Semantic-Image-Segmentation-model

Background-Foreground Separation (BFS) using Deep Learning

Overview

This project implements a semantic segmentation model using PyTorch to separate foreground objects (vehicles, pedestrians) from background in traffic camera images. The model is trained on a dataset of urban traffic scenes and aims to improve background-foreground separation (BFS) through deep learning techniques.

Features

• Custom PyTorch Dataset class for structured data loading
• UNet-based deep learning model for pixel-wise segmentation
• IoU (Intersection over Union) evaluation for performance assessment
• Optimized training with AdamW optimizer and ReduceLROnPlateau scheduler
• Visualization tools to display predictions overlaid on original images

Dataset

• Contains images from suburban traffic cameras in Chicago, IL
• Annotation masks highlight foreground objects (vehicles, pedestrians)
• Images are resized for faster training and inference

Model Architecture

• UNet-inspired architecture for semantic segmentation
• Encoder with convolutional layers, batch normalization, and dropout
• Decoder with transposed convolutions for upsampling
• Sigmoid activation to produce probability masks

Training & Evaluation

• Binary Cross-Entropy Loss (BCELoss) for training
• IoU (Intersection over Union) as primary metric
• Training pipeline supports GPU and CPU execution
• 400 epochs with dynamic learning rate adjustment

Inference & Visualization

• Predicts segmentation masks on new traffic camera images
• Outputs original images, binary masks, and overlayed predictions

Example Model Predictions

Training Metrics