newlib: High-Performance Python Image Processing

newlib is a modular, production-grade image processing library designed to sit comfortably between OpenCV and modern ML frameworks. It prioritizes clean APIs, pipeline-based processing, and seamless interoperability.

🚀 Key Features

• Fluent Image API: Wrapper around NumPy arrays with method chaining (img.resize().to_gray()).
• Pipeline Engine: Lazy execution, serialization (YAML/JSON), and benchmarking.
• Machine Learning Ready:
  • Seamless to_torch() and from_torch() bridges.
  • Augmentations like ColorJitter, RandomRotate.
  • ML normalization tools.
• Developer Friendly:
  • imgforge CLI tool.
  • Visual debugging (show, compare).
  • Built-in sample data loader (get_sample("lenna")).
• Web-Ready IO: Automatic caching of images loaded from URLs.

📦 Installation

pip install -e .
# Optional dependencies
pip install -e .[ml]   # For PyTorch/TensorFlow support
pip install -e .[dev]  # For testing/development

⚡ Quick Start

Python API

from newlib.core.image import Image
from newlib.core.pipeline import Pipeline
from newlib.transforms import geometric, color, filtering
from newlib.utils.data import get_sample

# 1. Load an image (auto-downloads sample)
img = get_sample("lenna")

# 2. Method Chaining
processed = img.resize(256, 256).to_gray()
processed.save("output_fast.jpg")

# 3. Pipeline API (Recommended for reproducible workflows)
pipe = Pipeline([
    geometric.Resize(224, 224),
    filtering.GaussianBlur(kernel_size=5),
    color.ToGray()
])

result = pipe(img)
print(f"Execution time: {pipe.benchmark()}")

CLI Tool (imgforge)

Resize an image:

imgforge resize input.jpg 512 512 output.jpg

Inspect image metadata & quality:

imgforge inspect https://example.com/image.jpg

Run a Pipeline from YAML:

# pipeline.yaml
pipeline:
  - name: Resize
    params:
      width: 256
      height: 256
  - name: ToGray

imgforge pipeline pipeline.yaml input.jpg output.jpg

🛠️ Modules Overview

🤝 ML Integration

Stop fighting with NumPy/Tensor conversions.

from newlib.core.interop import to_torch

img = Image.open("dataset/001.jpg")

# Automatically converts to (C, H, W) float tensor
tensor = to_torch(img.data, device="cuda") 

🧪 Running Tests

python -m pytest tests/

📄 License

MIT License. See LICENSE for details.