Documentation of torch-imagetools

torch-imagetools

An image processing library based on PyTorch that aim to provide some non deep learning-based tools for pre-/post-processing, data augmentation, or an alternative option when no data for training the network.

  • imgtools.balance: chromatic adaptations, illuminant estimations, correlate color temperature estimations.

  • imgtools.color: transformations between color spaces. The spaces including RGB, YUV, HSV, CIEXYZ, LMS, CIELAB, etc.

  • imgtools.core: discrete wavelet transform and its inverse transform, math tools (PCA, filter2d, matrix transform).

  • imgtools.enhance: basic intensity trnasform (linear, log, gamma, …), sharpening/unsharp masking, high dynamic range, low-light enhance.

  • imgtools.filters: spatial domain and frequency domain filters.

  • imgtools.restoration: dehaze.

  • imgtools.statistice: dehaze.

  • imgtools.utils: conversion np.ndarray <-> torch.Tensor (arrayize, tensorize),

This library

  • Docs: https://johnny95731.github.io/torch-imagetools/

Future plan

  • automatic contrast enhancement

  • highlight removal and shadow removal

  • style transfer

  • restoration

  • chromatic adaptation

  • frequency domain methods

  • gradient domain methods

  • statistics tools.

Installation

  1. Clone this repository

pip clone https://github.com/johnny95731/torch-imagetools.git
cd torch-imagetools
pip install -e .

  1. Install PyTorch >= 2.7.0. Since the dependency will keep covering the torch+cuda, torch is not in the dependencies of pyproject.toml.

Optional Dependencies

Install PyWavelets if you want to use the submodule imgtools.wavelets

pip install pywavelets

Usage

  • The library assume that the RGB image is in the range of [0, 1].

  • The functions will try to convert dtype and device to fit the first argument.

  • The tests only test float32 and float64 dtypes. Some features in torch.linalg does not support float16.

  • The image format should be CHW.

  • The docstring indicate the tensor shape and the asterisk * means an optional place. For example, the shape (*, C, H, W) allows (C, H, W) or (B, C, H, W).

import imgtools

hwc_img = cv2.imread(path)
chw_img = imgtools.utils.tensorize(hwc_img)

# Hist equalization
enhanced = imgtools.enhance.hist_equalize(chw_img)
# Color transfer
enhanced = imgtools.enhance.transfer_reinhard(img)
# Color space transformation
xyz = imgtools.color.rgb_to_xyz(img)

License

See LICENSE.md