top of page
Souradeep Mukhopadhyay
LATEST PROJECTS
Project | 01
Project | 01 Biomedical Image Analysis
I have worked extensively on biomedical image analysis projects, utilizing advanced algorithms and optimization techniques for various medical applications. My work includes cancer detection using breast cancer mammogram images from the DDSM dataset, where I applied the Social Ski Diver optimization algorithm with adaptive beta hill climbing for feature selection. Additionally, I have worked on kidney image segmentation from CT and SPECT images at Siemens Healthineers, using tools like MeVisLab. Another key project involved U-Net for cell nuclei segmentation. These projects have strengthened my expertise in medical imaging, deep learning, and optimization algorithms for healthcare applications.
Project | 02
Project | 02 Generalized Class Discovery
I developed a novel approach for Generalized Class Discovery (GCD) during my internship at IIT Bombay, under the guidance of Prof. Biplab Banerjee. I created GraphVL, a vision-language model leveraging CLIP and a graph convolutional network (GCN). This model effectively addresses GCD by integrating a GCN with CLIP’s text encoder to preserve class neighborhood structures and applies a visual projector with margin-based contrastive losses to map images to textual representations. The work enabled semi-supervised clustering and class discovery, with applications in molecular structure discovery for 3D Cryo-electron tomography, leading to a submission to 'Briefing in Bioinformatics’, Oxford University Press.
Project | 03
Project | 03 Open Set Domain Generalization
I have been working on Open Set Domain Generalization (OSDG) to tackle the challenges posed by domain and category shifts in training and testing data. To address this, I proposed MetaPrompt, a novel framework that leverages the strengths of CLIP’s vision-language capabilities and Meta-Learning’s adaptability. MetaPrompt formulates OSDG as a multi-class classification problem, utilizing domain-agnostic and domain-focused prompts for effective unknown class detection. Additionally, I introduced unsupervised contrastive loss during episodic Meta-Training to enhance generalization and unknown class awareness. This approach aims to improve model robustness in handling unseen domains and classes, with the manuscript submitted to TMLR.
Project | 04
Project | 04 Blind Image Quality Assessment
In the ongoing work, I exploit contrastive learning and CLIP to propose some algorithm that predicts blindly (no reference) Image quality. This work is supervised by my M.Tech Thesis guide Dr. Rajiv Soundararajan in his lab named ‘Visual Information Processing’ (VIP). Few outputs are: a."UPLIP-IQA: Integrating Unsupervised Prompt Learning and CLIP for Image Quality Assessment" is my Advanced Image Processing course final project.
b. I have started working on Prompt distribution Learning in a supervised way to predict the quality of an image blindly as my M.Tech Thesis.
bottom of page