Generation of non-primary-class samples from primary-class-only dataset
An ASDGS (Artificially Spiked Data Generation System) may comprise computing and mechanical systems for using a single class of data to generate artificially “spiked” data of a second class. To generate each spiked image, the ASDGS may randomly select a clean image an augmentation object (“AO”) from an object library, shape library, and/or hair library. The ASDGS may use a texture library to add or change the texture of the AO. The ASDGS may adjust the lighting and coloring of the AO to be similar to the clean image, and may then add the AO to the clean image to generate a spiked image.
1. A method, comprising:
obtaining an image from a first class;
obtaining an augmentation object, wherein the augmentation object is segmented and comprises at least one from the following list: a real-world object, a computer-generated shape, and an object from a hair library;
calibrating the augmentation object relative to the image from the first class to generate a calibrated augmentation object; and
adding the calibrated augmentation object to the image from the first class to generate a spiked image of a second class;
wherein:
calibrating the augmentation object comprises at least one from the following list:
adding noise to the augmentation object;
adding blurring to the augmentation object;
rotating the augmentation object;
flipping the augmentation object;
adjusting a lighting characteristic of the augmentation object to increase similarity to a lighting characteristic of the image from the first class; and
adjusting a coloring characteristic of the augmentation object to increase similarity to a coloring characteristic of the image from the first class; and
adding the calibrated augmentation object to the image from the first class comprises:
selecting a background anchor point in the image from the first class;
selecting an augmentation object anchor point in the calibrated augmentation object; and
superimposing the augmentation object onto the image from the first class with the augmentation object anchor point at the background anchor point.
2. The method of claim 1 , wherein obtaining an image from a first class comprises randomly selecting the image from the first class from a set comprising two or more images from the first class.
3. The method of claim 1 , wherein obtaining an augmentation object comprises randomly selecting an augmentation object.
4. The method of claim 1 , where calibrating the augmentation object comprises at least one from the following list:
adjusting a lighting characteristic of the augmentation object to increase similarity to a lighting characteristic of the image from the first class; and
adjusting a coloring characteristic of the augmentation object to increase similarity to a coloring characteristic of the image from the first class.
5. The method of claim 1 , further comprising applying at least one operation from the following list to the augmentation object:
randomly rotating the augmentation object;
randomly flipping the augmentation object.
6. The method of claim 1 , wherein adding the calibrated augmentation object to the image from the first class further comprises at least one from the following list:
resizing the calibrated augmentation object;
cropping the calibrated augmentation object;
adjusting transparency calibrated augmentation object;
applying edge blending to at least one of the calibrated augmentation object and the image from the first class;
de-noising the spiked image;
blurring the spiked-image; and
performing blur correction on the spiked image.
7. The method of claim 1 , wherein the object from the hair library is dynamically generated.
8. The method of claim 1 , wherein the computer-generated shape is dynamically generated.
9. The method of claim 1 , wherein the computer-generated shape comprises texture added from a texture library.
10. The method of claim 1 , wherein the object from the hair library comprises at least one from the following list: a segmented image of a hair and mathematically generated hair object.
11. The method of claim 1 , wherein the augmentation object comprises a real-world object.
12. The method of claim 1 , wherein the augmentation object comprises a computer-generated shape.
13. The method of claim 1 , wherein the augmentation object comprises an object from a hair library.