Welcome to my GitHub portfolio! I’m an AI Engineer currently working on Computer Vision & Deep Learning. Feel free to explore my work and reach out at [email protected].
- Global Rank: 6th out of 2,100+ participants
- Duration: 4 months
- Hosted by: Microsoft, NVIDIA, BMW Group, & QUT Design Academy
✍️ Summer Challenge on Writer Verification - National Conference on Computer Vision, Pattern Recognition, Image Processing and Graphics - 2nd Position
- Challenge: Verify handwritten signatures to detect potential fraud
- Problem: Tackling the natural variability in offline handwriting
- Achievement: Secured 2nd position in a competitive environment
🎥 Multi-Class Abnormality Classification for Video Capsule Endoscopy - International Conference on Computer Vision & Image Processing - Capsule Vision Challenge 2024
- Approach: Deep learning-based multiclass abnormality classification
- Data: Trained on 37,607 VCE frames and validated on 16,132 frames
- Performance: Mean AUC of 0.98 & Balanced Accuracy of 0.83
- Result: Achieved 8th place in the competition
Leverage Dreambooth & LORA to fine-tune the Stable Diffusion model for personalized text-to-image generation.
- Technique: Custom fine-tuning using a limited set of images to capture unique subject details,
- Pipeline: End-to-end process covering data preparation, fine-tuning, and inference.
- Applications: Ideal for creative projects, personalized image generation & marketing content generation.
A unified Python package for running inference on multiple object detection models (YOLOX, YOLOv3, YOLOv4, YOLOv6, YOLOv7) without the hassle of managing different codebases.
A user-friendly repository to fine-tune Faster R-CNN on any custom COCO dataset, streamlining the process of object detection training.
Implemented the CANet (Chained Context Aggregation Network) to accurately segment cells in microscopy images using the LIVECell dataset.
🥭 SAM-ONNX
A lightweight pip package designed to seamlessly integrate the capabilities of the SAM (Segment Anything Model) with minimal dependencies.
A CNN-based solution paired with a Streamlit app to recognize Indian currency denominations.
Potential Use Cases:
- Assistive tool for visually impaired individuals
- Currency verification system
LW-μDCNN: A Lightweight CNN Model for Human Activity Classification using Radar micro-Doppler Signatures
Abstract:
Developed a novel lightweight CNN model named LW-μDCNN with 438,998 parameters across 7 layers to classify human activities from radar micro-Doppler signatures.
- Model Size: 5.2 MB (reduced to 0.43 MB using quantization-aware training)
- Performance: Achieved 97% accuracy and superior F1-scores compared to state-of-the-art models
- Additional Studies: Employed transfer learning with InceptionV3 and MobileNetV1