HandGestureMaskPredictor

This repository contains code for a segmentation project using the UNet architecture. The project involves training a deep learning model to perform image segmentation on a custom dataset of hand gestures.

Overview

Our model has been trained for 80 epochs and evaluated on a dataset consisting of 196 images. The dataset is divided into 124 train samples, 32 validation samples, and 40 test samples. The performance metrics for the model on the test set are as follows:

Metric	Value
Test Loss	0.17
Test Dice Coefficient	0.89
Test Accuracy	0.93

Table of Contents

• 📚 Dataset
• 💻 Installation
• 🛠️ Data Preparation
• 🧠 Model
• 🔧 Utils
• 🚀 Training
• 📈 Result

📚Dataset

The dataset is organized into the following directories:

• Dataset

  • Fist
  • OpenPalm
  • PeaceSign
  • ThumbsUp

• New_Mask

  • Fist_Mask
  • OpenPalm_Mask
  • PeaceSign_Mask
  • ThumbsUp_Mask

💻Installation

1. Clone the repository:

   git clone https://github.com/shimaazizi/HandGestureMaskPredictor.git

   cd HandGestureMaskPredictor

2. Install the required packages:

   pip install -r requirements.txt

The librarie we need:

🛠️Data Preparation

This part provides a PyTorch-based pipeline for image segmentation, including data loading and preprocessing.

🧠Model

This part provides a PyTorch implementation of the U-Net architecture for image segmentation.

• Encoder: Captures image features at multiple scales using a series of convolutional blocks and pooling layers.

• Decoder: Upsamples feature maps and concatenates them with corresponding encoder outputs to reconstruct the segmented image.

• U-Net: Combines the encoder and decoder, with a final convolutional layer to produce the segmentation map.

🔧utils

This part includes functions to evaluate and visualize image segmentation model performance using PyTorch.

Functions:

• accuracy (pred, target): Computes the accuracy of predictions against the target masks.

• dice_score (pred, target, epsilon=1e-6): Calculates the Dice coefficient for segmentation tasks.

• visualize_prediction (model, test_loader, device, num_classes=4): Visualizes model predictions alongside true masks and original images

🚀Training

This module trains and evaluates a U-Net model for image segmentation using PyTorch.

Functions:

• train_model(model, train_loader, val_loader, test_loader, num_classes=4, num_epochs=80, device='cuda'): Trains the model and evaluates it on validation and test sets. Saves the model to unet_model.pth.

• evaluate_model(model, test_loader, device='cuda'): Evaluates the model on the test set

📈Result

in main.py script integrates dataset loading, model training, evaluation, and visualization.

• create_dataloaders: Loads and preprocesses the dataset.

• UNet: Defines the U-Net model architecture.

• train_model: Trains the model on the training set and evaluates on the validation set.

• evaluate_model: Evaluates the trained model on the test set.

• visualize_prediction: Visualizes the model predictions compared to the true masks.

Example of Results