reImage - Interactive Graph-Cut Image Segmentation

Professional image segmentation tool using graph-cut algorithm with AVX2-optimized C++ backend and PyQt6 GUI.

Features

• Interactive Scribble Interface - Draw foreground/background seeds with adjustable brush
• AVX2 Optimized - 3-4x faster with SIMD vectorization
• Auto-Display Results - See segmentation overlay instantly
• Export Results - Save segmented images with transparency
• Standalone Executable - No installation required (Windows .exe provided)

Quick Start (Windows)

Download reImage.exe from releases and double-click to run. No installation needed!

Algorithm

Uses graph-cut segmentation with:

• Boykov-Kolmogorov (BK) and Dinic max-flow algorithms for min-cut computation
• Gaussian Mixture Models (GMM) with 5 components per model (fitted via Expectation-Maximization) and fallback to 8×8×8 RGB histograms
• Adaptive β for pairwise smoothness terms
• 4-neighborhood graph structure

Building from Source

Prerequisites

C++ Backend:

• CMake 3.10+
• GCC/Clang (with AVX2 support) or MSVC
• C++23 compiler

Python GUI:

• Python 3.12+
• uv (package manager)

Build C++ Backend

Windows (MinGW):

cd cpp/build
cmake .. -G "MinGW Makefiles"
cmake --build . -j

Linux / macOS:

cd cpp
mkdir -p build && cd build
cmake .. -DCMAKE_BUILD_TYPE=Release
make -j$(nproc)

Build

Install Python dependencies:

uv sync
uv pip install pyinstaller pyqt6

Create standalone executable:

uv run --no-sync pyinstaller reImage.spec --clean

Output: dist/reImage.exe

# Rectangle mode
./cpp/build/segment image.bin W H rect x0 y0 x1 y1 output.bin

# Mask mode (with scribbles)
./cpp/build/segment image.bin W H mask seed.bin output.bin

Optimizations

AVX2 SIMD & Memory Layout

• pre-converted aligned double pixel array for direct SIMD loads
• FMA-powered polynomial minimax approximations for exp and log (removing slow scalar math calls)
• Batch color distance computations (4 pixels at once)
• Vectorized operations in beta calculation
• ~10-25x speedup on graph construction and cost evaluation

Compiler Flags

• -O3 - Maximum optimization

• -march=native - CPU-specific instructions

• -mavx2 - Enable AVX2 explicitly

• -ffast-math - Aggressive floating-point

• LTO enabled for minimal binary size

• Inline getColor() (called millions of times)

• alignas(32) memory alignment for SIMD

• Cache-friendly loop ordering (horizontal/vertical separation)

• Const/constexpr where applicable

Project Structure

reImage/
├── cpp/                    # C++ backend
│   ├── main.cpp           # CLI interface
│   ├── Image.{h,cpp}      # Image loading (AVX2-aligned)
│   ├── SeedMask.{h,cpp}   # Foreground/background seeds
│   ├── Gmm.{h,cpp}        # Gaussian Mixture Models (EM fitting)
│   ├── DataModel.{h,cpp}  # GMM/Histogram-based unary costs
│   ├── GraphBuilder.{h,cpp} # Graph construction (AVX2)
│   ├── BoykovKolmogorov.{h,cpp} # BK Max-flow algorithm
│   ├── Dinic.{h,cpp}      # Dinic Max-flow algorithm
│   ├── Segmenter.{h,cpp}  # Orchestration
│   ├── MinCut.h           # Min-cut extraction
│   ├── SimdOps.h          # AVX2 intrinsics
│   └── CMakeLists.txt
├── gui_app.py             # PyQt6 GUI application
├── reImage.spec           # PyInstaller build config
└── pyproject.toml         # Python dependencies

Performance

• Binary Size: 132KB (C++), 91MB (standalone GUI with dependencies)
• Typical Segmentation: <0.3s for 640x480 images, <1.7s for 1080p images using GMM + BK
• Max-Flow Speedup: Up to 25x speedup with Boykov-Kolmogorov compared to Dinic solver
• Memory: ~40MB for graph construction on HD inputs