SandraKorol/catguard-efficientnet

CatGuard EfficientNet

Overview

CatGuard EfficientNet is an image classification model designed to identify a specific domestic cat named Syrnyn from photographs. The model was developed as a first-year university computer vision project and serves as the first step toward a future IoT monitoring system capable of detecting when a specific cat enters a restricted area.

---

Problem

One of the household cats frequently enters the kitchen and attempts to eat food left unattended. Monitoring this behavior manually is inconvenient and inconsistent. The goal of this project is to automatically recognize the target cat from camera images.

---

Task

Binary image classification.

Input:

• Image containing a cat

Output:

• Naughty cat
• Other Cat

---

Dataset

Custom dataset collected from personal photographs.

Classes:

• Naught Cat (black cat)
• Other Cat

Dataset split:

• Train: 70%
• Validation: 15%
• Test: 15%

---

Model Architecture

The project uses a two-stage pipeline:

Image
 ↓
DETR Object Detector
 ↓
Cat Detection
 ↓
EfficientNet-B0
 ↓
Classification Head
 ↓
Naughty Cat / Other Cat

Stage 1: Object Detection

The system first uses DETR (DEtection TRansformer) (facebook/detr-resnet-50) to determine whether a cat is present in the image.

Possible outcomes:

• Cat detected → continue to classification
• No cat detected → return a warning message

Stage 2: Cat Classification

If a cat is detected, the image is passed to an EfficientNet-B0 classifier trained using transfer learning.

The classifier predicts one of two classes:

• Naughty Cat
• Other Cat

---

Results

Best validation accuracy:

89.7%

The model correctly identifies the target cat in approximately 9 out of 10 validation images.

---

Future Work

Current version:

Image
 ↓
Cat Classification

Planned extension:

Camera
 ↓
Cat Detection
 ↓
Cat Classification
 ↓
IoT Device Response

The future system may automatically detect the target cat entering the kitchen and trigger a connected IoT device.

---

Author

Sandra Korol Computer Vision Project