Rajatsubhra
Chakraborty

Building DM-QPMNet, a dual-modality fusion network for cell segmentation in quantitative phase microscopy. Developing physics-aware models that stay robust under low-visibility imaging conditions.

Proposed BiasMap, a model-agnostic pipeline that localizes and quantifies entanglement between demographic and semantic concepts in Stable Diffusion via cross-attention attribution. Designed energy-guided diffusion with a differentiable SoftIoU objective for plug-and-play bias mitigation without retraining.

Co-developed LLAVIDAL, a large language-vision model for activities of daily living, and curated the 100K-pair ADL-X dataset with 3D poses and object trajectories. Built a curriculum-based semi-supervised detector that lifted aerial object-detection mAP by 32% using only 20% ground images.

Engineered a U-Net model for cervical texture segmentation in ultrasound images, improving accuracy and clinical usability. Built an automatic multiclass segmentation pipeline that classified medical-imaging regions by tissue type.

Designed deep local descriptors using CNNs for instance-level recognition of historical architecture. Focused on robust feature matching for heritage-building and landmark recognition.

Developed transfer-learning models for handwritten character recognition, reaching 99.99% accuracy with a VGG16 backbone. Worked on Bangla and Devanagari script-recognition pipelines.

Designed an automatic text-detection pipeline for Indus Valley Civilization seals. Explored computer-vision approaches to localizing undeciphered scripts.