Equivalent for convert np.array input from PILImage to Tensor input in Java

I have successfully converted the emotion_ferplux.onnx model to a Tensorflow protobuf file and am trying to execute the model in java. In the python examples we have a code snippet:

input_shape = (1, 1, 64, 64)
img = Image.open(image_path)
img = img.resize((64, 64), Image.ANTIALIAS)
img_data = np.array(img) //Needs a dtype=float32
img_data = np.resize(img_data, input_shape)

to create the [1,1,64,64] input for the onnx model. From what I've been able to find over on the Java side in Tensorflow, there's something like this code snippet mentioned in the link here

[https://github.com/tensorflow/tensorflow/blob/master/tensorflow/java/src/main/java/org/tensorflow/examples/LabelImage.java]

      final Output<String> input = b.constant("input", imageBytes);
      final Output<Float> output =
          b.div(
              b.sub(
                  b.resizeBilinear(
                      b.expandDims(
                          b.cast(b.decodeJpeg(input, 3), Float.class),
                          b.constant("make_batch", 0)),
                      b.constant("size", new int[] {H, W})),
                  b.constant("mean", mean)),
              b.constant("scale", scale));

However this produces a [?][224][224][3]

I've been fiddling around to change the channel to use a value of 1 and also the dimensions to match the 64x64 input from the onnx model, however, when I inspect the actual values I'll be sending into the Tensorflow pb model over in java, they are very different.

I'm led to believe I have to create the input similar to how PILImage and Numpy are representing the image. I've also stumbled across some code from the org.bytedeco libraries, but haven't been able to piece together the equivalent.

Does anyone have any pointers on how to go about this?

Start with the Python preprocessing snippet in the issue and TensorFlow's Java example, LabelImage.java. Compare how each path loads, resizes, channels, reshapes, and types the image input. Done means documenting or demonstrating an agreed Java equivalent that produces the model's [1,1,64,64] float32 input.