Project Website: https://xknt21.github.io

This project is introduced in detail on the official project website above, including system design, implementation details, and experimental results.

Overview

This project is a mobile and AI-based system for stray cat re-identification, developed as part of a 15-week Project-Based Learning (PBL) course.
The system aims to support the Trap-Neuter-Return (TNR) program by preventing redundant medical treatments and improving coordination between volunteers, animal hospitals, and organizations in the Kansai region of Japan.

The application integrates:

  • A React Native mobile app
  • A Flask-based backend
  • A deep learning-based cat re-identification model

Motivation

In TNR activities, the same stray cat is often captured and treated multiple times due to the lack of a reliable identification system. This leads to:

  • Unnecessary medical procedures
  • Increased burden on animal hospitals
  • Inefficient use of resources

This project aims to solve these issues by enabling image-based identification of individual cats.

System Architecture

The system consists of three main components:

1. Frontend (Mobile & Web)

  • Built with React Native (Expo)
  • Allows users to:
    • Capture or upload cat images
    • View identification results
    • Access cat profiles and medical records

2. Backend (Flask API)

  • Handles:
    • Image processing requests
    • Model inference
    • Database communication
  • Provides RESTful APIs for both mobile and web clients

3. AI Model (Cat Re-Identification)

  • Siamese Neural Network for similarity learning
  • Generates embeddings and performs matching

Model Pipeline

  1. User uploads or captures a cat image
  2. YOLO-based model detects and crops the cat face
  3. Image is resized and normalized
  4. Siamese network generates a 128-dimensional embedding
  5. Embedding is compared with stored embeddings using Euclidean distance
  6. If distance < threshold (0.4), a match is returned; otherwise, no match

Model Architecture

Siamese Network

  • Backbone: EfficientNetB3 (default), VGG16 or MobileNetV2 (optional)
  • Embedding Head:
    • Global Average Pooling
    • Dense (128 units, ReLU)
    • Batch Normalization / Dropout
    • L2 normalization

Loss Functions

  • Contrastive Loss (primary)
  • Triplet Loss (experimental)

Dataset

  • Total Cats: 250
  • Total Images: 1,880
  • Average Images per Cat: 7.5

Sources:

  • Kaggle Cat Re-ID datasets
  • HelloStreetCat dataset
  • Custom scraped datasets

Training Details

  • Optimizer: Adam (lr = 0.0001)
  • Batch Size: 16
  • Epochs: 50 (production)
  • Image Size: 200×200
  • Data Augmentation:
    • Flip, rotation, noise

Performance

Contrastive Model

  • Accuracy: 69.4%
  • Precision: 67.7%
  • Recall: 69.4%
  • F1 Score: 68.2%

Key Insights

  • Contrastive learning performs well with limited data
  • Triplet loss requires more data and tuning
  • Model is suitable for practical deployment

Key Features

Cat Re-Identification

  • Image-based matching with confidence scores
  • Detects previously treated cats
  • Prevents duplicate medical interventions

Medical Record Management

  • Stores treatment history, vaccination status
  • Supports updates by animal hospitals

Role-Based System

  • Volunteers: upload images and check matches
  • Hospitals: manage medical records
  • Admins: system monitoring and analytics

Admin Dashboard

  • Match statistics
  • Activity tracking
  • Dataset insights

System Design Principles

  • No Auto-Registration Policy

    • Prevents incorrect or duplicate entries
    • Only authorized admins can register new cats

  • Privacy & Security

    • API key authentication
    • Secure data handling (TLS, encryption)

  • Scalability

    • Designed for up to 1000 users
    • Supports real-time inference

Tech Stack

  • Frontend: React Native, Expo
  • Backend: Python, Flask
  • AI: TensorFlow, Keras
  • Computer Vision: OpenCV, YOLO
  • Data: NumPy, Pandas

My Contributions

  • Designed and implemented the AI model pipeline
  • Developed the Flask backend for model inference
  • Contributed to system integration (mobile ↔ backend ↔ AI)
  • Evaluated model performance and improved training process

Challenges

  • Limited dataset size for deep metric learning
  • Difficulty in tuning triplet loss
  • Ensuring real-time performance on mobile environments
  • Maintaining data quality without auto-registration

Future Work

  • Improve accuracy with larger datasets
  • Optimize model for edge/mobile deployment
  • Enhance UI/UX for non-technical users
  • Expand deployment beyond Kansai region

Impact

This project demonstrates how AI can support animal welfare by:

  • Reducing redundant medical treatments
  • Improving efficiency in TNR operations
  • Enabling data-driven coordination between stakeholders

It reflects my interest in applying computer vision and machine learning to socially impactful real-world problems.