Module 02 Activity

Scenario

The images have to become arrays a model can consume, and every step of that discards something.

What you build

A loading pipeline that detects format from bytes, rejects with reasons, and matches the compression and resolution of production images.

Steps

1. Detect format from the file header rather than the extension. Find or construct one mismatched file and confirm it is handled.
2. Measure what your target resize destroys: take your smallest feature and report its size in pixels before and after.
3. Convert a sample to HSV and grayscale. Say which space suits your task and why.
4. Save one image at three JPEG qualities and measure the pixel error. Compare against the quality production images arrive at.
5. Write a loader that returns a reason for every rejection, and count rejections by kind over the whole folder.
6. Report the rejection rate per source.

Evidence to hand in

• The header-based detection with one mismatched file handled.
• Your smallest feature's size before and after resize.
• The colour space choice with its reason.
• The compression comparison against production quality.
• Rejection counts by kind and by source.

Review checklist

• Format comes from the bytes, not the filename.
• The smallest feature's post-resize size is measured, not assumed.
• Training compression matches production compression.
• Every rejection carries a reason.
• Rejection rates are reported per source, not in aggregate.

Mark complete