# Classifier

# Introduction

The task of identifying what an image represents is called image classification. An image classifier is a machine learning model trained to recognize various classes of images. When you give it an image, it responds with a category label for that image. For example, you may train a model to recognize images representing different types of animals like bears, lions, zebras and okapis. An example of image classification

# Using existing image classification model for prediction

EdgeAI Studio provides 50+ pre-built models that you can use right away for prediction or integrating with other applications. Here is a pipeline example using a pre-built model for animals classification.

An example of zebra prediction An example of okapi prediction

TIP

For details of the pre-built models and ready-to-use pipelines, please refer to the pre-built model section here (opens new window).

# Training your own image classifier

You can train an image classifier directly on EdgeAI Studio (opens new window) by showing samples of images you have labeled for each class. After the image classifier finishes training, you can perform a prediction and assess its accuracy by using the Detect block.

Datasets should be prepared and connected to the Train block like this.

There are some differences in capabilities in each engine in the Train block.

Engine Capabilities/Limitations
Azure Custom Vision Able to train 2 or more classes. Each class should contain 5 or more images. An Azure subscription is required and the training key should be configured in Settings
Apple M1/M1 Pro/M1 Max Able to train up to 2 classes with fixed name a1 and a2. Each class should contain 5 or more images. The model name must be called as test

To run a training, the details of training an image classifier on EdgeAI Studio is as follows:

  1. Expand the Input category from the menu
  2. Drag an input block to the canvas as the source of your input files
  3. Drag and drop the file to the input block (e.g. Drag and drop a video file to a "Video File" input block or an image file to the "Image File" input block)
  4. Expand the Train category from the menu
  5. Drag the "Dataset" block to the canvas
  6. Click +1 or the record button in "Dataset" block
  7. Input the label name in "Dataset" block
  8. Repeat Step 1-7 for another class of the dataset
  9. Expand the Train category from the menu
  10. Drag the "Train" block to the canvas
  11. Connect the "Dataset" blocks to "Train" block
  12. Select the Engine and Type from the dropdown menu
  13. Select "Classification" as the model type
  14. Input your model name in the Name textbox
  15. Update the number of training epochs if Apple M1/M1 Pro/M1 Max engine is used (default value is 50)
  16. Click the play button to start training

Image Classifier

TIP

Only 10 or 50 epochs are supported in the Apple M1/M1 Pro/M1 Max engine. For details on how to configure your engine, please refer to the description here (opens new window).

Last updated: 11/16/2021, 1:41:06 PM