Image processing apparatus and operating method thereof
A method of operating an image processing apparatus is provided. The method includes generating a first feature map by performing a convolution operation between a first image and a first kernel group, generating a second feature map by performing a convolution operation between the first image and a second kernel group, generating a first combination map based on the first feature map, generating a second combination map based on the first feature map and the second feature map, generating a second image based on the first combination map and the second combination map, and generating a reconstructed image of the first image, based on the second image and the first image, and generating a high-resolution image of the first image by inputting the reconstructed image to an upscaling model.
1. A method of operating an image processing apparatus, the method comprising:
generating a first feature map by performing a convolution operation between a first image and a first kernel group;
generating a second feature map by performing a convolution operation between the first image and a second kernel group;
generating a combination map by summing the first feature map and the second feature map;
generating a second image based on the first feature map and the combination map; and
generating a reconstructed image of the first image, based on the second image and the first image, and generating a high-resolution image of the first image by inputting the reconstructed image to an upscaling model,
wherein a size of at least one kernel included in the first kernel group and a size of at least one kernel included in the second kernel group are equal to each other,
wherein a dilation rate of the second kernel group is greater than a dilation rate of the first kernel group, and
wherein a number of kernels included in the first kernel group is equal to a number of kernels included in the second kernel group.
2. The method of claim 1 ,
wherein the size of the at least one kernel included in the first kernel group and the size of the at least one kernel included in the second kernel group are each 3×3.
3. The method of claim 1 ,
wherein the dilation rate of the first kernel group is 1, and
wherein the dilation rate of the second kernel group is 2.
4. The method of claim 1 , wherein the combination map is generated by weighted-summing the first feature map and the second feature map.
5. The method of claim 1 , further comprising:
generating a third feature map by performing a convolution operation between the first image and a third kernel group; and
generating a second combination map by summing the second feature map and the third feature map without including the first feature map,
wherein the generating of the second image comprises generating the second image based on the first feature map, the combination map, and the second combination map.
6. The method of claim 1 ,
wherein the second image is generated by inputting the first feature map and the combination map to an activation function, and
wherein the activation function comprises at least one of a rectified linear unit (ReLU) function, a sigmoid function, a Tanh function, or an exponential linear unit (ELU) function.
7. The method of claim 1 , wherein the generating of the second image comprises:
generating a first activation map based on the first feature map and the combination map;
generating a fourth feature map by performing a convolution operation between the first activation map and a fourth kernel group;
generating a fifth feature map by performing a convolution operation between the first activation map and a fifth kernel group;
generating a fourth combination map based on the fourth feature map;
generating a fifth combination map based on the fourth feature map and the fifth feature map; and
generating the second image based on the fourth combination map and the fifth combination map,
wherein a size of at least one kernel included in the fourth kernel group and a size of at least one kernel included in the fifth kernel group are equal to each other, and
wherein a dilation rate of the fifth kernel group is greater than a dilation rate of the fourth kernel group.
8. The method of claim 7 ,
wherein the second image is generated by inputting the fourth combination map and the fifth combination map to an activation function, and
wherein the activation function comprises at least one of a rectified linear unit (ReLU) function, a sigmoid function, a Tanh function, or an exponential linear unit (ELU) function.
9. The method of claim 1 , wherein the generating of the high-resolution image of the first image comprises:
identifying an upscaling coefficient of the first image; and
generating the high-resolution image of the first image by inputting the reconstructed image of the first image to the upscaling model corresponding to the identified upscaling coefficient among a plurality of upscaling models.
10. The method of claim 1 , wherein at least one of the first kernel group, the second kernel group, or the upscaling model is trained using a training data set comprising low-resolution images and high-resolution images respectively corresponding to the low-resolution images.
11. An image processing apparatus comprising:
at least one processor including processing circuitry; and
memory storing at least one instruction that, when executed by the at least one processor individually or collectively, cause the image processing apparatus to:
generate a first feature map by performing a convolution operation between a first image and a first kernel group,
generate a second feature map by performing a convolution operation between the first image and a second kernel group,
generate a combination map by summing the first feature map and the second feature map,
generate a second image based on the first feature map and the combination map, and
generate a reconstructed image of the first image, based on the second image and the first image, and generate a high-resolution image of the first image by inputting the reconstructed image to an upscaling model,
wherein a size of at least one kernel included in the first kernel group and a size of at least one kernel included in the second kernel group are equal to each other,
wherein a dilation rate of the second kernel group is greater than a dilation rate of the first kernel group, and
wherein a number of kernels included in the first kernel group is equal to a number of kernels included in the second kernel group.
12. The image processing apparatus of claim 11 ,
wherein the size of the at least one kernel included in the first kernel group and the size of the at least one kernel included in the second kernel group are each 3×3.
13. The image processing apparatus of claim 11 ,
wherein the dilation rate of the first kernel group is 1, and
wherein the dilation rate of the second kernel group is 2.
14. The image processing apparatus of claim 11 , wherein the combination map is generated by weighted-summing the first feature map and the second feature map.
15. The image processing apparatus of claim 11 , wherein the at least one instruction that, when executed by the at least one processor individually or collectively, cause the image processing apparatus to:
generate a third feature map by performing a convolution operation between the first image and a third kernel group,
generate a second combination map by summing the second feature map and the third feature map without including the first feature map, and
generate the second image based on the first feature map, the combination map, and the second combination map.
16. The image processing apparatus of claim 11 ,
wherein the second image is generated by inputting the first feature map and the combination map to an activation function, and
wherein the activation function comprises at least one of a rectified linear unit (ReLU) function, a sigmoid function, a Tanh function, or an exponential linear unit (ELU) function.
17. The image processing apparatus of claim 11 , wherein to at least one instruction that, when executed by the at least one processor individually or collectively, cause the image processing apparatus to:
generate a first activation map based on the first feature map and the combination map,
generate a fourth feature map by performing a convolution operation between the first activation map and a fourth kernel group,
generate a fifth feature map by performing a convolution operation between the first activation map and a fifth kernel group,
generate a fourth combination map based on the fourth feature map,
generate a fifth combination map based on the fourth feature map and the fifth feature map, and
generate the second image based on the fourth combination map and the fifth combination map,
wherein a size of at least one kernel included in the fourth kernel group and a size of at least one kernel included in the fifth kernel group are equal to each other, and
wherein a dilation rate of the fifth kernel group is greater than a dilation rate of the fourth kernel group.
18. The image processing apparatus of claim 17 ,
wherein the second image is generated by inputting the fourth combination map and the fifth combination map to an activation function, and
wherein the activation function comprises at least one of a rectified linear unit (ReLU) function, a sigmoid function, a Tanh function, or an exponential linear unit (ELU) function.
19. The image processing apparatus of claim 11 , wherein at least one instruction that, when executed by the at least one processor individually or collectively, cause the image processing apparatus to:
identify an upscaling coefficient of the first image, and
generate the high-resolution image of the first image by inputting the reconstructed image of the first image to the upscaling model corresponding to the identified upscaling coefficient among a plurality of upscaling models.
20. At least one non-transitory computer-readable recording media having recorded thereon a program for executing the method of claim 1 .