Image scaling using a convolutional neural network

A method for image scaling for a digital magazine comprising:

• receiving an image from a digital magazine server, the image having a first number of pixels;
• increasing the resolution of the image using a convolutional neural network, where increasing the resolution comprises:
• dividing the image into a plurality of subregions having associated subregion matrices, each subregion matrix having a first number of entries, each entry of each subregion matrix corresponding to a pixel of the image,
• determining similarity values by performing a convolution operation between each of the plurality of subregion matrices and a first set of kernel matrices representing common image features, each of the kernel matrices in the first set of kernel matrices having a first number of entries equal to the first number of entries of each subregion matrix,
• increasing the resolution of the subregions by expressing each subregion as a combination of two or more of a second set of kernel matrices representing the common image features, wherein each of the kernel matrices in the second set of kernel matrices has a second number of entries that is higher than the first number of entries, wherein the combination is based on the similarity values and one or more parameters determined by training the convolutional neural network, and
• reassembling the image using the subregions with increased resolution;
• and
• outputting the image for display in a digital magazine.

A system for image scaling for a digital magazine comprising:

• a processor configured to execute modules;
• and
• a memory storing the modules, the modules executable to perform steps comprising:
• receiving a training image;
• scaling down the training image by reducing a resolution thereof;
• increasing the resolution of the scaled training image using a convolutional neural network to form an output image;
• comparing the training image to the output image;
• and
• determining one or more parameters for the convolutional neural network based on the comparing, the convolutional neural network used to increase the resolution of images for presentation in a digital magazine.

A system for image scaling for a digital magazine comprising:

• a processor configured to execute modules;
• and
• a memory storing the modules, the modules executable to perform steps comprising:
• receiving an image from a digital magazine server, the image having a first number of pixels;
• increasing the resolution of the image using a convolutional neural network, where increasing the resolution comprises:
• dividing the image into a plurality of subregions having associated subregion matrices, each subregion matrix having a first number of entries, each entry of each subregion matrix corresponding to a pixel of the image,
• determining similarity values by performing a convolution operation between each of the plurality of subregion matrices and a first set of kernel matrices representing common image features, each of the kernel matrices in the first set of kernel matrices having a first number of entries equal to the first number of entries of each subregion matrix,
• increasing the resolution of the subregions by expressing each subregion as a combination of two or more of a second set of kernel matrices representing the common image features, wherein each of the kernel matrices in the second set of kernel matrices has a second number of entries that is higher than the first number of entries, wherein the combination is based on the similarity values and one or more parameters determined by training the convolutional neural network, and
• reassembling the image using the subregions with increased resolution;
• and
• outputting the image for display in a digital magazine.

A non-transitory computer-readable storage medium storing executable computer program instructions for image scaling for a digital magazine, the instructions executable to perform steps comprising:

• receiving a training image;
• scaling down the training image by reducing a resolution thereof;
• increasing the resolution of the scaled training image using a convolutional neural network to form an output image;
• comparing the training image to the output image;
• and
• determining one or more parameters for the convolutional neural network based on the comparing, the convolutional neural network used to increase the resolution of images for presentation in a digital magazine.

A non-transitory computer-readable storage medium storing executable computer program instructions for image scaling for a digital magazine, the instructions executable to perform steps comprising:

• receiving an image from a digital magazine server, the image having a first number of pixels;
• increasing the resolution of the image using a convolutional neural network, the increasing the resolution comprising:
• dividing the image into a plurality of subregions having associated subregion matrices, each subregion matrix having a first number of entries, each entry of each subregion matrix corresponding to a pixel of the image,
• determining similarity values by performing a convolution operation between each of the plurality of subregion matrices and a first set of kernel matrices representing common image features, each of the kernel matrices in the first set of kernel matrices having a first number of entries equal to the first number of entries of each subregion matrix,
• increasing the resolution of the subregions by expressing each subregion as a combination of two or more of a second set of kernel matrices representing the common image features, wherein each of the kernel matrices in the second set of kernel matrices has a second number of entries that is higher than the first number of entries, wherein the combination is based on the similarity values and one or more parameters determined by training the convolutional neural network, and
• reassembling the image using the subregions with increased resolution;
• and
• outputting the image for display in a digital magazine.