tinyscaler

A small CPU image scaling library with SIMD support on x86_64 and Arm (Neon). This project is aimed to replace OpenCV for image resizing, resolving installation inconveniences and compatibility issues. We developed this for future use in Gymnasium and PettingZoo wrappers.

Installation

You can install from PyPI using pip install tinyscaler. Linux and macOS with Python >=3.8 are supported.

Usage

Tinyscaler contains a single external function, scale that using a numpy array input for the image and the new resized shape, returns the resized image.

import numpy as np
import tinyscaler

img = np.random.rand(64, 64, 4).astype(np.float32)

resize_img = tinyscaler.scale(img, (32, 32))
print(resize_img.shape, resize_img.dtype)  # (32, 32) np.float32

TinyScaler supports mode='area', mode='bilinear', and mode='nearest' filtering. It also allows one to pass a destination buffer in order to avoid duplicate memory allocations.

Area filtering is only really useful for downscaling, bilinear will be used even when area filtering is set if upscaling. Area filtering is also likely not worth it when downscaling less than or equal to 2x.

TinyScaler is used through a single function. The full signature is:

scale(src : np.ndarray, size : tuple, mode='area', dst : np.ndarray = None)

Note that the size tuple parameter is (width, height). However, the numpy arrays have dimensions ordered as (height, width, channels). This is similar to OpenCV.

TinyScaler expects a contiguous numpy array. If it is not contiguous, it will throw an error. You can make a non-contiguous numpy array contiguous by calling np.ascontiguousarray. Usually a numpy array will already be contiguous.

If the final array dimension is not 4 (RGBA), it will automatically convert to it. Further, if the array is uint8, it will be converted to float32. So the prefered array has a shape (height, width, 4) and dtype=np.float32.

Finally, downscaling is the focus of TinyScaler. It can also upscale, but it will not be as fast as a more complex separable algorithm in that case.

Performance

In a simple benchmark, we resized the same image (4928x3279) down to (852x567) 100 times using bilinear filtering with several libraries. Here are the times (in seconds) spent (measured with Python's perf_counter) on a AMD 1950x:

Time elapsed for tinyscaler: 0.7968465110002398
Time elapsed for OpenCV: 0.48667862100001
Time elapsed for Pillow: 12.672875003999707
Time elapsed for skimage: 164.45401711399973

And with area filtering (just TinyScaler and OpenCV):

Time elapsed for tinyscaler: 4.34793155800071
Time elapsed for OpenCV: 8.118138265999733

All methods were forced to use a single thread. OpenCV is slightly faster than TinyScaler for bilinear filtering, but TinyScaler remains very fast regardless.

Interestingly, for area filtering, TinyScaler is faster (almost 2x).