Flower Classification using CNNs

Project Description

Transfer learning project using VGG19 and YOLOv5 for 102-category flower classification.

Dataset

• Oxford 102 Flowers Dataset
• 8,189 images across 102 flower categories
• Split: 50% training / 25% validation / 25% testing
• Two random seeds (42 and 123) used to evaluate model robustness

Results

Model	Seed 42	Seed 123	Trainable Params
VGG19	~84%	~85.3%	~124M
YOLOv5-cls	91.75%	91.89%	~130K

Requirements

• Python 3.8+
• PyTorch 2.8.0
• matplotlib -sickcit learn
• See requirements.txt for full list

Installation

Please Use VENV!

1. Clone the repository

git clone https://github.com/Jacob170/flower_classification.git
cd flower_classification

2. Create virtual environment (recommended)

python3 -m venv venv
source venv/bin/activate  # Linux/Mac
# or
venv\Scripts\activate     # Windows

pip install -r requirements.txt

Code Stracture

1. data_preparation.py - splits the dats 25%|25%|50%
2. evaluate - prints the plots of accuracy and loss
3. train_vgg19.py - training class for vgg
4. train_yolov5.py - training class for yolo
5. preproccessing - augmentation normalization of iamges before training
6. main - main program runs everything (except the data_preparation.py so no split)

Usage

Step 1: Prepare data (run once)

python src/data_preparation.py

This creates stratified train/val/test splits for seeds 42 and 123.

Step 2: Train and evaluate

python src/main.py

This runs preprocessing, training, and evaluation for both models.

Note !! Important !! If you are running this on CPU (without GPU)

in the main change the DataLoader arguments
to work with num_workers=1 or 2 and the pin_memory=False or just remove it This code optimized to work with cuda

Model Architectures

VGG19

• Frozen: 16 convolutional layers (feature extractor)
• Trainable: 3 fully-connected classifier layers
• Modified: Final layer changed from 1000 → 102 classes
• Optimizer: SGD (lr=0.01, momentum=0.9)
• Epochs: 35

YOLOv5-cls

• Frozen: Backbone layers 0-8
• Trainable: Classification head (layer 9)
• Modified: Final linear layer changed from 1000 → 102 classes
• Optimizer: Adam (lr=0.001)
• Epochs: 20

Key Findings

1. Modern architectures win: YOLOv5-cls (2020) outperformed VGG19 (2014) by ~12% accuracy
2. Efficiency matters: YOLOv5-cls trained 1000x fewer parameters
3. Both models are robust: < 0.2% variance between different random seeds
4. Transfer learning works: Achieved 92% accuracy with only ~4,000 training images

Results

• VGG19 Test Accuracy: 85.30%
• YOLOv5 Test Accuracy: 91.89%

Authors

Jacob Yaacubov & Barak Milshtein

GitHub Repository

https://github.com/Jacob170/flower_classification

Dataset

• Primary: Oxford 102 Flowers
• Additional: [if you used any]