image_enhancement

Library of NumPy-based and CUDA-based Image Processing functions for various types of Image Enhancement, including:

Spatial Tone Mapping
Local Contrast Enhencement
Color Correction (White Balance)
Color Saturation Adjustment
Exposure Fusion

Examples:

Functions:

contains:

get_photometric_mask(): Estimates local brightness distribution (edge-aware blurring).
apply_local_contrast_enhancement(): Adjusts strength of local details.
apply_spatial_tonemapping(): Adjusts brightness levels in dark and bright regions.
transfer_graytone_to_color(): Transfers tones from a tone-mapped grayscale image to a color image.
change_color_saturation(): Adjusts the color vibrance of an image.
correct_colors(): Removes color casts from images.
adjust_brightness(): Adjusts global brightness of the image.
srgb_to_linear(): Transforms image to the linear color space.
linear_to_srgb(): Transforms image to the gamma-corrected color space.
enhance_image(): Applies multiple stages of enhancement to an image.
blend_expoures(): Fuses a collection of image exposures to a single well exposed image.

├── source                                     [Directory: Source code]
│   ├── image_enhancement.py                   [Main script with all the functions]
│   ├── image_enhancement_cuda.py              [Script with ~most~ of the functions in CUDA]
│   ├── example_color_correction.py            [Example of applying color correction]
│   ├── example_enhance_image.py               [Example of combined image enhancement]
│   ├── example_local_contrast_enhancement.py  [Example of applying increasing local details]
│   ├── example_blend_exposures.py             [Example of blending multiple image exposures]
│   ├── example_medical_image.py               [Example of processing medical images]
│   ├── cuda
│   │  ├── image_enhancement.cu.h              [Header file for CUDA implementation]
│   │  └── image_enhancement.cu                [Implementation of image enhancement in CUDA]
└── images                                     [Directory: Sample test images]

Dependences

numpy
imageio
skimage (can be easily bypassed if needed)
Pillow (optional: only needed for running image_enhancement_cuda.py)
pycuda (optional: only needed for running image_enhancement_cuda.py)

Dataset

If you want to try this code in real-life challenging travel photos, please try the following dataset:

TM-DIED: The Most Difficult Image Enhancement Dataset

Building CUDA code

The Python code loads a compiled cuda binary (.cubin) and runs the CUDA kernels using numpy arrays. Replace [arch] in the code below to compile the CUDA code (e.g. -arch=sm_72).

nvcc --cubin -arch=[arch] -use_fast_math -O3 source/cuda/image_enhancement.cu -o source/image_enhancement.cubin

With image_enhancement.cubin in source, you can run source/image_enhancement_cuda.py to generate an enchanced version of alhambra1.jpg.

Citations

If you use this code in your research please cite the following papers:

bbonik/image_enhancement