Frequency-Aware Facial Deepfake Detection
Multi-stream framework using RGB, FFT, and DCT representations for detecting GAN-generated, diffusion-generated, edited, and real face images. Built spectral feature pipelines, CNN/ViT fusion modules, ablation studies, Grad-CAM style interpretability, and robustness evaluation.
Physics-Guided MRI Reconstruction and Tumor Segmentation
Dual-head attention U-Net and hybrid ViT direction for accelerated MRI. Simulated Cartesian and radial undersampling; optimized reconstruction with PSNR, SSIM, NMSE, and segmentation with Dice, IoU, Recall, Specificity, and HD95.
Transformer Models and Vision Transformers
Implemented sequence-to-sequence Transformers, attention mechanisms, positional encoding, and Vision Transformers from scratch. Extended these ideas toward token-based image, k-space, mask, and optional text-conditioning frameworks.
Semantic 3D Reconstruction and Scene Understanding
Exploring geometry-aware pipelines that combine visual encoders, point-cloud reasoning, semantic priors, and Gaussian-splatting-style scene representations for interpretable 3D reconstruction from visual data.
Big Data Search and ML Systems
Built Hadoop-oriented search and analytics prototypes using TF-IDF, MapReduce, approximate search, Bloom filters, sketching ideas, and lightweight interfaces for querying retrieved documents and model outputs.
