Granted Patent B2

US 12,561,967 · App. 18/267,095 · Granted Feb 24, 2026

Inventors: Lok Won Kim (Seongnam-si, KR); Shin Woo Jeon (Seoul, KR)

Assignee: DEEPX CO., LTD.

G06V10/82G06V10/87

Abstract

Claims (68)

1 . An image processing method comprising:

receiving a first image including at least one object;

receiving, from a head-mounted display (HMD) device, data indicative of a gaze of a user viewing the first image;

determining a region of the at least one object by inputting the first image to a first model based on an artificial neural network trained to classify an object corresponding to a predetermined category;

determining, based on the received gaze data, a region in which the gaze of the user stays relatively long among regions of each of the at least one object as a region of interest (ROI) by using the first model;

classifying a category of each of the at least one object by re-inputting each determined ROI to the first model;

determining at least one second model among a plurality of second models based on an artificial neural network trained to output an image to which a category-specific image processing corresponding to an object category is applied by inputting the first image to the at least one second model, the determined at least one second model corresponding to the category of each of the at least one object; and

outputting a second image having improved quality by using the at least one second model such that an ROI-specific image processing corresponding to an object category of each determined ROI is applied to the first image.

2 . The image processing method of claim 1 ,

wherein the at least one object includes an object belonging to one category selected from among a plurality of categories.

3 . The image processing method of claim 1 ,

wherein the first model includes an input layer and an output layer comprising a plurality of nodes, and

wherein the number of the plurality of second models corresponds to the number of the plurality of nodes of the output layer of the first model.

4 . The image processing method of claim 1 , wherein the at least one second model includes at least one of a denoising model, a deblurring model, an edge enhancement model, a demosaicing model, a color tone enhancing model, a white balancing model, a super resolution model, a wide dynamic range model, a high dynamic range model, and a decompression model.

5 . The image processing method of claim 1 ,

wherein the at least one second model includes an ensemble model in which at least two second models selected from among the plurality of second models are combined and are connected in parallel or in series, and

wherein the second image is output from one of the parallel connection and the series connection.

6 . An image processing method comprising:

receiving a first image including at least one object, the at least one object including an object belonging to one category selected from among a plurality of categories;

receiving, from a head-mounted display (HMD) device, data indicative of a gaze of a user viewing the first image;

determining a region of the at least one object by inputting the first image to a first model based on an artificial neural network trained to classify an object corresponding to a predetermined category;

determining, based on the received gaze data, a region in which the gaze of the user stays relatively long among regions of each of the at least one object as a region of interest (ROI) by using the first model;

classifying a category of each of the at least one object by re-inputting each determined ROI to the first model;

applying a parameter among a plurality of parameters predetermined for each of the plurality of categories to a second model based on an artificial neural network trained to output an image to which a category-specific image processing corresponding to an object category is applied by inputting the first image to the second model, the applied parameter corresponding to the category of each of the at least one object; and

outputting a second image having improved quality by using the second model such that an ROI-specific image processing corresponding to an object category of each determined ROI is applied to the first image.

7 . A neural processing unit comprising:

an internal memory configured to store a first model, at least one second model, and at least part of data of a first image, the first image including at least one object and a region of interest (ROI) determined based on gaze data acquired from a head-mounted display (HMD) device, the gaze data indicative of a gaze of a user viewing the first image;

a processing element (PE) array configured to access the internal memory and to process convolution of the first model and the at least one second model; and

a controller operatively coupled to the internal memory and the processing element,

wherein the first model includes an artificial neural network-based model trained to determine a region of the at least one object by inputting the first image to the first model and to classify an object corresponding to a predetermined category,

wherein the at least one second model includes a plurality of artificial neural network-based models respectively trained to output an image to which a category-specific image processing corresponding to an object category is applied by inputting the first image to the at least one second model, and

wherein the controller is configured to

induce the PE array to determine, as the ROI, a region in which the gaze of the user stays relatively long among regions of each of the at least one object, the ROI determined based on the gaze data by using the first model,

re-input each determined ROI to the first model to classify a category of each of the at least one object, and

output a second image having improved quality by using a second model corresponding to the category of each of the at least one object among the plurality of second models such that an ROI-specific image processing corresponding to an object category of each determined ROI is applied to the first image.

8 . The neural processing unit of claim 7 , further comprising:

a main memory configured to store the first model and the at least one second model,

wherein the internal memory is further configured to read the first model and the at least one second model stored in the main memory.

9 . The neural processing unit of claim 8 ,

wherein each of the first model and the at least one second model includes a parameter, and

wherein the internal memory is further configured to read, from the main memory, the parameter of the first model or the parameter of the at least one second model, the parameter being tiled to a predetermined size based on a capacity of the internal memory.

10 . The neural processing unit of claim 8 ,

wherein each of the first model and the at least one second model includes a parameter, and

wherein the internal memory is further configured to store the parameter of the first model, and optionally read the parameter of the at least one second model from the main memory.

11 . The neural processing unit of claim 7 ,

wherein the at least one object belongs to one category selected from among a plurality of categories.

12 . The neural processing unit of claim 7 ,

wherein the first model includes an input layer and an output layer comprising a plurality of nodes, and

wherein the number of the plurality of second models corresponds to the number of the plurality of nodes of the output layer of the first model.

13 . The neural processing unit of claim 7 , wherein the at least one second model includes at least one of a denoising model, a deblurring model, an edge enhancement model, a demosaicing model, a color tone enhancing model, a white balancing model, a super resolution model, a wide dynamic range model, a high dynamic range model, and a decompression model.

14 . The neural processing unit of claim 7 ,

wherein the at least one second model includes an ensemble model in which at least two second models selected from among the plurality of second models are combined and are connected in parallel or in series, and

wherein the second image is output from one of the parallel connection and the series connection.

15 . The neural processing unit of claim 7 , wherein the controller is further configured to combine regions processed by each of the at least one second model to output the second image having improved quality.

16 . The neural processing unit of claim 7 ,

wherein the at least one second model includes a parameter,

wherein the internal memory is further configured to store the parameter of the at least one second mode and a plurality of images including the first image, and

wherein the parameter stored in the internal memory corresponds to a classification result of an object for a previous image when a classification result of an object for a selected image among the plurality of images by the first model is the same as the classification result of the object for the previous image.

17 . A neural processing unit comprising:

an internal memory configured to store a first model, at least one second model, and at least part of data of a first image, the first image including at least one object belonging to one category selected from among a plurality of categories and a region of interest (ROI) determined based on gaze data acquired from a head-mounted display (HMD) device, the gaze data indicative of a gaze of a user viewing the first image;

a processing element (PE) array configured to access the internal memory and to process convolution of the first model and the at least one second model; and

a controller operatively coupled to the internal memory and the processing element,

wherein the first model includes a plurality of artificial neural network-based model trained to determine a region of the at least one object by inputting the first image to the first model and to classify an object corresponding to a predetermined category,

wherein the at least one second model includes a plurality of artificial neural network-based models respectively trained to output an image to which a category-specific image processing corresponding to an object category is applied by inputting the first image to the at least one second model, and

wherein the controller is configured to

induce the PE array to determine, as the ROI, a region in which the gaze of the user stays relatively long among regions of each of the at least one object, the ROI determined based on the gaze data by using the first model,

re-input each determined ROI to the first model to classify a category of each of the at least one object, and

output a second image having improved quality by applying a parameter corresponding to the category of each of the at least one object to the second model such that an ROI-specific image processing corresponding to an object category of each determined ROI is applied to the first image, the applied parameter selected from among a plurality of parameters predetermined for each of the plurality of categories.

Assignments (1)

Priority Claims (1)

Continuity (1)

References Cited (18)