FIG. 8 illustrates three columns of images 802a-c. Column 802b of FIG. 8 shows an original extracted handwriting image 804a. Below the original image 804a are two warped versions of the image 804b and 804c. Each image 804a-c is rotated left and right through 1-degree increments up to +/?8-degrees in various embodiments. The left and right rotations of the images 804a-c are shown in the left and right columns as images 806a-c and 808a-c respectively. One or more of the enhanced images 804b-c, 806a-c, and 808a-c, generated by the original image 802a, may be included in a training set for a model configured to recognize handwriting.
Some embodiments utilize a Speeded Up Robust Features (SURF) algorithm to identify key points in the training set. With this approach, key points are selected to create a visual “bag-of-words” which is used in a Support Vector Machine to distinguish language classes. This approach works well considering the SURF algorithm is an “off-the-shelf” algorithm that has no specific customization for handwriting language recognition.
