Segmenting letters in words written on a tablet during a dictation

The goal of this Master Thesis is, given writing data recorded by a tablet during a dictation task as a temporal sequence of samples that measure, as a function of time, the spatial position of the pen tip on the tablet, as well as "Pen-Up" and "Pen-Down" switching, be able to segment that data into smaller units, starting from a higher level, like lines of text, until a lower level, the letters. On the one hand, given a series of samples corresponding to texts written by children, a temporary analysis of these samples is carried out in order to segment the text into lines. With this lines of text, the links between strokes are analyzed and classified into two groups, inter-letter and intra-letter links, by means of a binary classifier based on SVM with the aim of segmenting each line of text into words. On the other hand, to perform the segmentation of words into letters, in addition to using the data provided by the tablet, it is used the position of all the templates given by a method implemented in a previous Master Thesis performed by Clara Sansalvadó that first warps reference letter templates using Lucas-Kanade optimization algorithm and then defines an optimal sequence of reference templates and warping parameters, based on a shortest-path algorithm. With this, it is defined a minimization problem which finds the best set of endpoints of the letters which minimizes the distance between the warped templates and the function of time that defines the tracing of the word in order to segment the word into letters. It is defined two different strategies to perform that, one based on the use of the center of each template and another on the used of the distance map of each template. Finally, an analysis of the results of both word and letter segmentation will be presented as well as an analysis of the weak points that the proposed methods have.