CUDA-based Seam Carving

A high-performance implementation of content-aware image resizing using CUDA. This project implements the seam carving algorithm to intelligently resize images while preserving important content.

Example

Original Image

Image after Seam Carving

Features

• CUDA Acceleration: Leverages GPU parallelism for fast energy map computation and seam carving.
• Seam Removal and Insertion: Supports both removing and inserting seams for image resizing.
• Multiple Resizing Modes:
  • Vertical and horizontal seam removal/insertion.
  • Target resizing to specific dimensions.
• Seam Visualization: Option to save and visualize seam paths on the image.
• Experimentation Script: Automated script to run predefined experiments and generate outputs.

Prerequisites

• NVIDIA GPU with CUDA support
• CUDA Toolkit
• C++ compiler
• Make build system

Installation

1. Clone the repository:

git clone https://github.com/FILALIHicham/seam-carving-cuda.git
cd seam-carving-cuda

2. Build the project:

make

Usage

Basic Commands

./build/seam_carving -i <input_image> -o <output_image> [options]

Options

• -n <number>: Number of seams to remove/insert
• --horizontal: Operate on horizontal seams instead of vertical
• --insert: Insert seams instead of removing them
• --save-seams: Save visualization of seam paths
• --target <width>x<height>: Resize to specific dimensions
• --optimized: Use optimized computation methods

Examples

1. Remove 50 vertical seams:

./build/seam_carving -i data/tower.jpg -o output.png -n 50

2. Insert 30 horizontal seams:

./build/seam_carving -i data/tower.jpg -o output.png -n 30 --horizontal --insert

3. Resize to specific dimensions:

./build/seam_carving -i data/tower.jpg -o output.png --target 800x600

Running Experiments

The project includes an automated experiment script (exp.sh) that runs various seam carving operations and saves the results:

1. Make the script executable:

chmod +x exp.sh

2. Run experiments:

./exp.sh

The script will:

• Clean and rebuild the project
• Create an outputs directory
• Run multiple seam carving operations with different parameters
• Save results and visualizations

Experiment Output

Results will be saved in the outputs directory with the following naming convention:

• output_remove_vertical_n{number}.png: Vertical seam removal
• output_remove_horizontal_n{number}.png: Horizontal seam removal
• output_insert_vertical_n{number}.png: Vertical seam insertion
• output_insert_horizontal_n{number}.png: Horizontal seam insertion
• output_target_resize_{number}.png: Target size resizing

Project Structure

└── seam-carving-cuda/
    ├── README.md
    ├── LICENSE
    ├── Makefile
    ├── exp.sh
    ├── data/           # Input images
    ├── include/        # External libraries
    ├── build/         # Compiled binaries
    └── src/           # Source code

Implementation Details

The project uses several optimization techniques:

• Parallel computation of energy maps
• Efficient cumulative energy calculation
• GPU memory management
• Shared memory utilization for faster access

Acknowledgments

• Seam Carving for Content-Aware Image Resizing (Avidan and Shamir)
• STB Image library for image I/O operations