Recuperação da trajetória online de caracteres latinos utilizando deep learning

Abstract

Studies on the recovery of trajectories manuscript have gained space in the field of offline handwriting recognition research. The reason is in the use of online recognition resources, creating techniques to simulate the writing of the handwritten word and inserting the simulated coordinates of the pixels in systems that recognize words online. The principle of these techniques is to find an orderly pattern similar to that made by a person during writing; this process is known as handwriting trajectory recovery (HTR). Several works have presented the use of graphs to perform the HTR, skeletonizing the characters and tracing the correct path of the graph, which this graph representing the skeleton of the character. However, recent studies characterize the use of deep learning artificial neural networks to realize the HTR. The great advantage of using deep learning networks is to take their generalization capacity to achieve better accuracy rates in the recovery of handwriting trajectories. Although these works present promising results, their results are difficult to compare with each other, since they use different evaluation metrics. Based on these factors, this work aims to present a proposal to perform the HTR of Latin characters of the IRONOFF dataset through the use of deep learning networks, in addition, present a new evaluation model called SWO (Segmentation-sliding Window-Ordering) to evaluate the sequence of coordinates predicted by the deep learning networks. Compared with the existing metrics in the literature, the SWO evaluation proved to be more effective in both quantitatively and qualitatively aspect, being able to verify whether the sequence of the strokes and the geometric shape of the character were recovered. This work also identifies that the variation in the number of coordinate points per character (𝑝/𝑐) affects the performance of deep learning networks for the problem of HTR, evidencing an approach to improve the evaluation rates. The experiments are conducted from the transformation of information online into offline. This data goes through the normalization and data augmentation process. Five network configurations are implemented and undergo the training, validation and testing process. The best results are achieved with a 95.31% accuracy rate in predicting the trajectory of single-stroke character and 92.93% accuracy in predicting the trajectory of multi-strokes character, thus presenting good results for the HTR problem.