PIL vs NumPy Images¶

Image Representation¶

1. PIL Image Object¶

from PIL import Image

img = Image.open('photo.jpg')
print(type(img))  # <class 'PIL.Image.Image'>
print(img.size)   # (width, height)
print(img.mode)   # 'RGB', 'RGBA', 'L', etc.

2. NumPy Array¶

import numpy as np

arr = np.array(img)
print(type(arr))  # <class 'numpy.ndarray'>
print(arr.shape)  # (height, width, channels)
print(arr.dtype)  # uint8

3. Key Difference¶

PIL: Format-aware object NumPy: Raw pixel data

When to Use Each¶

1. Use PIL For¶

# I/O operations
img = Image.open('input.jpg')
img.save('output.png')

# Format conversion
img_rgb = img.convert('RGB')

# Resizing
img_small = img.resize((100, 100))

# Rotation
img_rot = img.rotate(45)

2. Use NumPy For¶

# Pixel arithmetic
arr = np.array(img)
bright = np.clip(arr * 1.5, 0, 255).astype(np.uint8)

# Filtering
kernel = np.ones((5, 5)) / 25
filtered = convolve2d(arr[:,:,0], kernel)

# Batch processing
batch = np.array([arr1, arr2, arr3])  # (3, H, W, C)

3. Workflow¶

# Load with PIL
img = Image.open('photo.jpg')

# Convert to NumPy for processing
arr = np.array(img)
processed = arr * 0.8  # Darken

# Convert back to PIL for saving
result = Image.fromarray(processed.astype(np.uint8))
result.save('output.jpg')

Deep Learning¶

1. PyTorch Format¶

import torch

# PIL → NumPy → PyTorch
img = Image.open('photo.jpg')
arr = np.array(img) / 255.0  # Normalize
tensor = torch.from_numpy(arr).permute(2, 0, 1)  # (C, H, W)

2. Preprocessing¶

from torchvision import transforms

transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),  # PIL → Tensor
    transforms.Normalize(mean=[0.485, 0.456, 0.406],
                         std=[0.229, 0.224, 0.225])
])

tensor = transform(img)

3. Batch Loading¶

# DataLoader expects PIL or Tensor
from torch.utils.data import Dataset

class ImageDataset(Dataset):
    def __init__(self, image_paths):
        self.paths = image_paths

    def __getitem__(self, idx):
        img = Image.open(self.paths[idx])
        return transform(img)  # Returns Tensor