7.9.36. Machine Readable Code Classification

Classify a 28x28 image as a QR code, a barcode or something else, on the MSPM0G5187 with a LeNet-5 CNN.

Example folder: examples/machine_readable_code_classification (GitHub)

7.9.36.1. Overview

Task: image classification (image_classification, target_module: vision), 3 classes
Application: tiny visual pre-classification of machine readable codes on low-memory, low-power edge devices. The goal is not to decode the code, only to classify its visual structure. At 28x28 many generated codes would not be scannable by a real decoder, which is acceptable here.
Device: MSPM0G5187 (config_MSPM0.yaml, the only config in the folder)
Model: Lenet5

7.9.36.2. The Dataset

A synthetic dataset similar in size and format to MNIST: 28x28, one channel, binary black and white PNG images. The default size is 3000 images per class, 9000 in total.

Class	How it is generated
`qr`	QR codes from the Python `qrcode` package (version 1, error correction L, box size 4, border 1) with a random alphanumeric payload
`barcode`	Code128 barcodes from the `python-barcode` package with the text under the barcode disabled, so the model learns the line structure and not printed text
`other`	Blank white, blank black, random binary noise, random line patterns and random block patterns, so the class is not a single trivial pattern

All images go through the same conversion: grayscale, nearest-neighbor resize to 28x28 (to keep hard edges), threshold at 127, pixels set to 0 or 255. The script uses a fixed seed (42), so the dataset is reproducible with the same package versions.

The config downloads https://software-dl.ti.com/C2000/esd/mcu_ai/datasets/machine_readable_code_classification.zip. To regenerate the data, install the dependencies and run the generator:

python -m pip install "qrcode[pil]" "python-barcode" pillow numpy
cd examples/machine_readable_code_classification
python generate_machine_readable_code_28x28.py

The dataset layout is one folder per class (qr, barcode, other). Point dataset.input_data_path at the generated zip or folder to use it.

7.9.36.3. Running the Example

Install Tiny ML ModelZoo first (Installation, Quickstart), then run from the tinyml-modelzoo folder:

./run_tinyml_modelzoo.sh examples/machine_readable_code_classification/config_MSPM0.yaml

See Running Examples for all launchers, for example:

run_tinyml_modelzoo.bat examples\machine_readable_code_classification\config_MSPM0.yaml

7.9.36.4. Configuration

common:
  target_module: vision
  task_type: image_classification
  target_device: MSPM0G5187
dataset:
  dataset_name: machine_readable_code_classification
  input_data_path: https://software-dl.ti.com/C2000/esd/mcu_ai/datasets/machine_readable_code_classification.zip
data_processing_feature_extraction:
  feature_extraction_name: MachineReadable_Default
training:
  model_name: Lenet5
  batch_size: 64
  training_epochs: 10
  num_gpus: 0
  quantization: 2
  learning_rate: 0.1
  weight_decay: 1e-5
testing: {}
compilation: {}

The preset MachineReadable_Default converts to grayscale, resizes to 28x28 and normalizes with mean 0.5 and scale 0.5. The image models enabled in 1.5.0 are Lenet5, MobileNetV1_28k_NPU and MobileNetV2_58k_NPU. For the image task type see Image Classification.

7.9.36.5. Results

The readme does not publish accuracy numbers. In the automated 1.5.0 config test run (test set of 900 images, 300 per class) the quantized model reached 100.00% accuracy and an AUC ROC of 1.000. The classes are visually very distinct and the data is synthetic, so expect lower accuracy on photographs of real codes.

7.9.36.6. Output Location

Results are written to ./data/projects/machine_readable_code_classification/run/<date-time>/Lenet5/. See Understanding the Run Output and Project Compilation Guide: Timeseries Classification Examples for deployment.