How would you build an image classifier with dirty data?

## Scenario You are asked to build an **image classification** model (single-label, multi-class) for a product team. The image dataset is known to be **dirty** (e.g., corrupted files, wrong labels, duplicates, irrelevant images, inconsistent formats). Compared with text classification, image inputs often require additional preprocessing and validation. ## Tasks 1. **Design the end-to-end approach** to train and evaluate an image classifier. 2. Describe how you would **measure the “dirty rate”** of the image data (what counts as dirty, how to estimate it reliably). 3. **Follow-up:** After training a baseline, you discover performance is worse than expected. Explain how you would **identify data problems** (not just model problems) and propose concrete **data and pipeline improvements**. ## Constraints / clarifications (you may state assumptions) - You may assume typical real-world constraints: limited labeling budget, heterogeneous image sources, and the need for reproducible training. - You should specify what metrics you would use (overall and per-class) and how you would validate improvements.