Plant Disease Classifier with a TensorFlow Lite Model

Image classifiers learn from thousands of labeled photos. You train a small model on a public leaf-disease dataset on your computer or in a free notebook, export it as a .tflite file, and the Pi runs it locally with LiteRT in a fraction of a second.

Mission steps

1. Train a model

   Train an image classifier on a public dataset such as PlantVillage (or your own labeled leaf photos) and export it to TensorFlow Lite with a labels.txt file.

2. Install the runtime

   In a venv created with --system-site-packages, run pip install ai-edge-litert numpy.

3. Copy the files

   Put plant_model.tflite and labels.txt next to the script on the Pi.

4. Classify a leaf

   Fill the frame with one leaf on a plain background in daylight and run the script.

5. Check the result

   Treat low-confidence answers as 'unknown' and confirm any disease with a local extension service.

Code

import numpy as np
from picamera2 import Picamera2
from ai_edge_litert.interpreter import Interpreter

labels = [l.strip() for l in open('labels.txt')]
interp = Interpreter(model_path='plant_model.tflite')
interp.allocate_tensors()
inp, out = interp.get_input_details()[0], interp.get_output_details()[0]
h, w = inp['shape'][1:3]

cam = Picamera2()
cam.configure(cam.create_still_configuration(main={'size': (int(w), int(h)), 'format': 'RGB888'}))
cam.start()
img = cam.capture_array()[..., ::-1]               # BGR -> RGB

x = np.expand_dims(img, 0)
if inp['dtype'] == np.float32:
    x = x.astype(np.float32) / 127.5 - 1.0         # match how your model was trained
else:
    x = x.astype(inp['dtype'])
interp.set_tensor(inp['index'], x)
interp.invoke()
scores = interp.get_tensor(out['index'])[0].astype(float)
best = int(np.argmax(scores))
print(f"{labels[best]}  ({scores[best] / scores.sum():.0%})")

Tip: test every script on the bench before installing it in a moving machine.

Questions