🏆 Multi-Domain AI Suite featuring High-Precision Segmentation and Edge-Ready Classification
This repository contains two production-ready Computer Vision engines developed to solve complex visual tasks in high-stakes industries. By bridging Medical Diagnostics and Precision Agriculture, this workspace demonstrates a versatile mastery of modern Deep Learning architectures, including Residual U-Net (ReUNet) and MobileNetV2.
Due to the significant computational resources required for training (approx. 3 days), this repository includes pre-trained weights and detailed performance reports for immediate verification.
- Architecture: Implemented a Residual U-Net (ReUNet) to handle fine-grained boundary detection in 2D MRI slices.
- Optimization: Developed custom Dice Similarity Coefficient (DSC) and Dice Loss functions to overcome the high class-imbalance inherent in medical imaging.
- Rigor: Evaluation includes Pixel-wise Confusion Matrices and Heatmaps generated via Seaborn to ensure clinical-grade precision.
- Architecture: Utilized MobileNetV2 with Transfer Learning, optimized for lightweight deployment on mobile or edge devices.
- Scale: Trained on a dataset of 70,000+ images across 38 distinct plant disease classes.
- Results: Achieved a robust 91.73% Validation Accuracy with a final validation loss of 0.2475.
- Deep Learning: TensorFlow, Keras
- Computer Vision: OpenCV, Matplotlib
- Data Science: NumPy, Pandas, Scikit-Learn
- Infrastructure: Kagglehub, Google Colab