The Python Imaging Ecosystem: OpenCV, scikit-image & Pillow

"OpenCV calls me a BGR Mat, scikit-image calls me a float in [0, 1], and Pillow just wants to know if I have an EXIF rotation. I contain multitudes, all of them the same bytes."

A Thoroughly Benchmarked Test Image

In the previous section you learned what an image is in Python. This section is about who you will be working with. The ecosystem can look chaotic from outside: at least six major libraries overlap on basic operations, each with its own defaults. Viewed from inside, the structure is simple, and it is the structure shown in Figure 0.2.1: NumPy in the center, everything else orbiting it.

1. The Lay of the Land Beginner

Practically every imaging library in Python either operates directly on NumPy arrays or converts to and from them in one line. That single design decision, made independently by different communities over two decades, is why the ecosystem composes so well. The map looks like this:

The boxes in Figure 0.2.1 divide into three tiers of involvement. The big three (OpenCV, scikit-image, Pillow) are the subject of the rest of this section. The utility tier (SciPy's ndimage for N-dimensional filtering, imageio for uniform file access, Matplotlib for display) appears throughout the book wherever it is the simplest tool. The deep learning tier, torchvision and friends, takes over in Chapter 18; for now you only need to know that the handoff from NumPy to tensors is cheap and routine.

2. OpenCV: The Industrial Workhorse Beginner

OpenCV is a C++ library born at Intel in 1999, and cv2 (installed as opencv-python) is its Python binding. When a function exists in OpenCV, it is usually the fastest CPU implementation you can call from Python without writing native code yourself: the hot paths are hand-optimized, SIMD-vectorized, and often multithreaded. Coverage is enormous, from JPEG decoding through video capture, classical feature detectors (the stars of Chapter 10), camera calibration, and a DNN inference engine.

The price of that power is a set of conventions that predate the rest of the Python ecosystem. OpenCV stores color as BGR rather than RGB. It defaults to uint8 arithmetic with saturation. Its size arguments are $(width, height)$ tuples while NumPy shapes are $(height, width)$. None of these are bugs; all of them are tripwires, and Section 0.4 is devoted to walking through them in slow motion.

import cv2
import numpy as np

print(cv2.__version__)                  # e.g. 4.10.0

img = np.zeros((200, 300, 3), dtype=np.uint8)
img[:, :, 2] = 255                      # in OpenCV's BGR world, channel 2 is RED

small = cv2.resize(img, (150, 100))     # note: (width, height), not (h, w)!
print(small.shape)                      # (100, 150, 3)  -> rows, cols, channels

blurred = cv2.GaussianBlur(img, (5, 5), 1.0)   # 5x5 kernel, sigma = 1.0
print(blurred.dtype)                    # uint8: OpenCV keeps your dtype

3. scikit-image: The Scientist's Library Beginner

scikit-image grew out of the SciPy community with a different set of values: algorithmic transparency, NumPy-native interfaces, and well-cited reference implementations. Its functions are organized into readable submodules (filters, transform, measure, morphology, restoration), its documentation reads like a short textbook, and when you want to know exactly what an algorithm does, the source is clean Python and Cython you can actually study. Where OpenCV speaks uint8 by default, scikit-image thinks in floats: most functions accept any dtype but return float64 images scaled to $[0, 1]$.

import numpy as np
from skimage import filters, transform

rng = np.random.default_rng(0)
img = rng.integers(0, 256, (100, 150), dtype=np.uint8)

resized = transform.resize(img, (50, 75))      # anti-aliased by default
print(resized.dtype, resized.min().round(3), resized.max().round(3))
# float64 0.137 0.864   -> note: float in [0, 1] now, not uint8!

edges = filters.sobel(img)                     # gradient magnitude
print(edges.dtype, edges.max().round(3))       # float64 0.633

That automatic float conversion embodies a real philosophical difference. scikit-image prioritizes numerical correctness (anti-aliasing on by default, no integer overflow possible) over preserving your storage format. OpenCV prioritizes throughput and in-place processing of camera streams. Neither is wrong; mixing them blindly is. The Sobel filter shown above, by the way, is our first meeting with image gradients, which become edge detection in Chapter 9.

4. Pillow: The File-Format Specialist Beginner

Pillow is the maintained fork of PIL, the Python Imaging Library, and it is the elder statesman of this ecosystem: its lineage predates NumPy itself. Uniquely among the big three, Pillow's central abstraction is not an array but an Image object that knows its format, color mode, and metadata. This makes Pillow the right tool for the file layer of an application: opening anything (it speaks dozens of formats), honoring EXIF orientation from phone cameras, generating thumbnails, converting palettes, saving with fine format control. Web frameworks, Django among them, lean on Pillow for exactly these jobs.

import numpy as np
from PIL import Image

# Pillow -> NumPy: one call each way.
pil_img = Image.new("RGB", (300, 200), color=(255, 60, 0))  # size is (W, H)!
arr = np.asarray(pil_img)
print(arr.shape, arr.dtype)         # (200, 300, 3) uint8  -> back to (H, W, C)

arr = arr.copy()                    # np.asarray may give a read-only view
arr[:100] = (0, 120, 255)
back = Image.fromarray(arr)         # NumPy -> Pillow
print(back.size, back.mode)         # (300, 200) RGB

back.thumbnail((128, 128))          # in-place, preserves aspect ratio
print(back.size)                    # (128, 85)

Note the small landmine in that snippet: Pillow's size is $(W, H)$, the geometric convention, while the array that comes out of np.asarray has shape $(H, W, C)$, the matrix convention. You have now seen this width-height swap twice in one section. It is not a coincidence; it is the oldest schism in computer graphics, and Section 0.4 gives it a full subsection.

5. Choosing Your Tools: A Decision Guide Intermediate

Table 0.2.1 condenses the section into the reference card you will actually use. The dialect columns (color order, default dtype) matter more than the feature columns, because features overlap but dialects clash.

A reasonable default policy for this book and for real projects: use OpenCV as the backbone (it will be our main tool from Chapter 2 onward), borrow scikit-image when you need an algorithm OpenCV lacks or want a trustworthy reference implementation, and let Pillow or imageio own the file boundary when formats get exotic. Whatever you choose, write the choice down: a comment like # contract: uint8 BGR (H, W, 3) at the top of a pipeline file is worth an hour of debugging.

small = cv2.resize(img, None, fx=0.25, fy=0.25,
                   interpolation=cv2.INTER_AREA)   # 1 line, SIMD-fast

6. Interoperation in Practice Intermediate

Let us close with the canonical mixed pipeline: each library doing the one job it is best at, with explicit conversions at the seams. This pattern, Pillow for the file layer, OpenCV for processing, scikit-image for a specialty algorithm, Matplotlib for display, is one you will write hundreds of times.

import numpy as np
import cv2
from PIL import Image
from skimage import img_as_ubyte
from skimage.restoration import denoise_tv_chambolle

# 1. File layer: Pillow opens anything and fixes phone-camera rotation.
#    (Here we synthesize instead, so the snippet runs without assets.)
pil_img = Image.new("RGB", (320, 240), (180, 120, 60))
rgb = np.asarray(pil_img).copy()                # uint8, RGB, (H, W, 3)

# 2. Processing layer: OpenCV expects BGR; convert AT THE BOUNDARY.
bgr = cv2.cvtColor(rgb, cv2.COLOR_RGB2BGR)
bgr = cv2.GaussianBlur(bgr, (5, 5), 1.2)

# 3. Specialty algorithm: scikit-image's TV denoiser (no cv2 equivalent).
float_rgb = cv2.cvtColor(bgr, cv2.COLOR_BGR2RGB).astype(np.float64) / 255.0
den = denoise_tv_chambolle(float_rgb, weight=0.05, channel_axis=-1)
result = img_as_ubyte(den)                      # back to the uint8 contract

print(result.shape, result.dtype)               # (240, 320, 3) uint8

The total-variation denoiser smuggled into step 3 is a preview of Chapter 7, where denoising gets a proper treatment, and of the remarkable arc by which denoising later becomes the engine of generative models. For now the point is structural: four libraries, one array, three explicit conversions, zero surprises. The next section descends from ecosystem cartography to the most basic operational skill of all: getting images into and out of your program without losing data on the way.