Atayaz/bsds500_transfer_learning

🧠 BSDS500 Image Segmentation with CNN and Transfer Learning

This project implements two deep learning approaches to perform image segmentation on the Berkeley Segmentation Dataset (BSDS500):

1. 1.`bsds500_segmentation_cnn.py` – A U-Net-based Convolutional Neural Network (CNN) built from scratch
2. 2.`bsds500_transfer_learning.py` – A Transfer Learning model using ResNet50 as encoder with a custom decoder

Both models are deployed using Streamlit, enabling interactive training visualizations and real-time segmentation predictions.

📁 Dataset

Dataset Used: BSDS500 - Berkeley Segmentation Dataset

The dataset must be structured as follows:

• —archive/images/train/ → input images (.jpg or .png)
• —archive/ground_truth/train/ → corresponding ground truth masks (.mat files with groundTruth segmentation data)

Each .jpg or .png file must have a .mat file with the same base name.

⚙️ How to Run

Install dependencies:

pip install -r requirements.txt

Run CNN model:

streamlit run bsds500_segmentation_cnn.py

Run Transfer Learning model:

streamlit run bsds500_transfer_learning.py

🧠 Model 1: CNN (bsds500_segmentation_cnn.py)

Key Features:

• —Custom U-Net with 5+ convolutional layers
• —3+ max pooling layers
• —Dropout regularization
• —Binary mask prediction via sigmoid output
• —Training and validation accuracy/loss plots
• —Real-time segmentation of uploaded images
• —Model saved as bsds500_cnn_model.h5

🧠 Model 2: Transfer Learning (bsds500_transfer_learning.py)

Key Features:

• —Encoder: Pretrained ResNet50 (imagenet weights)
• —Decoder: Custom CNN-based U-Net-like layers
• —Skip connections for semantic segmentation
• —Binary mask prediction via sigmoid output
• —Training and validation metric visualization
• —Real-time mask prediction for uploaded images
• —Model saved as bsds500_resnet50_transfer.h5

🖼️ Streamlit Interface Includes:

• —Training graphs (Accuracy & Loss)
• —File uploader to select image
• —Predicted segmentation mask display

💾 Model Output Files