Image processing apparatus and method of generating model for image processing

1 . A method of generating a model for image processing, the method comprising:

increasing, using a receptive field (RF) increasing module of at least one processor, a receptive field of an input image frame;

generating, using a feature extraction (FE) module of the at least one processor, a feature map based on the input image frame with the increased receptive field;

generating, using a super resolution (SR) module of the at least one processor, a target image having a target resolution, based on the feature map; and

generating, using a model generation (MG) module of the at least one processor, a replacement model that replaces at least one of the RF increasing module, the FE module, and the SR module,

wherein the replacement model comprises a lookup table that maps inputs of a module to be replaced to outputs of the module to be replaced.

2 . The method of claim 1 , further comprising:

preprocessing, using a preprocessing module of the at least one processor, the input image frame.

3 . The method of claim 2 , wherein the input image frame comprises a first image frame and a second image frame,

wherein the preprocessing of the input image frame comprises converting the first image frame and the second image frame based on at least one of an average value of the first image frame and the second image frame and an interpolated value between the first image frame and the second image frame, and

wherein the increasing of the receptive field of the input image frame comprises increasing, using the RF increasing module, a first receptive field of the first image frame and a second receptive field of the second image frame.

4 . The method of claim 1 , wherein the increasing of the receptive field of the input image frame comprises performing at least one of a down shuffle, a channel-wise concatenation, and a dimension reduction.

5 . The method of claim 1 , further comprising:

training, using a training module of the at least one processor, at least one of the RF increasing module, the FE module, and the SR module, the training module being based on the target image and a ground truth image,

wherein the generating of the replacement model comprises generating, based on completion of the training, the replacement model based on a training result.

6 . The method of claim 1 , wherein the generating of the replacement model comprises generating one or more replacement models for each module to be replaced, based on context of the input image frame.

7 . The method of claim 1 , further comprising:

generating, using a weight generation (WG) module of the at least one processor, weights for each layer of a feature extraction neural network based on a result of the increasing of the receptive field,

wherein the generating of the feature map comprises:

applying the weights to the feature extraction neural network; and

providing the result of the increasing of the receptive field to the feature extraction neural network.

8 . The method of claim 7 , wherein the generating of the feature map further comprises quantizing output values of the feature extraction neural network, and

wherein the generating of the target image comprises:

dividing the quantized output values into a plurality of units of a predetermined size; and

generating a predicted pixel value of the target image for each unit of the plurality of units.

9 . The method of claim 8 , wherein the generating of the replacement model comprises:

resizing the lookup table that replaces the SR module; and

performing, for each unit of the plurality of units, at least one of adjustment of a size unit and sampling.

10 . The method of claim 7 , wherein the generating of the weights comprises applying a predetermined sparsity to the weights of each layer of the feature extraction neural network.

11 . The method of claim 7 , wherein the generating of the weights comprises clustering, using a clustering algorithm, the result of the increasing of the receptive field.

12 . The method of claim 1 , further comprising:

performing at least one of super resolution, frame interpolation, and frame interpolation after super resolution.

13 . The method of claim 12 , wherein the generating of the target image comprises, based on performing only the frame interpolation, converting the target image having the target resolution to have a same resolution as at least one of a first frame of the input image frame and a second frame of the input image frame.

14 . A non-transitory computer-readable recording medium storing one or more computer-executable programs including instructions for generating a model for image processing that, when executed by at least one processor of a device, cause the device to:

increase, using a receptive field (RF) increasing module of the at least one processor, a receptive field of an input image frame;

generate, using a feature extraction (FE) module of the at least one processor, a feature map based on the input image frame with the increased receptive field;

generate, using a super resolution (SR) module of the at least one processor, a target image based on the feature map, the target image having a target resolution; and

generate, using a model generation (MG) module of the at least one processor, a replacement model that replaces at least one of the RF increasing module, the FE module, and the SR module,

wherein the replacement model comprises a lookup table that maps inputs of a module to be replaced to outputs of the module to be replaced.

15 . An image processing apparatus, comprising:

a memory configured to store a model for image processing; and

a processor configured to:

generate the model for image processing;

store the model for image processing in the memory; and

perform image processing by using the model for image processing,

wherein the processor comprises:

a receptive field (RF) increasing module configured to increase a receptive field of an input image frame;

a feature extraction (FE) module configured to generate a feature map based on the input image frame with the increased receptive field;

a super resolution (SR) module configured to generate a target image having a target resolution, based on the feature map; and

a model generation (MG) module configured to generate the model for image processing and to replace at least one of the RF increasing module, the FE module, and the SR module with the model for image processing,

wherein the model for image processing comprises a lookup table that maps inputs of a module to be replaced to outputs of the module to be replaced.

16 . The image processing apparatus of claim 15 , wherein the processor further comprises:

a training module configured to train at least one of the RF increasing module, the FE module, and the SR module, based on the target image and a ground truth image.

17 . The image processing apparatus of claim 15 , wherein the processor further comprises:

a weight generation (WG) module configured to generate weights for each layer of a feature extraction neural network based on a result of the RF increasing module.

18 . The image processing apparatus of claim 15 , wherein the MG module is further configured to:

generate one or more models for each module to be replaced, based on context of the input image frame.

19 . The image processing apparatus of claim 15 , wherein the processor is further configured to determine a module to be replaced with the model for image processing based on computing resources of the image processing apparatus.