Granted Patent B2

US 12,488,556 · App. 18/172,774 · Granted Dec 2, 2025

Inventors: Rajath Elias Soans (Bengaluru, IN); Pradeep Nelahonne Shivamurthappa (Mahadevapura, IN); Kuladeep Marupalli (Bengaluru, IN); Alladi Ashok Kumar Senapati (Bangalore, IN); Ananya Paul (Katihar, IN)

Assignee: Samsung Electronics Co., Ltd.

G06V10/267G06V10/25G06V10/82

Abstract

Claims (68)

1 . A method performed by an electronic device for in-place transformation of an input frame for an on-device artificial intelligence (AI) model, the method comprising:

obtaining an input frame;

building at least one kernel independent of a scale of the input frame by passing input variables to the at least one kernel using preprocessor directives independent of the scale of the input frame;

inputting the input frame to the on-device AI model including the at least one kernel independent of the scale of the input frame; and

processing the input frame in the on-device AI model,

wherein the processing of the input frame comprises:

performing neural network operation, including convolution operation or pooling operation, for a region identified as at least one region of interest (RoI) within the input frame;

copying the region identified as a non-RoI representing the remaining region of the at least one RoI within the input frame; and

generating an output image based on the performing of the neural network operation for the region identified as the at least one RoI and the copying of the region identified as the non-RoI.

2 . The method of claim 1 , wherein the processing of the input frame comprises:

identifying at least one region of interest (the at least one RoI) including at least one pixel included in the input frame.

3 . The method of claim 2 , wherein the identifying of the at least one RoI comprises:

obtaining a remaining region after removing a padding area in the input frame as RoI; or

obtaining at least one region designated by a user as at least one RoI based on a user input including information on the region designated by the user.

4 . The method of claim 2 , wherein the identifying of the at least one RoI comprises:

obtaining a residual frame representing a difference between the input frame included in successive frames and a previous frame of the input frame as the at least one RoI.

5 . The method of claim 1 , wherein the generating of the output image comprises:

obtaining the at least one RoI by performing the neural network operation for the region identified as the at least one RoI within the input frame; and

returning the at least one Rol to expected location in an output frame taking into account spatial changes.

6 . The method of claim 5 , wherein the obtaining of a remaining region after removing a padding area in the input frame as RoI comprises:

identifying the padding area within the input frame by traversing the input frame in a row-wise and column-wise;

obtaining information associated with the scale of an image area, the scale of the input frame, the scale of the padding area and the input frame; and

removing the padding area within the input frame based on the information.

7 . An electronic device for in-place transformation of an input image for an on-device artificial intelligence (AI) model by the electronic device, the electronic device comprises:

a memory, comprising one or more storage media, storing instructions; and

at least one processor communicatively coupled to the memory,

wherein the instructions, when executed by the at least one processor individually or collectively, cause the electronic device to:

obtain an input frame,

build at least one kernel independent of a scale of the input frame by passing input variables to the at least one kernel using preprocessor directives independent of the scale of the input frame,

input the input frame to the on-device AI model including the at least one kernel independent of the scale of the input frame, and

process the input frame in the on-device AI model,

wherein the at least one processor is further configured to:

perform neural network operation, including convolution operation or pooling operation, for a region identified as at least one region of interest (RoI) within the input frame,

copy the region identified as a non-RoI representing the remaining region of the at least one Rol within the input frame, and

generate an output image based on the performing of the neural network operation for the region identified as the at least one RoI and the copying of the region identified as the non-RoI.

8 . The electronic device of claim 7 , wherein the at least one processor is further configured to:

identify the at least one RoI including at least one pixel included in the input frame.

9 . The electronic device of claim 8 , wherein the at least one processor is further configured to:

obtain a remaining region after removing a padding area in the input frame as Rol, or

obtain at least one region designated by a user as the at least one RoI based on a user input including information on the region designated by the user.

10 . The electronic device of claim 8 , wherein the at least one processor is further configured to:

obtain a residual frame representing a difference between the input frame included in successive frames and a previous frame of the input frame as the at least one RoI.

11 . The electronic device of claim 7 , wherein the at least one processor is further configured to:

obtain the at least one RoI by performing the neural network operation for the region identified as the at least one RoI within the input frame, and

return the at least one RoI to expected location in an output frame taking into account spatial changes.

12 . The electronic device of claim 11 , wherein the at least one processor is further configured to:

identify a padding area within the input frame by traversing the input frame in a row-wise and column-wise,

obtain information associated with the scale of an image area, the scale of the input frame, the scale of the padding area and the input frame, and

remove the padding area within the input frame based on the information.

13 . One or more non-transitory machine-readable media storing media storing one or more computer programs including computer-executable instructions that, when executed at least one processor of an electronic device individually or collectively cause the electronic device to perform operations, the operations comprising to:

obtaining an input frame,

building at least one kernel independent of the a scale of the input frame by passing input variables to the at least one kernel using preprocessor directives independent of the scale of the input frame,

inputting the input frame to an on-device artificial intelligence (AI) model including the at least one kernel independent of the scale of the input frame, and

processing the input frame in the on-device AI model,

wherein the operations further comprising:

performing neural network operation, including convolution operation or pooling operation, for a region identified as at least one region of interest (RoI) within the input frame,

copying the region identified as a non-RoI representing the remaining region of at least one RoI within the input frame, and

generating an output image based on the performing of the neural network operation for the region identified as the at least one RoI and the copying of the region identified as the non-RoI.

14 . The one or more non-transitory machine-readable media of claim 13 , wherein the operations further comprising:

identifying the at least one RoI including at least one pixel included in the input frame.

15 . The one or more non-transitory machine-readable media of claim 13 , wherein the operations further comprising:

obtaining a remaining region after removing the a padding area in the input frame as RoI, or

obtaining at least one region designated by a user as the at least one RoI based on a user input including information on the region designated by the user.

16 . The one or more non-transitory machine-readable media of claim 13 , wherein the operations further comprising:

obtaining a residual frame representing a difference between the input frame included in successive frames and a previous frame of the input frame as at least one region of interest (RoI).

17 . The one or more non-transitory machine-readable media of claim 13 , wherein the operations further comprising:

obtaining at least one region of interest (RoI) by performing a neural network operation for a region identified as the at least one RoI within the input frame, and

returning the at least one RoI to expected location in an output frame taking into account spatial changes.

Assignments (1)

Priority Claims (1)

Continuity (2)

References Cited (21)