Image Classification
The image classification workflow and the reference results.
What you will learn
- Building an image classifier
- Understanding the limits of results on clean data
Classification gives one label to the whole image. Steps: collect balanced images for each class, split, train, and test on new images from different conditions.
Our actual result on the 8×8 digit images with 10-fold cross-validation (scikit-learn): SVM 0.984 and kNN 0.979. These are small, clean images, so do not expect the same numbers on real-world images.
Implementation steps with the tool
- Teachable Machine: an image project.
- Orange: the image classification project workflow (a ready-made .ows file).
Quick review question
Why might a model trained on images with a white background collapse when tested in the street?
