Font recognition and font similarity learning using a deep neural network

A computer-implemented method comprising:

• synthesizing, by a computing device, a test image comprising text by introducing variances, the variances including one or more of character spacing, noise, shading, affine transformation, compression artifacts, aspect ratio, or Gaussian blurring;
• cropping, by the computing device, a plurality of test image patches from the test image;
• inputting, by the computing device, each test image patch of the plurality of test image patches to a convolutional neural network (CNN) that has been trained with a plurality of training image patches;
• obtaining, by the computing device, a test output for each test image patch of the plurality of test image patches, wherein the test output includes one or more dimensions;
• performing, by the computing device, a max operation along each of the one or more dimensions of the test output, a result of the max operation being a feature representation for the test image;
• and
• determining, by the computing device, a font similarity for the test image.

A computerized system comprising:

• one or more processors;
• and
• one or more computer storage media storing computer-useable instructions that, when used by the one or more processors, cause the one or more processors to:
• synthesize one or more text images comprising a font to introduce variances based on variances that exist in real-world text images;
• generate one or more training images that include the variances in the one or more text images;
• crop the one or more training images into one or more training patches that are utilized as an input to a convolutional neural network (CNN);
• train the CNN with the one or more training patches, the CNN comprising a plurality of convolutional layers and a plurality of fully connected layers;
• produce a distribution of classified text images;
• utilize the CNN to extract a feature representation from each of the one or more text images;
• utilize the CNN to extract a feature representation from a test image comprising an unknown font;
• and
• determine a font similarity between the unknown font of the test image and one or more fonts that comprise one or more of the classified text images.