🐾 Smart Animal Classifier

A Hierarchical Multi-Model CNN System for Animal Identification

---

1️⃣ Project Overview

Smart Animal Classifier (AnimaAI) is a deep-learning system that identifies animals from images using a hierarchical pipeline of specialized convolutional neural networks (CNNs). Instead of relying on a single classifier to recognize dozens of species and breeds, our model routes an image through multiple networks to produce a more accurate prediction.

The system is capable of:

• 🐶 Classifying 70+ dog breeds
• 🦋 Classifying 70+ butterfly species
• 🐱 Recognizing multiple cat breeds
• 🐦 Identifying various bird families
• 🐾 Classifying 14 general animal categories

This makes it suitable for wildlife research tools, educational apps, and automated image-sorting systems.

---

2️⃣ Installation

Install required packages:

pip install torch torchvision pillow matplotlib gradio

---

3️⃣ Running the Classifier App

Start the Gradio interface:

python myApp.py

This will:

• Load the router model
• Load all available specialist models
• Launch an interactive web app
• Allow users to upload images for real-time predictions

4️⃣ Problem Statement

Image classification is one of the most fundamental tasks in machine learning. Animals are one of the most common real world targets: from wildlife monitoring and ecological research to educational apps and pet-ID tools. However animals are visually diverse and often have many breeds with similar appearances. A single monolithic neural network that tries to classify everything at once can struggle to distinguish each animal due to things like: large label space, imbalance datasets between species, fine-grained distinctions between visually similar classes, etc. Our project, AnimalAI, tackles this by combining a general router CNN with a set of specialized CNNs. Users upload an image and the system: uses a router model to predict a coarse animal category (e.g. Cat, dog, bird, butterfly, etc.) then, it routes the images to specialist models(e.g. Dog breeds, cat breeds, birds vs butterflies, etc.) finally, the network returns a final prediction with probabilities via Streamlit.

---

5️⃣ Datasets

Datasets Used

General Animals

• Cats, Dogs, Elephants, Horses, Lions
  https://www.kaggle.com/datasets/antobenedetti/animals

  https://www.kaggle.com/datasets/yapwh1208/cats-breed-dataset?resource=download

Dog Breeds

• 70 Dog Breeds Image Dataset
  https://www.kaggle.com/datasets/gpiosenka/70-dog-breedsimage-data-set

Butterfly Species

• Butterfly Image Classification Dataset
  https://www.kaggle.com/datasets/phucthaiv02/butterfly-image-classification
  Training data extracted from Training_set.csv

Bird Species

• Bird Species Dataset
  https://www.kaggle.com/datasets/linh123456/birds-dataset

• Additional Bird Dataset
  https://www.kaggle.com/datasets/samuelayman/bird-dataset

Wolves

• Dogs vs Wolves Dataset
  https://www.kaggle.com/datasets/harishvutukuri/dogs-vs-wolves

Reptiles and Amphibians

• Chameleon, Crocodile, Alligator, Frog, Snake
  https://www.kaggle.com/datasets/vencerlanz09/reptiles-and-amphibians-image-dataset

Big Cats

• Lions and Cheetahs Image Classification Dataset
  https://www.kaggle.com/datasets/mikoajfish99/lions-or-cheetahs-image-classification

---

6️⃣ Training the Models

Training is performed through train_animals.py and additional scripts for specialist models.

Dataset Structure

The main dataset is organized as:

Animals/
  cats/
  dogs/
  birds/
  butterfly/
  snake/
  elephant/
  ...

Each folder contains images for a single coarse class.

Specialist datasets exist for:

• Cat breeds
• Dog breeds
• Butterflies
• Birds
• Big cats
• Snake / Chameleon
• Birds vs butterflies (ambiguity resolution)

Training Details

• Architecture: ResNet-18 (ImageNet pretrained)
• Loss: CrossEntropy
• Optimizer: Adam (lr = 1e-4)
• Batch size: 32
• Epochs: 5–20 depending on dataset size
• Train/Val split: 80/20

Data Preprocessing

All images are:

• Resized to 224×224
• Normalized using ImageNet mean and std
• Randomly flipped horizontally during training

Model Saving

Each trained model is saved as:

{
  "model_state_dict": ...,
  "class_names": ...,
  "label_to_idx": ...
}

Models are stored in the Models/ directory.

---

7️⃣ Deployment & Real-World Integration

This project demonstrates how a trained neural network can be integrated into an actual software application.
Using Gradio, we built a functional web-based classifier that:

• Accepts user-uploaded images
• Feeds them through the router → specialist pipeline
• Outputs the predicted category or breed with probabilities

This shows how machine learning systems can move from development into real, accessible tools.

---

⭐ Credits

Developed for:

CSCI 4050U – Machine Learning
Ontario Tech University

Team Members:

• Christopher Kiige
• Tony Akinniranye
• Lucas Fenkam
• Jedidiah Dennis