Vertex interpolation in one-shot learning for object classification

A non-transitory computer-readable storage medium storing computer-executable instructions that, in response to executing, cause a device comprising a processor to perform operations, comprising:

• receiving a source image depicting an object for identification;
• determining a type and shape of the depicted object;
• overlaying the source image with a N×M grid of vertices and horizontal and vertical edges to fit the shape of the depicted object;
• for each vertex in the grid, adding perturbations to (x,y) coordinates of the vertex;
• and
• interpolating pixel values in a range between original and final (x,y) coordinates to generate an item of synthetic training data.

A computer-implemented method comprising:

• generating a plurality of items of synthetic training data comprising:
• receiving, by a processor, a source image depicting an object for identification;
• determining, by the processor, a type and shape of the depicted object;
• overlaying, by the processor, the source image with an N×M grid of vertices and horizontal and vertical edges to fit the shape of the depicted object;
• for each vertex in the grid, adding perturbations to (x,y) coordinates of the vertex;
• and
• interpolating pixel values in a range between original and final (x,y) coordinates to generate an item of synthetic training data;
• repeating the addition of perturbations and pixel interpolations on the source image overlayed with the N×M grid to create a plurality of items of synthetic training data, wherein the (x,y) coordinates of the vertices are manipulated differently for each item of synthetic training data;
• training a convolutional neural network using the plurality of items of synthetic training data;
• and
• identifying the object with the trained convolutional neural network comprising:
• receiving, by a processor, an input image depicting the object;
• identifying the object in the input image;
• retrieving metadata associated with the object, based on the identification of the object;
• and
• outputting the retrieved metadata.