Image upsampling with training images

A method of upsampling an image using an existing knowledge base, the existing knowledge base including a plurality of patch pairs derived from a set of high-resolution training images, each patch pair including a candidate high-frequency patch representing high-frequency aspects of a plurality of neighboring pixels and a corresponding candidate low-frequency patch representing low-frequency aspects of the plurality of neighboring pixels, the method comprising:

• interpolating a low-resolution, full-frequency image into a high-resolution, low-frequency image;
• decomposing the high-resolution, low-frequency image into a low-frequency component comprising one or more low-frequency primitives;
• defining for each low-frequency primitive, a sample low-frequency patch including the low-frequency primitive;
• identifying for each sample low-frequency patch a plurality of corresponding candidate low-frequency patches from the existing knowledge base, and for each candidate low-frequency patch selecting from its patch pair an associated candidate high-frequency patch;
• identifying for each sample low-frequency patch an inferred first-tier high-frequency patch selected from the corresponding candidate high-frequency patches;
• constructing a coarse high-frequency component comprising the inferred first-tier high-frequency patches;
• constructing a high-compatibility component from the low-resolution, full-frequency image without using the existing knowledge base;
• combining the high-compatibility component with the coarse high-frequency component to yield an enhanced high-frequency component, the enhanced high-frequency component comprising a plurality of enhanced high-frequency patches;
• identifying for each enhanced high-frequency patch a plurality of corresponding candidate high-frequency patches from the existing knowledge base;
• identifying for each enhanced high-frequency patch an inferred third-tier high-frequency patch selected from the corresponding candidate high-frequency patches;
• constructing an improved high-frequency component comprising the inferred third-tier high-frequency patches;
• and
• blending the improved high-frequency component with the high-resolution, low-frequency image to yield a high-resolution image.

A method of using a high-resolution, high-frequency image including a plurality of high-frequency patches in an upsampling process, the method comprising:

• identifying for each high-frequency patch a set of corresponding candidate high-frequency patches from an existing knowledge base;
• selecting from each set of candidate high-frequency patches, an inferred high-frequency patch;
• and
• constructing an improved high-frequency component comprising the inferred high-frequency patches.