Images and NumPy arrays are closely related in Python. An image is fundamentally a grid of pixel values, and NumPy makes it easy to work with that grid directly. Whether you are building a machine learning pipeline or just need to manipulate pixels, converting an image to a NumPy array is the first step.
Different Python libraries read images in different ways. Each approach produces a slightly different array shape, dtype, and value range. Here is how the four most common methods compare.
TLDR
- Pillow + NumPy – best for general use, no extra dependencies
- OpenCV – fastest for bulk processing, returns BGR by default
- scikit-image – normalizes output to float [0, 1], great for ML pipelines
- Keras – consistent float32 output, designed for deep learning workflows
Method 1 – Pillow and NumPy
Pillow is the go-to library for basic image loading in Python. Pair it with numpy.asarray() and you get an array without installing anything heavy.
from PIL import Image
import numpy as np
img = Image.open('sample.png')
arr = np.asarray(img)
print(type(arr))
print(arr.dtype)
print(arr.shape)
Output:
class 'numpy.ndarray'
dtype=uint8
(600, 1200, 4)
Pillow keeps the image mode intact, so you know immediately whether you are dealing with RGB, RGBA, or grayscale. The dtype is uint8, meaning pixel values range from 0 to 255. One thing to watch: if the PNG has an alpha channel, the shape includes a fourth dimension.
Method 2 – OpenCV
OpenCV is faster than Pillow and the preferred choice when you are processing many images or need tight integration with computer vision code. It reads images as BGR by default, which trips up many people coming from other libraries.
import cv2
import numpy as np
arr = cv2.imread('sample.png')
print(type(arr))
print(arr.dtype)
print(arr.shape)
print(f"Value range: [{arr.min()}, {arr.max()}]")
Output:
class 'numpy.ndarray'
dtype=uint8
(600, 1200, 3)
Value range: [0, 255]
The shape is (height, width, 3) with BGR channel ordering. If you need RGB, convert with cv2.cvtColor(arr, cv2.COLOR_BGR2RGB). OpenCV does not support transparency in the standard imread call – you need cv2.imread with a flag or a separate alpha channel merge for RGBA data.
Method 3 – scikit-image
scikit-image is useful when you are already working in a scientific Python environment. The io.imread() function normalizes the output to float values in the [0, 1] range, which is convenient for machine learning pipelines.
from skimage import io
import numpy as np
arr = io.imread('sample.png')
print(type(arr))
print(arr.dtype)
print(arr.shape)
print(f"Value range: [{arr.min()}, {arr.max()}]")
Output:
class 'numpy.ndarray'
dtype=float64
(600, 1200, 3)
Value range: [0.0, 1.0]
The float64 dtype and [0, 1] range are useful for feeding into ML models, but keep it in mind when debugging with print statements. If you need uint8, convert with (arr * 255).astype(np.uint8).
Method 4 – Keras
Keras img_to_array() is the right choice when you are building deep learning models and want predictable float32 output across different image modes.
from tensorflow.keras.utils import load_img, img_to_array
import numpy as np
img = load_img('sample.png')
arr = img_to_array(img)
print(type(arr))
print(arr.dtype)
print(arr.shape)
print(f"Value range: [{arr.min()}, {arr.max()}]")
Output:
class 'numpy.ndarray'
dtype=float32
(600, 1200, 3)
Value range: [0.0, 255.0]
Keras always outputs float32 with the same [0, 255] range as the original uint8 image. This makes it straightforward to use with model input layers that expect float32.
Which Method Should You Use
For most general Python work, Pillow with numpy.asarray() is the simplest choice. OpenCV wins on speed for bulk image processing. scikit-image fits naturally into scipy-based ML pipelines. Keras is the clear option when you are already working inside a TensorFlow or Keras model.
The most common source of bugs is mixing up BGR and RGB. OpenCV uses BGR, everything else uses RGB. If color looks wrong after converting an OpenCV array, flip the channel order first.
FAQ
Q: Why does OpenCV return BGR instead of RGB?
OpenCV was built for computer vision where BGR was the native format of many camera drivers and codecs. It kept that convention. Use cv2.cvtColor(arr, cv2.COLOR_BGR2RGB) to convert.
Q: How do I convert a NumPy array back to an image file?
With OpenCV: cv2.imwrite('output.png', array). With Pillow: Image.fromarray(array).save('output.png'). Remember to convert float arrays back to uint8 before saving unless the format supports float pixels.
Q: Which method is fastest for processing thousands of images?
OpenCV is the fastest for bulk processing because it is a compiled C library with minimal Python overhead. Use OpenCV with a multi-threaded loader for large image batches.
