preprocessor_config.json

{
  "image_processing": {
    "size": {
      "width": 320,
      "height": 320
    },
    "resample": "bilinear",
    "normalize": true,
    "mode": "densenet_preprocess_input",
    "channel_order": "RGB",
    "mean": [123.675, 116.28, 103.53],
    "std": [58.395, 57.12, 57.375],
    "scale": 1.0,
    "description": "Use tf.keras.applications.densenet.preprocess_input for automatic preprocessing. For manual TF-Lite: RGB->BGR, subtract ImageNet mean. Note: preprocessor_config.json reserves 'BGR' for DenseNet ImageNet references, but densenet.preprocess_input() works on RGB input and handles the internal conversion."
  },
  "input_name": "input_1",
  "output_name": "output_1",
  "input_shape": [1, 320, 320, 3],
  "output_shape": [1, 8]
}