US 12,003,746 B2
Joint forward and backward neural network optimization in image processing
Guan-Ming Su, Fremont, CA (US)
Assigned to DOLBY LABORATORIES LICENSING CORPORATION, San Francisco, CA (US)
Appl. No. 17/800,886
Filed by DOLBY LABORATORIES LICENSING CORPORATION, San Francisco, CA (US)
PCT Filed Feb. 17, 2021, PCT No. PCT/US2021/018407
§ 371(c)(1), (2) Date Aug. 18, 2022,
PCT Pub. No. WO2021/168001, PCT Pub. Date Aug. 26, 2021.
Claims priority of provisional application 62/978,638, filed on Feb. 19, 2020.
Claims priority of application No. 20158278 (EP), filed on Feb. 19, 2020.
Prior Publication US 2023/0084705 A1, Mar. 16, 2023
Int. Cl. H04N 19/436 (2014.01); H04N 19/119 (2014.01); H04N 19/136 (2014.01); H04N 19/186 (2014.01)
CPC H04N 19/436 (2014.11) [H04N 19/119 (2014.11); H04N 19/136 (2014.11); H04N 19/186 (2014.11)] 15 Claims
OG exemplary drawing
 
1. A computer-implemented method comprising:
receiving a source image of a source color grade and a destination image of a destination color grade as input to an end-to-end image mapping pipeline comprising a forward path and a backward path,
partitioning the forward path into multiple sub-nets represented by a plurality of sets of forward neural networks, with each of the multiple sub-nets in the forward path represented by a corresponding set of forward neural networks in the plurality of sets of forward neural networks,
partitioning the backward path into multiple sub-nets represented by a plurality of sets of backward neural networks, with each of the multiple sub-nets in the backward path represented by a corresponding set of backward neural networks in the plurality of sets of backward neural networks, wherein the forward path and the backward path are concatenated together;
applying the plurality of sets of forward neural networks in the forward path to forward reshape the source image of the source color grade to generate a forward reshaped image of the destination color grade;
applying the plurality of sets of backward neural networks in the backward path to backward reshape the forward reshaped image of the destination color grade to generate a backward reshaped image of the source color grade;
computing a joint neural network cost function specified for both the forward path and the backward path, wherein the joint neural network cost function comprises a forward cost portion that computes a first difference between the forward reshaped image and the destination image, wherein the joint neural network cost function further comprises a backward cost portion that computes a second difference between the backward reshaped image and the source image; and
determining operational parameters for the plurality of sets of forward neural networks and for the plurality of sets of backward neural networks by back propagation using the joint neural network cost function; wherein the plurality of sets of forward neural networks convert the source image of the source color grade to a sequence of successive forward reshaped images for a plurality of successively lower quality destination color grades, wherein the plurality of sets of backward neural networks convert a forward reshaped image last generated in the sequence of successive forward reshaped images in the forward path to a plurality of backward reshaped images for the plurality of destination color grades, and wherein the forward reshaped image last generated in the sequence of successive forward reshaped images in the forward path and the plurality of backward reshaped images are used in the joint neural cost function.