for continuous monitoring.
- Quality improvement of the recovered signals using a newly developed Kronecker-based recovery technique.
- Comparative study of recovery quality using random sensing techniques, as opposed to deterministic sensing.
- Performance evaluation of the deterministic sensing framework using different sparsifying basis, like DCT, wavelets (at different levels of decomposition).
- Verification robustness of deterministic sensing technique in presence of noise at different levels for different compression ratios.
- Presenting statistical analysis on a large dataset, with an emphasis on pathologically significant signals.
Measurement Framework for 2-D Signal (non-aspect ratio preserved)
- Reduction in size of sensing matrices using segmentation based parallel CS approach, for traditional column/row wise CS compression of images.
- Extending 1-D Kronecker-based approach developed to improve the quality of the recovered signals to 2-D.
- Comparative analysis between the random and deterministic sensing matrices.
- Evaluation of the quality of the recovered images while changing segmentation length.
- Structural similarity analysis for the recovered images (between Kronecker-based and non Kronecker-based technique) for various compression levels.
- Presenting statistical analysis on a large MRI dataset.
Measurement Framework for 2-D Signal (aspect ratio preserved)
- A deterministic framework for aspect ratio and morphology preserving image compression.
- Analysis of structural similarity on the compressed images using Frobenius norm.
- Compressed domain analysis of feature based template matching using SURF algorithm.
Projects:
Distributed Stochastic Gradient Descent with Quantized Compressive Sensing.
Please find the presentation here.
Optimal Resource Allocation in Federated Learning: A Stackelberg Game Approach.
User Selection in Federated Learning.
Please see my presentation here.
Investigation of Kronecker-based Recovery in Compressive Sensing.
In many real life applications, such as recording of bio-signals using wearable sensors or surveillance drones, signals are continuously acquired and transmitted wirelessly to a remote system for 'processing on demand'. One of the efficient ways of continuous measurement and transmission while using the resource constrained sensors is to acquire signals in smaller segments. Although segmentation has advantages in the sensing side, the quality of the reconstructed signals reduce as an effect. The aim of the work was to explore the possibility of 'joint recovery' of the individual segments, in order to improve the reconstruction signal quality. A Kronecker-based solution is adopted exploring the orthonormality of the sparse basis to perform one time recovery while increasing the quality of the reconstructed signals, as opposed to the standard technique.
Check the related publications here.
Find a detailed presentation here. (Partially presented at 2019 IEEE International Instrumentation & Measurement Technology Conference (I2MTC) at a tutorial: 'Signal Quality: From Wearables to Hospitals')
Check the related publications here.
Find a detailed presentation here. (Partially presented at 2019 IEEE International Instrumentation & Measurement Technology Conference (I2MTC) at a tutorial: 'Signal Quality: From Wearables to Hospitals')
Segmentation of Heart Rate from Breathing Rate Based on UWB Radar Estimation.
From the measured dataset of breathing rate estimation of a subject by UWB Radar, we noticed that the output signal remains flat/doesn't show any variation in certain regions. I am currently investigating whether we can find heart rate from those areas of the signal, where breathing is absent.
Remote Heart & Breathing Rate Detection by Skin Colour Variation.
Blood chromatophores flowing through the veins and arteries, during systole and diastole, reflect light differently. As the amount of blood in the arteries varies during cardiac activity, the skin color changes. I worked to develop a novel algorithm, for video signal processing, to achieve remote heartbeat which can bypass the impact of skin colour and ambient light effects. Too add another and more robust dimension to this, the work was extended to the application of radar signal processing. From the measured dataset of breathing rate estimation of a subject by UWB Radar, we noticed that the output signal remains flat/doesn't show any variation in certain regions. I investigated whether we could find heart rate from those areas of the signal, where breathing is absent.
Human Face & Eye Detection Using Haar Cascade Classifier for Video Analysis.
Haar Cascade based object detection is one of the most efficient and widely used object detection technique. It's a machine learning based approach where Haar Wavelets are trained from a pool of positive and negative images, to identify the desired object. Two desired classifiers, front face and eye, were needed to be trained in this case. I used the pre-trained classifiers available in Github. Then the second challenge lies with the feature point matching. This process is as followed. Pick a scale (eg. 24X24 pixels) for the feature, slide it across the image, compute the average pixel values under the white and black area, if difference between these areas is above some threshold; the feature point match is found. But, Haar classifier is a very weak classifier. Hence, many classifiers are needed to be used to make an array of weak Haar Classifiers, which in tern makes it strong. So, the next objective of this project will be extending the work to bring Adaboost Algorithm into frame.
Study of Illumination in a Closed Room Under Different Conditions & Determination of Light Sensor Characteristics.
This is my final year project. The main objectivity of the project is energy saving by controlling room light intensity automatically. For the purpose, the a standard day and a standard room, the Lux Meter readings are taken under different lighting conditions at different time. From the Lux values, the Lux value curve has been plotted. This curve gives a precise idea about the range of Lumens that the desired sensor must operate in between to meet the purpose. Theoretical calculations have also been done to calculate the solid angle for verifying the theoretical result with the practical one. We analysed the characteristics of PV cell & LDR, found R2 error from trend-line observation, equating short ckt. current with Lux. Finally, the composite function is found and characteristic curve of LDR is plotted. Finally the main objective, study closed room illumination under different lighting conditions & determination of light sensor characteristics, was met.
Read the full project report here.
Read the full project report here.
Leaf Disease Detection, Support Vector Machine Approach.
Leaf disease and its detection play a major role in agricultural research. But several times due to lack of professional experience proper disease prediction isn't done. In this project we are working on developing an algorithm which will implement the precise accuracy based disease detection. Firstly computer will be trained with standard data sets of diseased leafs and then by applying pre processing, enhancement, extraction and feature based matching we will be getting the desired outcome, the disease. For convenience a MATLAB GUI will be designed.
Bio-metric face recognition system.
The program efficiently detects face/faces from a picture. It can distinguish the face of human & nonhuman creature. Standard Viola Jones algorithm has been used. Basic drawback of this project is, it doesn't perform well for tilted image or in darkness. Further research will include feature extraction for face detection or using a machine learning approach.
Feature extraction from an image.
In this hobby project, I have designed a program in MATLAB for executing Gabor Function. By selecting specific values of parameters we can extract different features from an image. This algorithm can further be extended for Face Detection Using Gabor Wavelet. Please contact for details.
Optical Character Recognition (OCR).
In this hobby project, I have taken a standard data set of characters. Then by matching the template image with the standard data set the desired outcome has been displayed. The algorithm works fine for a specific allowance of tilt and for both upper and lower case. I would like to extend this work further by training neural network. Coding and execution of this algorithm has been done in MATLAB.
Making of a voice controlled robot.
It’s an extension of the previous project. The robot follows voice command which involves the concept of visual basic through UB 2008 Exp. Edition and Arduino. This program simply recognizes voice commands, like- front, right, left, back etc. Then the program translates voice command and then it’s fed to the MATLAB which communicates with Arduino controlled robot.
Object detection and tracking based on image processing.
In this project a robot has been made driven by Arduino UNO, communicated by MATLAB and standard webcam. The robot follows an object of specific color (Red/Yellow/Blue) with higher level of accuracy. Further research on this topic will include object avoidance mechanism in this specific robot. Please contact for details.
Making of a Photoshop like image editor for execution of multiple pre-processing functions.
Attractive software has been developed for image pre-processing operations by necessary algorithms. It contains a GUI which efficiently performs some pre-processing functions like image rotation, color frame changing, image addition and lots more. We will be developing an app for further use of this project. Please contact for details.
Making of a Gesture Controlled Robot.
Interfacing of MATLAB with Arduino UNO Board has been done. The robot consists of a standard webcam which captures the gesture and moves as per the command. The work has been further extended to iRobot platform over Wired Serial and Wireless Bluetooth communication protocols. Please contact for details.
2014-23 © Dipayan Mitra