Atish020/butterfly-semantic-segmentation-unet

🦋 Semantic Image Segmentation of Butterfly Images using U-Net

  ![U-Net]()  ![HTML5]() ![CSS3]() ![JavaScript]() ![Docker]() ![Mean IoU]() ![Dice]()

This project implements a Computer Vision–based semantic image segmentation system to accurately extract butterfly regions from complex backgrounds at the pixel level. The solution uses a U-Net architecture, deployed as a real-time web application using FastAPI, HTML/CSS/JavaScript, Docker, and Hugging Face Spaces.

🚀 Demo

• —Upload a butterfly image
• —Get real-time segmentation mask
• —View original image & predicted mask side-by-side [image]

The application runs fully inside a Docker container and exposes a FastAPI backend for inference.

📌 Problem Statement

Traditional object detection methods provide only bounding boxes, which are insufficient for tasks requiring precise object boundaries. Butterfly images pose additional challenges due to:

• —Complex and cluttered backgrounds (flowers, leaves, sky)
• —Color similarity between butterfly wings and background
• —Fine-grained wing structures

The goal is to classify each pixel as either:

• —Butterfly (foreground)
• —Background

🧠 Proposed Solution

This problem is solved using Semantic Image Segmentation with a U-Net architecture, which combines:

• —Encoder path for contextual feature extraction
• —Decoder path for precise spatial localization
• —Skip connections to preserve fine details

The trained model predicts a binary segmentation mask highlighting only the butterfly region.

🧩 Model Architecture

• —Architecture: U-Net
• —Input size: 256 × 256 × 3
• —Loss Function: Binary Cross-Entropy
• —Metrics: Accuracy, Dice Coefficient, Mean IoU
• —Framework: TensorFlow / Keras

📊 Performance

• —Validation Accuracy: ~96%
• —Mean IoU (Validation): ~0.89
• —Dice Coefficient: ~0.94
• —Inference Time: ~100 ms per image (CPU)

🗂️ Dataset

Leeds Butterfly Dataset (BMVC 2009)

• —832 butterfly images
• —Pixel-level segmentation masks
• —10 butterfly species
• —Images collected from Google Images and manually filtered

🧱 Project Structure

Semantic-image-segmentation-using-UNet/
│
├── app.py
├── Dockerfile
├── requirements.txt
├── README.md
│
├── backend/
│   └── main.py        # FastAPI backend logic
│
├── frontend/
│   ├── index.html     # UI
│   ├── style.css      # Styling
│   └── script.js      # Client-side logic
│
├── saved_model/
│   └── final_UNET_Butterfly_Segmentation.keras
=======
title: Butterfly Semantic Segmentation Unet
emoji: 👀
colorFrom: yellow
colorTo: blue
sdk: docker
pinned: false
license: apache-2.0
short_description: Pixel-level butterfly image segmentation using U-Net
---