No abstract

No abstract

A new description method is proposed for sketched symbol recognition. It incorporates local direction information into the Zernike moments which represent only the spatial distribution of sample points. A symbol is decomposed into several component patterns according to the local direction of sample points before Zernike moments computation. The resulting descriptor inherits from the traditional Zernike moments descriptor the invariability to stroke number, stroke order and symbol rotation. Moreover, the fusion of both types of data makes it more informative and discriminative, resulting in better performances in both rotation-invariant classification and rotation angle estimation.Introduction: Challenges in sketch recognition lie mainly in the variations and deformations of hand-drawn symbols [1]. A practical recognition system should take into consideration issues like stroke number, stroke order, as well as the symbol's orientation. In many applications, a graphical symbol can be drawn towards different orientations; hence, the recognition algorithm should be rotation free: either to classify a symbol rotation invariantly, or to compensate for the rotation angle in pre-processing.The Zernike moment (ZM) feature is a powerful image-based descriptor that has been widely used in image processing. It has been introduced into the domain of handwriting recognition [2] and online sketch recognition [1]. Traditionally, only the magnitudes of ZMs were extracted as rotation-invariant features in recognition [1,3]. In recent studies [4,5], the moment phase is also exploited for improving recognition performance, or even estimating the rotation angle between the symbols being matched. The ZMs essentially encode the twodimensional spatial distribution of sample points composing the sketched strokes. In other words, the ZM descriptor mainly describes the symbol's visual appearance but not the way in which it was drawn. Therefore this representation is insensitive to stroke number, stroke order, and it is free of stroke segmentation, in contrast to structural recognition approaches.Besides the offline pattern, the online data of how a symbol is drawn also provide valuable information for recognition. In fact, the offline information and online information are usually complementary since they describe the sketched symbol from different perspectives. Efforts have been made to utilise both of them using combination methods such as Mean rule, the Dempster-Shafer combination rule and Naive Bayes [3].We present in this Letter a new descriptor called directional ZMs (DZMs) for rotation-free recognition of online sketched symbols. DZM represents not only the spatial distribution of sample points, but also their local directional information. As a result, it is more informative and consequently more effective than the traditional ZM descriptor in sketched symbol recognition.

No abstract

No abstract