US 11,893,709 B2
Image quantization using machine learning
Mohammad Sadegh Norouzzadeh, Pittsburgh, PA (US); Renan Alfredo Rojas Gomez, Champaign, IL (US); Anh Nguyen, Auburn, AL (US); and Filipe J. Cabrita Condessa, Pittsburgh, PA (US)
Assigned to Robert Bosch GmbH
Filed by Robert Bosch GmbH, Stuttgart (DE)
Filed on Dec. 9, 2021, as Appl. No. 17/546,391.
Prior Publication US 2023/0186429 A1, Jun. 15, 2023
Int. Cl. G06T 3/40 (2006.01); G06T 7/90 (2017.01); G06N 3/08 (2023.01); H04N 1/64 (2006.01); G06T 9/00 (2006.01); G06N 3/084 (2023.01); G06N 3/045 (2023.01)
CPC G06T 3/4046 (2013.01) [G06N 3/08 (2013.01); G06N 3/084 (2013.01); G06T 7/90 (2017.01); G06T 9/002 (2013.01); H04N 1/644 (2013.01); G06N 3/045 (2023.01); G06T 2207/10028 (2013.01); G06T 2207/20081 (2013.01); G06T 2207/20084 (2013.01)] 20 Claims
OG exemplary drawing
 
1. A computer-implemented method for training a machine-learning image quantization system, the computer-implemented method comprising the following steps:
(i) receiving a plurality of input images from an image source, wherein each input image includes a plurality of pixels;
(ii) utilizing an image-to-image machine-learning model to assign each pixel a new pixel color;
(iii) utilizing a mixer model to map each new pixel color to one of a fixed number of colors to produce a quantized image corresponding to each input image;
(iv) feeding the input images to a pre-trained reference model to produce a first set of activations;
(v) feeding the quantized image to the pre-trained reference model to produce a second set of activations;
(vi) computing a loss function based on a comparison between the first set of activations and the second set of activations;
(vii) backpropagating the loss function into the image-to-image machine learning model and the mixer model; and
(viii) outputting a trained image-to-image machine learning model and a trained mixer model after repeating steps (ii)-(vii) until convergence.