Structure and training for image classification

Abstract

A computer implemented method of training an image classifier, comprising: receiving training images data labeled according to image classes; selecting reference points of the images; and constructing a set of voting convolutional tables and binary features on a patch surrounding each reference point by performing, for each calculation stage: creating a voting table by: creating first candidate binary features; calculating a global loss reduction for each first candidate binary feature; selecting one first candidate binary feature having minimal global loss reduction; and repeating to select stage-size binary features; and performing a tree split using the voting table by: creating second candidate binary features; calculating a combined loss reduction for each stage-split size group of the second candidate binary features; selecting one of the groups having a maximal combined loss reduction; and creating a child-directing table using the selected binary features.