Kamlesh21/preprocessing-lab

Preprocessing Lab

A teaching app about one idea: a model only works on the input it was trained on. One photo goes through preprocessing steps into ResNet50 (classify a pet) or YOLO11s-seg (find, outline and count objects). Every step shows its image, the exact Python code that ran, and why it exists.

• —Preparation steps are on: BGR→RGB, resize, crop or letterbox, normalize, channels first. Skip one and watch the model break (skip Normalize: the boxer becomes a "spotlight").
• —Break it switches do what the real world does to photos: rotate, dark, blur, noise, low resolution.
• —Fix it switches are classic enhancement and restoration: gamma, histogram equalization, CLAHE, denoising, median filter, sharpening, straightening. Some bring the answer back; some can't (lost detail is gone for good).

Keys: ← → walk the pipeline, 1–6 break-it switches, H histogram, E explain & code.

Source, notebook and tests: https://github.com/JustKamlesH21/image-preprocessing-demo Images: Oxford-IIIT Pet (CC BY-SA 4.0) and COCO 2017 (CC BY 4.0 annotations, Flickr image licenses). Models: ResNet50 (torchvision, BSD-3-Clause), YOLO11s-seg (Ultralytics, AGPL-3.0).