VetVision

VetVision is a desktop AI assistant for dog breed recognition and breed-aware veterinary guidance. The public repository combines a CustomTkinter desktop application, an EfficientNet-based training script, local label assets, PDF export, and optional Gemini-powered report generation.

The opening visual above is assembled from the original report screens and evaluation artifacts so the repository leads with the real product surfaces.

What the application does

• Loads a dog image from disk with file picker support and optional drag-and-drop
• Produces the top breed predictions from the trained model
• Shows the primary breed label and confidence score in the desktop UI
• Generates a veterinary-style text report with Gemini when an API key is available
• Exports the current result set as a PDF document
• Includes a training script for rebuilding the classification pipeline

Repository reality

• The current desktop app uses a clean light CustomTkinter interface.
• The current training script is EfficientNetB0-based.
• Model, evaluation, and dataset artifacts are grouped under artifacts/ so the source tree stays clearer without losing the original delivery assets.

Academic context

• Course: YBS 4015 Yapay Zeka
• Project title: VetVision: Yapay Zeka Destekli Köpek Irkı Analizi ve Veteriner Asistanı
• Project window: Oct 2025 - Dec 2025
• Team: Batuhan Yüksel, Yusuf Yılmaz, Savaş Avcı, Ekin Çelik

Tech stack

Area	Tools
Desktop app	Python, CustomTkinter, Pillow
Inference	TensorFlow, NumPy
Report generation	Google Gemini API, ReportLab
Training	TensorFlow Keras, pandas, matplotlib
UX extras	tkinterdnd2 for drag-and-drop
Environment	python-dotenv

Repository structure

.
|-- app.py
|-- train_model.py
|-- llm_test_api.py
|-- create_charts.py
|-- requirements.txt
|-- artifacts/
|   |-- model/
|   |   |-- vetvision_model.h5
|   |   `-- labels.txt
|   |-- dataset/
|   |   |-- labels.csv
|   |   `-- sample_submission.csv
|   `-- evaluation/
|       |-- confusion_matrix.png
|       `-- report.txt
`-- docs/assets/

Running locally

1. Create a virtual environment:

   python3 -m venv .venv
   source .venv/bin/activate

2. Install the dependencies:

   pip install -r requirements.txt

3. Create an environment file if you want Gemini-backed report generation:

   cp .env.example .env

4. Add your Gemini key to .env if you want AI-generated veterinary summaries:

   GEMINI_API_KEY=your-key-here

5. Launch the desktop application:

   python app.py

If GEMINI_API_KEY is not set, the app still works for local breed recognition and PDF export, but the LLM-backed veterinary narrative remains disabled.

Training pipeline

• train_model.py organizes the dataset into train and validation folders
• the current script uses transfer learning with EfficientNetB0
• augmentation is applied through ImageDataGenerator
• the best model is saved as artifacts/model/vetvision_model.h5

Evaluation artifact

The repository already includes the confusion matrix generated during the academic delivery:

Notes on scope

• This repository is strongest as a course project handoff and desktop AI demo.
• The public snapshot still includes the original heavy model and evaluation artifacts, but they are grouped under artifacts/ to reduce root-level clutter.
• A production packaging pass would normally move the trained model to a release asset or model registry instead of versioning it in the source tree.

License

Released under the MIT License.