CNN Projects — Deep Learning with TensorFlow/Keras

A collection of Convolutional Neural Network projects built from scratch and using transfer learning — covering image classification, production deployment, and model explainability.

---

Projects

1. 🌿 Plant Disease Detection (EfficientNetB0)

Transfer learning model detecting 15 plant diseases from leaf images

• Accuracy: 93.09% validation accuracy
• Model: EfficientNetB0 (ImageNet pretrained, frozen)
• Classes: 15 (Tomato, Potato, Pepper Bell diseases)
• Trainable Params: 19,215 out of 4,000,000
• Explainability: Grad-CAM implemented
• Deployed: Flask REST API → Docker → Render (live)
• Key decisions: Confidence thresholding (0.60), two-phase training strategy

🔗 Live API: https://plant-disease-api-l7mi.onrender.com 🔗 Deployment Repo: plant-disease-api

---

2. CIFAR-10 Image Classifier (Custom CNN)

Built a custom CNN from scratch achieving 80.6% validation accuracy on 10-class image classification

• Accuracy: 80.6% validation accuracy
• Architecture: Conv→BN→Pool ×3, Dropout(0.25), Dense(128), Softmax
• Dataset: CIFAR-10 (60,000 images, 10 classes)
• Key learnings:
  • BatchNorm placement and effect
  • Data augmentation strategy
  • GAP vs Flatten trade-offs
  • Output shape calculation
  • Transfer learning limitations on low-resolution images (32×32)
  • Confusion matrix analysis

---

Tech Stack

TensorFlow · Keras · Flask · Docker · Gunicorn · Render · NumPy · Matplotlib · Grad-CAM

---

Skills Demonstrated

Skill	Project
Custom CNN architecture design	CIFAR-10
Transfer learning & fine-tuning	Plant Disease
Production API development	Plant Disease
Docker containerization	Plant Disease
Cloud deployment	Plant Disease
Model explainability (Grad-CAM)	Plant Disease
Confidence thresholding	Plant Disease
BatchNorm, Dropout, Augmentation	Both
Confusion matrix analysis	Both

---

👤 Author

Kuldip Lakhtariya — B.Tech Student | ML & DL Engineer

🔗 GitHub · Live Plant Disease App