Turn labelled photo folders into a reusable image sorter

A growing photo collection becomes easier to use when each file has a useful category. Orange can learn those categories from your examples. This workflow produces a CSV of image filenames, paths and predicted labels, plus a reusable model for later batches.[10][8][9]

This is a documentation-based walkthrough, checked against Orange’s current widget guides on 19 October 2026; we have not trained or exported this example. The official desktop download is Orange 3.40.0 for Windows and macOS. Install Image Analytics through the add-on manager in the Options menu.[1]

Give the sorter a vocabulary

For an illustrative neighbourhood photo project, use three categories: bicycles, benches and plant pots. Make a Training folder with one subfolder for each category. Put clearly labelled examples in those subfolders, using photographs of several different objects. Keep a separate Review folder containing additional examples you can recognise, and a New Batch folder for the collection you want labelled.

Use a task that actually fits those labels. A close-up of a bicycle belongs readily in Bicycles; a whole street containing all three objects needs a decision about which subject matters. Resolve that choice while preparing the collection, so the category means something consistent.

On Orange’s canvas, add Import Images and choose Training. Orange creates one row per image and uses subfolder names as class values. Connect its Data output to a Data Table and check that the filenames and categories match the collection you prepared. If you change the folder contents, use Reload.[2]

Build the model from image features

Connect Import Images to Image Embedding. Choose SqueezeNet, then apply the embedding if it has not started automatically. This option computes on your computer; the other documented embedders use a remote server. The output is a table of numerical image descriptors alongside the image information. Check the completed and skipped-image counts before continuing.[3]

The neural network supplies a visual representation. A second, much smaller learning step associates those representations with your labels. Add Logistic Regression and connect Image Embedding’s Embeddings output to its Data input. Leave the default regularisation settings for this first pass and apply. Use its Model output in the next step: that is the trained classifier.[3][4]

Run a separate batch through it

Add a second Import Images widget for Review, followed by a second Image Embedding widget. Select SqueezeNet again. The training and prediction branches must use the same representation. Add Predictions, connect this second Embeddings output to Data, and connect Logistic Regression’s Model output to Predictors.[3][5]

Predictions displays the category and optional class probabilities. Connect its Predictions output to Image Viewer and select the predicted-category column as the image title. Compare the pictures and labels. If benches repeatedly become plant pots, add clearer bench examples to Training, reload and retrain, then examine additional review images before proceeding.[5][6]

Once the review is satisfactory, point the second Import Images widget at New Batch and run its embedding. The trained model now supplies categories for those files. It chooses from the vocabulary you gave it, so photographs outside that task need their own review rather than an assumed new category.[10]

Export decisions you can actually use

Connect the full Predictions output to Select Columns. Keep the image name, image path, predicted category and any probabilities you want to inspect; move the numerical embedding columns to the ignored area. The prediction is added as a meta attribute, which can remain in the exported table. Use Data Table to inspect this reduced output.[5][7]

Connect Select Columns to Save Data and use Save as to create a CSV. For a conventional spreadsheet with one header row, leave Add type annotations to header unticked. Open the saved file and check that its row count matches the successfully embedded New Batch images. Confirm that a filename and its predicted label stay together when you sort the spreadsheet.[8]

Finally, connect Logistic Regression’s Model output to Save Model and save the classifier as a .pkcls file. Keep the labelled training folders and the Orange workflow too. For a later batch, Load Model can supply the predictor; the new photographs still pass through Image Embedding with SqueezeNet. The useful asset is now a repeatable route from your categories to a labelled collection.[9][10]

Report an error or suggest a correction