# US 2020/0005098 A1 · Method and apparatus for building image model

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2020/0005098 A1
- Document ID: US20200005098A1
- Status: Published application
- Kind: A1
- Application number: 16/452,879
- Filed: 2019-06-26
- Published: 2020-01-02
- Earliest claimed: 2018-07-02
- Assignees: Samsung Electronics Co., Ltd.
- Applicants: Samsung Electronics Co., Ltd.
- Inventors: Jae Mo SUNG; Changhyun KIM
- Web page: https://patentgrep.com/patent/US20200005098A1

## CPC

- G06V 10/454: {Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]}
- G06F 18/2413: based on distances to training or reference patterns
- G06V 10/764: using classification, e.g. of video objects
- G06V 10/774: Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
- G06V 10/82: using neural networks
- G06V 30/1916: {Validation; Performance evaluation}
- G06V: IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)

## Abstract

A method and apparatus for building an image model, where the apparatus generates a target image model that includes layers duplicated from a layers of a reference image model and an additional layer, and trains the additional layer.

## Claims

### Claim 1

A method of building an image model, the method comprising:
generating a target image model comprising an additional layer and remaining layers that are same as layers of a reference image model and; and
training the additional layer of the target image model based on the reference image model.

### Claim 2 (depends on 1)

The method of claim 1, wherein the generating of the target image model comprises generating the target image model by connecting the additional layer to a layer located on an input side of the remaining layers in the target image model.

### Claim 3 (depends on 1)

The method of claim 1, wherein the generating of the target image model comprises initializing the additional layer by assigning a random value to each nodes of the additional layer.

### Claim 4 (depends on 1)

The method of claim 1, wherein the training of the additional layer comprises:
determining a reference model output from an image with a converted training input, based on the reference image model; and
training the target image model based on the reference model output.

### Claim 5 (depends on 4)

The method of claim 4, wherein the determining of the reference model output comprises:
generating a conversion image by converting the training input based on an input layer of the reference image model; and
computing the reference model output from the conversion image.

### Claim 6 (depends on 1)

The method of claim 1, wherein the training of the additional layer comprises training the target image model based on a reference model output and a target model output that are based on the reference image model and the target image model, respectively.

### Claim 7 (depends on 6)

The method of claim 6, wherein the training of the additional layer comprises:
computing an error based on the reference model output and the target model output; and
updating a parameter of at least a portion of the remaining layers and the additional layer in the target image model based on the computed error.

### Claim 8 (depends on 7)

The method of claim 7, wherein the updating of the parameter comprises repeating updating of the parameter until an error between the reference model output and the target model output converges.

### Claim 9 (depends on 1)

The method of claim 1, wherein the training of the additional layer comprises updating a parameter of the additional layer.

### Claim 10 (depends on 1)

The method of claim 1, wherein the training of the additional layer comprises updating a parameter of the additional layer while maintaining parameters of the remaining layers in the target image model.

### Claim 11 (depends on 1)

The method of claim 1, wherein the generating of the target image model comprises:
generating the remaining layers of the target image model by duplicating parameters and a layer structure of the reference image model; and
connecting the additional layer to the remaining layers.

### Claim 12 (depends on 1)

The method of claim 1, further comprising:
updating parameters of the remaining layers and the additional layer in the target image model, in response to a completion of training of the additional layer.

### Claim 13 (depends on 1)

The method of claim 1, wherein a number of nodes in the additional layer is greater than or equal to a number of nodes in one of the layers of the reference image model and is less than or equal to a number of nodes in an input layer of the target image model.

### Claim 14 (depends on 1)

The method of claim 1, further comprising:
acquiring an input image;
determining a label of the input image; and
additionally training the target image model based on the input image and the label.

### Claim 15 (depends on 14)

The method of claim 14, wherein the determining of the label comprises determining the label based on a user input associated with the acquiring of the input image.

### Claim 16 (depends on 1)

The method of claim 1, further comprising:
generating an additional image model comprising an additional layer connected to the target image model, in response to a completion of training of the target image model; and
training the additional layer of the additional image model based on an output of the target image model.

### Claim 17 (depends on 1)

The method of claim 1, wherein the target image model is configured to receive an image with a resolution higher than an input resolution of the reference image model.

### Claim 18 (depends on 1)

The method of claim 1, wherein the training of the additional layer comprises training the additional layer based on feature data output from at least a portion of the layers of the reference image model and feature data output from at least a portion of the remaining layers of the target image model.

### Claim 19 (depends on 5)

The method of claim 5, wherein the resolution of the image is greater than a resolution of the conversion image.

### Claim 20 (depends on 1)

The method of claim 1, wherein a number of nodes in the additional layer is greater than a number of nodes in any of the layers of the reference image model.

### Claim 21 (depends on 1)

A non-transitory computer-readable storage medium storing instructions that, when executed by a processor, cause the processor to perform the method of claim 1.

### Claim 22

An apparatus for building an image model, the apparatus comprising:
a processor configured to generate a target image model that comprises an additional layer and remaining layers that are same as layers of a reference image model, and to train the additional layer of the target image model based on the reference image model; and
a memory configured to store the trained target image model.
