How to reduce processing time of TensorImageUtils.bitmapToFloat32Tensor function

After converting the bitmap generated by Android into a Tensor through the TensorImageUtils.bitmapToFloat32Tensor function, we have been receiving the result value through inference.

However, the processing time of that function is about 2 seconds, so it takes a lot of time to get the full result.

In bitmap, an image is drawn in real time, and each time a bitmap is obtained and inferred.
The result is delayed compared to the real-time video, so matching is not possible.

@IvanKobzarev Wondering if you may have some thoughts on this. Thank you!