# US 12,602,740 B2 · Unsupervised learning-based scale-independent blur kernel estimation for super-resolution > Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt - Number: US 12,602,740 B2 - Document ID: US12602740B2 - Status: Granted - Kind: B2 - Application number: 17/813,551 - Filed: 2022-07-19 - Published: 2023-02-09 - Granted: 2026-04-14 - Earliest claimed: 2021-08-03 - Assignees: Samsung Electronics Co., Ltd. - Applicants: Samsung Electronics Co., Ltd. - Inventors: Kamal Jnawali; Tien Cheng Bau - Web page: https://patentgrep.com/patent/US12602740B2 ## CPC - G06T 3/4053: based on super-resolution, i.e. the output image resolution being higher than the sensor resolution - G06T 3/4046: using neural networks - G06T 5/70: Denoising; Smoothing - G06T 5/73: Deblurring; Sharpening - G06T 2207/20081: Training; Learning - G06T 2207/20084: Artificial neural networks [ANN] - G06T: IMAGE DATA PROCESSING OR GENERATION, IN GENERAL ## Abstract One embodiment provides a method generating a first image crop and a second image crop randomly extracted from a low-quality image and a high-quality image, respectively. The method further comprises comparing the first image crop and the second image crop using a plurality of loss functions including pixel-wise loss to calculate losses, and optimizing a model trained to estimate a realistic scale-independent blur kernel of a low-resolution (LR) blurred image by minimizing the losses. ## Claims ### Claim 1 A method comprising: generating a first image crop comprising a first crop of a low-quality image that is randomly extracted from the low-quality image; generating a second image crop comprising a second crop of a high-quality image that is randomly extracted from the high-quality image, wherein the low-quality image and the high-quality image are different images having different image resolutions; comparing the first image crop and the second image crop using a plurality of loss functions including pixel-wise loss to calculate losses between the first image crop and the second image crop; optimizing a model by minimizing the losses between the first image crop and the second image crop; and estimating, utilizing the optimized model, a realistic scale-independent blur kernel of the low-quality image without different scaled versions of the low-quality image. ### Claim 2 (depends on 1) The method of claim 1, further comprising: mapping, using the optimized model, a high-resolution (HR) blur-free image to a blurred output image, wherein the blurred output image has a blur distribution that is similar to the low-quality image. ### Claim 3 (depends on 1) The method of claim 1, wherein the low-quality image is a low-resolution (LR) blurred image, and the high-quality image is a high-resolution (HR) blur-free image. ### Claim 4 (depends on 3) The method of claim 3, further comprising: estimating blur kernels of a plurality of LR blurred images using the optimized model; and estimating a range of standard deviation for Gaussian blur present in the plurality of LR blurred images based on the estimated blur kernels. ### Claim 5 (depends on 4) The method of claim 4, further comprising: generating one or more LR images from a same single high-resolution (HR) image based on a standard deviation randomly selected from the range of standard deviation for Gaussian blur. ### Claim 6 (depends on 5) The method of claim 5, further comprising: generating one or more training pairs for training a deep convolutional super-resolution network, wherein each of the one or more training pairs comprises a LR image and a corresponding HR image, and at least one LR image of the one or more training pairs is included in the one or more LR images. ### Claim 7 (depends on 1) The method of claim 1, wherein the plurality of loss functions further include boundary loss, centralized loss, and sparsity loss. ### Claim 8 A system comprising: at least one processor; and a non-transitory processor-readable memory device storing instructions that when executed by the at least one processor causes the at least one processor to perform operations including: generating a first image crop comprising a first crop of a low-quality image that is randomly extracted from the low-quality image; generating a second image crop comprising a second crop of a high-quality image that is randomly extracted from the high-quality image, wherein the low-quality image and the high-quality image are different images having different image resolutions; comparing the first image crop and the second image crop using a plurality of loss functions including pixel-wise loss to calculate losses between the first image crop and the second image crop; optimizing a model by minimizing the losses between the first image crop and the second image crop; and estimating, utilizing the optimized model, a realistic scale-independent blur kernel of the low-quality image without different scaled versions of the low-quality image. ### Claim 9 (depends on 8) The system of claim 8, wherein the operations further include: mapping, using the optimized model, a high-resolution (HR) blur-free image to a blurred output image, wherein the blurred output image has a blur distribution that is similar to the low-quality image. ### Claim 10 (depends on 8) The system of claim 8, wherein the low-quality image is a low-resolution (LR) blurred image, and the high-quality image is a high-resolution (HR) blur-free image. ### Claim 11 (depends on 10) The system of claim 10, wherein the operations further include: estimating blur kernels of a plurality of LR blurred images using the optimized model; and estimating a range of standard deviation for Gaussian blur present in the plurality of LR blurred images based on the estimated blur kernels. ### Claim 12 (depends on 11) The system of claim 11, wherein the operations further include: generating one or more LR images from a same single high-resolution (HR) image based on a standard deviation randomly selected from the range of standard deviation for Gaussian blur. ### Claim 13 (depends on 12) The system of claim 12, wherein the operations further include: generating one or more training pairs for training a deep convolutional super-resolution network, wherein each of the one or more training pairs comprises a LR image and a corresponding HR image, and at least one LR image of the one or more training pairs is included in the one or more LR images. ### Claim 14 (depends on 8) The system of claim 8, wherein the plurality of loss functions further include boundary loss, centralized loss, and sparsity loss. ### Claim 15 A non-transitory processor-readable medium that includes a program that when executed by a processor performs a method comprising: generating a first image crop comprising a first crop of a low-quality image that is randomly extracted from the low-quality image; generating a second image crop comprising a second crop of a high-quality image that is randomly extracted from the high-quality image, wherein the low-quality image and the high-quality image are different images having different image resolutions; comparing the first image crop and the second image crop using a plurality of loss functions including pixel-wise loss to calculate losses between the first image crop and the second image crop; optimizing a model by minimizing the losses between the first image crop and the second image crop; and estimating, utilizing the optimized model, a realistic scale-independent blur kernel of the low-quality image without different scaled versions of the low-quality image. ### Claim 16 (depends on 15) The non-transitory processor-readable medium of claim 15, wherein the method further comprises: mapping, using the optimized model, a high-resolution (HR) blur-free image to a blurred output image, wherein the blurred output image has a blur distribution that is similar to the low-quality image. ### Claim 17 (depends on 15) The non-transitory processor-readable medium of claim 15, wherein the low-quality image is a low-resolution (LR) blurred image, and the high-quality image is a high-resolution (HR) blur-free image. ### Claim 18 (depends on 17) The non-transitory processor-readable medium of claim 17, wherein the method further comprises: estimating blur kernels of a plurality of LR blurred images using the optimized model; and estimating a range of standard deviation for Gaussian blur present in the plurality of LR blurred images based on the estimated blur kernels. ### Claim 19 (depends on 18) The non-transitory processor-readable medium of claim 18, wherein the method further comprises: generating one or more LR images from a same single high-resolution (HR) image based on a standard deviation randomly selected from the range of standard deviation for Gaussian blur. ### Claim 20 (depends on 19) The non-transitory processor-readable medium of claim 19, wherein the method further comprises: generating one or more training pairs for training a deep convolutional super-resolution network, wherein each of the one or more training pairs comprises a LR image and a corresponding HR image, and at least one LR image of the one or more training pairs is included in the one or more LR images.