Process and copy a sample image into the Alif E8 CMSIS project

Set up Python image tools

You’ll convert a handwritten digit image into a C header that the firmware can include at build time.

To start, create a Python virtual environment for image preprocessing:

    

        
        

  cd ~/mnist_alif
  python3 -m venv venv_image_prep
  source venv_image_prep/bin/activate
  python -m pip install --upgrade pip
  python -m pip install numpy pillow
  

    

    

        
        

  cd ~\mnist_alif
  py -m venv venv_image_prep
  .\venv_image_prep\Scripts\Activate.ps1
  python -m pip install --upgrade pip
  python -m pip install numpy pillow
  

    

Save your image

Create a directory for the input image and preprocessing script:

    

        
        

mkdir -p ~/mnist_alif/image
cd ~/mnist_alif/image
  

    

    

        
        

New-Item -ItemType Directory -Force -Path ~\mnist_alif\image
cd ~\mnist_alif\image
  

    

Place a PNG or JPEG image of a handwritten digit in the directory and name it mnist_image.jpg.

Download the preprocessing script

The MNIST model expects one grayscale 28 × 28 image. You’ll use a preprocessing script to resize the image, convert it to grayscale, scale pixel values to the range 0 to 127, and write the result to input_mnist.h.

Download prepare_mnist_image.py into the image directory:

    

        
        
cd ~/mnist_alif/image
curl -o prepare_mnist_image.py https://raw.githubusercontent.com/arm-education/alif-ethos-u85-npu-mnist/main/prepare_mnist_image.py

    

Open prepare_mnist_image.py and inspect.

The script defines command-line arguments, checks that the input image exists, then converts the image to 28 × 28 grayscale pixels. It also handles image inversion so black-on-white images can be converted to the white-on-black style used by MNIST.

The key conversion step scales each pixel into a value within the non-negative int8 range (0 to 127) and flattens the output into a one-dimensional array:

    

        
        
pixels = np.clip(np.rint(pixels * 127.0 / 255.0), 0, 127).astype(np.int8)
flat = pixels.reshape(-1)

    

The generated header will contain a total of 784 values, one for each pixel in the 28 × 28 input image.

Generate the input header

Run the preprocessing script:

    

        
        

cd ~/mnist_alif/image
python prepare_mnist_image.py mnist_image.jpg --output input_mnist.h
  

    

    

        
        

cd ~\mnist_alif\image
python .\prepare_mnist_image.py .\mnist_image.jpg --output .\input_mnist.h
  

    

Verify the generated file:

    

        
        

ls -lh input_mnist.h
head -n 8 input_mnist.h
  

    

    

        
        

Get-Item .\input_mnist.h
Get-Content .\input_mnist.h -TotalCount 8
  

    

Open input_mnist.h and look at the generated input_mnist array. This is the numeric pixel data that main.cpp passes into the ExecuTorch runner.

Copy the header into the firmware project

Copy input_mnist.h into the application assets directory:

    

        
        

cp ~/mnist_alif/image/input_mnist.h ~/mnist_alif/alif_vscode-template/mnist_executorch/assets/
  

    

    

        
        

Copy-Item "$HOME\mnist_alif\image\input_mnist.h" "$HOME\mnist_alif\alif_vscode-template\mnist_executorch\assets\"
  

    

Verify that both firmware assets are present:

    

        
        

ls -lh ~/mnist_alif/alif_vscode-template/mnist_executorch/assets/
  

    

    

        
        

Get-ChildItem "$HOME\mnist_alif\alif_vscode-template\mnist_executorch\assets"
  

    

The output is similar to:

    

        
        input_mnist.h
mnist_model_data.h

        
    

What you’ve accomplished and what’s next

You’ve now converted a test digit image into input_mnist.h and copied it into the firmware project.

Next, you’ll build and flash the application to the Alif Ensemble E8 DevKit.