CV

An explainable AI project for renal structure segmentation in CT imaging. SegResNet was fine-tuned on multi-phase renal CT scans, reaching a Dice score of 0.78 for organ segmentation. To make the model’s decisions transparent, 3D Grad-CAM and Guided Backpropagation were integrated to produce volumetric, class-specific visualizations that show which anatomical regions drive each prediction.

A deep learning pipeline for automatic annotation and retrieval of scientific figures. The system builds on CLIP and BiomedCLIP — including fine-tuned variants — to embed biomedical images and text into a shared semantic space, and integrates Wikidata-based knowledge graphs to enrich that space with structured domain knowledge. The result is stronger semantic alignment between figures and their descriptions, along with interpretable visualisations that make the annotation and retrieval behaviour of the models easier to inspect.

A medical image denoising system that fuses vision and language: a UNet-based denoiser is conditioned on CLIP text embeddings derived from the VQA-RAD dataset, letting textual context guide the reconstruction of noisy scans. Compared to a vision-only baseline, the vision-language fusion improved reconstruction quality by 12.5% in PSNR and 4.6% in SSIM.

Built a real-time tracking system in C++ with OpenCV for cue ball and table detection. Visualized ball trajectories across angles using dynamic 2D minimaps.

Enhanced PointNet with SHOT descriptors and custom triplet loss. Achieved 94.12% accuracy for robust 3D object recognition on noisy datasets.

Boosted DenseNet121 classification accuracy to 98.67% via PCA, DCT, and noise-based augmentation (baseline - 93.6%).

Implemented ICP with KD-Tree and SVD for faster, precise alignment in industrial 3D reconstruction tasks.

Built a 3D reconstruction pipeline using ORB matching, RANSAC, and geometric validation from multi-view images.

A stereo depth estimation project that combines classical and learning-based approaches: a monocular disparity prediction is used as the initial guess, which Semi-Global Matching (SGM) then refines into the final disparity map. Seeding SGM with the monocular estimate substantially improved accuracy over the baseline, reducing mean squared error by 75%.

Programmed ROS1-based routines for autonomous movement and pick-place using AprilTag recognition.