7.9.38. Image Classification Examples Guide

In 1.4.0 this page was a generic image classification tutorial that pointed to an examples/image_classification folder. That folder does not exist in 1.5.0. This page is now a short guide to the three image classification examples and the settings they share. Each example has its own page with the full configuration. For the task type itself see Image Classification.

7.9.38.1. Image Examples in 1.5.0

All use task_type: image_classification with target_module: vision on the MSPM0G5187 and are run from the tinyml-modelzoo folder with ./run_tinyml_modelzoo.sh examples/<folder>/<config>.yaml (Windows: see Running Examples).

Config	Model	Input	Classes	Page
`MNIST_image_classification/config_image_classification_mnist.yaml`	`Lenet5`	28x28 grayscale	10	MNIST Image Classification
`machine_readable_code_classification/config_MSPM0.yaml`	`Lenet5`	28x28 binary	3	Machine Readable Code Classification
`coffee_bean_classification/config_MSPM0.yaml`	`MobileNetV1_28k_NPU`	128x128 RGB	4	Coffee Bean Classification

7.9.38.2. Dataset Format

Image datasets use the classification layout: one folder per class under classes/ containing the images (PNG or similar).

dataset.zip/
|-- classes/
      |-- class_a/
      |     |-- img_001.png
      |-- class_b/
            |-- img_001.png

The training, validation and test split is made automatically.

7.9.38.3. Feature Extraction Presets

Image preprocessing is selected with feature_extraction_name. The presets used by the examples are:

Preset	Preprocessing
`Mnist_Default`	Grayscale, resize to 28x28, mean 0.1307, scale 0.3081
`MachineReadable_Default`	Grayscale, resize to 28x28, mean 0.5, scale 0.5
`CoffeeBean_Default`	RGB, resize to 128x128, ImageNet mean and scale
`CoffeeBean_Augmentation_Default`	As `CoffeeBean_Default` plus random flips, rotation and color jitter

Image size drives memory use: the input tensor is height x width x channels and is the largest early activation. Small inputs (28x28) fit the smallest devices, while 128x128 RGB images need an NPU class model.

7.9.38.4. Available Models

The image models enabled in 1.5.0 are:

Model	Description
`Lenet5`	Classic LeNet-5 CNN, about 60,000 parameters according to the MNIST readme
`MobileNetV1_28k_NPU`	NPU-compliant MobileNetV1-inspired tiny model
`MobileNetV2_58k_NPU`	NPU-compliant MobileNetV2-inspired tiny model

Quantization is set with quantization: 2 in all three examples.

7.9.38.5. Results

From the automated 1.5.0 config test run: MNIST 98.93% test accuracy, machine readable code 100.00% test accuracy (synthetic data), coffee bean 99.38% test accuracy on a small test set. See the example pages for details.

7.9.38.6. Limitations

These are tiny models for small, simple images. High resolution images, many classes or fine-grained detail need a larger model than a microcontroller NPU can run. For deployment see Project Compilation Guide: Timeseries Classification Examples.