Video Summary - Getting Started with Orange 15: Image Analytics

Main ideas and concepts

The video continues from a prior lesson about transforming images into vectors (embeddings) and using them for clustering.
It demonstrates that embeddings can also be used for image classification—specifically, distinguishing different flower types (e.g., tulips vs. orchids).

Create a top-level folder named flowers.
Inside it, create 9 subfolders, each corresponding to a flower class label (e.g., daisies, roses, tulips, etc.).
Each subfolder contains images for that class.

Use import images to load the main flowers folder.
Orange automatically treats each subfolder as an image class label.
Verify that the expected number of images is loaded (the video reports 82 images across 9 classes).
Use image viewer to visually confirm the images loaded correctly.

Pass the images to image embedding.
The widget returns a vector representation for each image.
Inspect the results in a data table:
- Each image keeps its class label
- Each image has ~248 additional features derived from a deep network embedding

Use cross-validation via test and score.
Choose a classifier suitable for the task:
- Logistic regression is selected as the learner.
Connect:
- test and score to image embedding (so it can train/test on embeddings).
Run 10-fold cross validation.
Review performance metrics:
- AUC (Area Under the ROC Curve) is reported as “really high”
- Classification accuracy is described as “not too bad”

Add confusion matrix connected to test and score.
Interpret the confusion matrix:
- Diagonal cells = correct predictions
- Off-diagonal cells = misclassifications
Inspect specific misclassified categories. Example errors mentioned include:
- The model predicted callus when the true class was tulips
- The model predicted roses for cases that were actually callus (the label in the subtitles appears possibly misspelled; exact categories may be unclear)
To investigate why:
- Select a misclassification cell
- Connect image viewer to confusion matrix
- Observe that the “wrong” predictions often still show meaningful visual overlap (e.g., bouquets may contain mixed flower types, so the model is not entirely “confused”)

Embeddings are versatile: beyond clustering, they support effective classification.
Cross-validation helps estimate predictive performance reliably.
Confusion matrices clarify where and how the model fails.
Misclassifications can sometimes be explained by data ambiguity (e.g., images containing mixed flower types).