MRG.
Initializing Experience...
Case StudyComputer Vision / AIJun 2026 — Present
CONFIDENTIAL COMPUTER VISION PROJECT @ CHRIST UNIVERSITY
Agricultural Vision & Spectral Analytics System
An advanced computer vision curation engine and multimodal deep learning platform engineered for agricultural specimen analysis. Integrates deep salient object segmentation, anti-aliased Lanczos-3 resizing, dynamic bounding-box auto-cropping, and Vis-NIR diffuse reflectance spectroscopy into a synchronized, standardized dataset pipeline.
Key Highlights
Key Highlights | Stipend-Funded Research · Agricultural Vision · Christ University
- Confidential software platform developed under Christ (Deemed to be University) guidance (Stipend-Funded Research Grant)
- 4-stage deterministic image curation engine converting >12MP uncalibrated benchtop photography into standardized ML artifacts
- U2-Net salient foreground extraction generating high-fidelity continuous alpha mattes with persistent VRAM session caching (110ms/image)
- 11,500 multimodal synchronized artifacts across 2,300 physical specimens with exact 50:50 binary organic/conventional balance
Tech Stack & Tools Used
Python
PyTorch
U2-Net
Vis-NIR
OpenCV
01. Data Acquisition & Scaling
Ingested macro photography and Theremino Vis-NIR 400nm–1000nm diffuse reflectance spectral curves. Implemented proportional Lanczos-3 downsampling (a=3 sinc kernel) bounded at 1200px, eliminating aspect distortion while preserving micro-lenticels and skin striations.
02. Neural Background Removal
Deployed U2-Net nested ReS-UNet architecture for salient object detection, computing continuous floating-point alpha mattes to cleanly separate specimens from laboratory bench clutter, ceramic plates, and diffuse contact shadows.
03. Compositing & Adaptive Cropping
Formulated Porter-Duff Over alpha compositing onto pure white (#FFFFFF) canvases to eliminate dark border halos, coupled with vectorized bounding-box detection adding 15% dynamic padding for auto-centered thumbnail generation.
04. Outcome & Dataset Synchronization
Delivered 11,500 paired dataset artifacts across 2,300 specimens (Groundnut, Soybean, Banana, Sapota) with automated canonical regex renaming (CU_SDD_...), ensuring seamless data integrity for downstream multimodal classification.
Outcome
Engineered an automated 4-stage computer vision pipeline standardizing 11,500 multimodal dataset artifacts.
Eliminates shortcut learning from ambient laboratory clutter, ensuring robust visual-spectral alignment across diverse fruit and legume geometries.