KAUST AI Training — Deep Learning Projects

A collection of computer vision projects completed as part of the KAUST AI training program, covering classification, semantic segmentation, and transfer learning using PyTorch.

---

Projects

🥔 Potato Disease Classification

Classifying potato leaf diseases from the PlantVillage dataset using a CNN built from scratch.

• 5-layer CNN with BatchNormalization
• Residual (skip) connection from layer 2 → layer 4
• Best Val Accuracy: 98.61% (10 epochs)
• Dataset: PlantVillage — 3 classes (Early Blight, Late Blight, Healthy)

---

🌊 Underwater Image Segmentation

Pixel-level segmentation of underwater scenes into 8 semantic categories.

• U-Net with EfficientNet-B1 encoder (pretrained on ImageNet)
• Dataset: SUIM — divers, fish, reefs, robots, and more
• Fine-tuned classifier head, frozen encoder backbone

---

🚭 Smoking Detection Classifier

Binary classifier detecting whether a person is smoking or not.

• ResNet-18 fine-tuned for binary classification
• BCEWithLogitsLoss with sigmoid thresholding
• Separate train / validation / test evaluation

---

🛰️ Water Body Segmentation

Detecting and segmenting water bodies in satellite imagery.

• U-Net with EfficientNet-B1 encoder
• Binary segmentation (water vs. non-water)
• Mean IoU tracked alongside loss

---

Stack

• Framework: PyTorch + torchvision
• Models: CNN from scratch, ResNet-18, U-Net + EfficientNet-B1
• Tools: segmentation-models-pytorch, kagglehub, tqdm, matplotlib

---

KAUST AI Training Program — 2025