Image Segmentation with UNet++ and EfficientNet

This project focuses on image segmentation, specifically for segmenting humans in various environments and settings. The goal is to generate accurate masks for human figures in images using a deep learning model. The project leverages the UNet++ architecture with EfficientNet as the encoder, trained on a dataset of images and corresponding masks.

Table of Contents

• Overview
• Dataset
• Data Preprocessing
• Model Architecture
• Training
• Results

Overview

Key features of the project include:

• Dataset: Images of humans in diverse environments with corresponding masks.
• Data Augmentation: Enhanced the data using the albumentations library.
• Model: UNet++ architecture with EfficientNet as the encoder, pretrained on ImageNet.
• Training: Trained using a combination of Dice Loss and Binary Cross Entropy, with Adam as the optimizer.
• Performance: Significant reduction in validation loss achieved in just 10 epochs.

Dataset

The dataset consists of:

• Images: RGB images of humans in various environments.
• Masks: Corresponding binary masks for human segmentation.

The dataset is preprocessed and split into training and validation sets for model training.

Data Preprocessing

To improve model generalization and robustness, the following steps were taken:

1. Data Augmentation: Applied using the albumentations library. Techniques include:
   • Random cropping
   • Horizontal flipping
   • Rotation
   • Brightness/contrast adjustments
2. Preprocessing:
   • Images and masks are resized and reshaped to (Channels, Height, Width) format supported by PyTorch.
   • Normalization: Pixel values are scaled to the range [0, 1] by dividing by 255.

Model Architecture

The model is based on the UNet++ architecture, which is an extension of the classic UNet model. Key components include:

• Encoder: EfficientNet (pretrained on ImageNet) is used as the encoder to extract feature maps.
• Decoder: The decoder path in UNet++ is designed to capture fine-grained details for accurate segmentation.
• Input: 3-channel RGB images.
• Output: Binary masks for human segmentation.

Training

The model was trained with the following configuration:

• Loss Function: A combination of Dice Loss and Binary Cross Entropy to handle class imbalance and improve segmentation accuracy.
• Optimizer: Adam optimizer with a learning rate of 0.001.
• Batch Size: 32.
• Epochs: 10.

Results

• Validation Loss: A significant reduction in validation loss was observed within 10 epochs, demonstrating the model's ability to learn effectively.
• Performance: The model achieves high accuracy in generating segmentation masks.

Credits

Credits to Vikram Shenoy for providing the dataset. https://github.com/VikramShenoy97/Human-Segmentation-Dataset