torchimage

PyTorch utilities for classic image processing and evaluation

Highlights:

from torchimage.utils import NdSpec introduces a NdSpec class. It automatically converts user input in parameters like kernel_size, which can be a scalar or list of scalars (unknown until runtime), into a wrapped class with a __getitem__ method.
from torchimage.padding import Padder offers the most versatile Pytorch padding functionalities to date, including zeros, constant, replicate, smooth, circular, periodize, symmetric, reflect, antisymmetric, odd_reflect, odd_symmetric, linear_ramp, maximum, mean, median, minimum, empty.
from torchimage.pooling import GaussianPoolNd, AvgPoolNd offers faster gaussian pooling and average pooling at arbitrary dimensions. (It is approximately 10x faster than PyTorch thanks to separable filtering.)
from torchimage.metrics import SSIM, MS_SSIM, PSNR, MSE brings differentiable metrics in image processing.
from torchimage.filtering import Sobel, Prewitt, Farid, Scharr, GaussianGrad, Laplace, LaplacianOfGaussian offers a range of edge detection modules.
from torchimage.filtering import UnsharpMask implements image sharpening algorithm
from torchimage.misc import poly1d calculates fast single-variable polynomial with constant coefficients.

Motivation:

We might want to use some classic image processing algorithms together with a neural network (e.g. loss, preprocessing). By making them differentiable, torchimage allows for their seamless integration into your PyTorch pipeline, which also avoids frequently moving tensors between devices.
Some of these algorithms don't previously have a Python version at all.
Many algorithms cannot process high-dimensional data (such as image batches or video batches) and therefore do not scale.
Being written in PyTorch means that torchimage can be run on both CPU and GPU using the same code, eliminating any need to adapt code implementation.
One of the main reasons for inconsistent behaviors in different image processing libraries is that some functions are implemented slightly differently, the most prominent example being padding (extending border of a signal). Not all packages have the same set of padding options; to make things worse, the same name may refer to different things. torchimage therefore aims to provide a standard for all methods to compare with.

miaotianyi/torchimage