Directional Hinge Features for Writer Identification: The Importance of the Skeleton and the Effects of Character Size and Pixel Intensity

Directional features such as Skeleton Hinge Distribution are not computationally expensive, are fast, easy to explain and very efficient in identifying the writer of a handwritten text. Hinge Distribution techniques are responsible for several works that have been researched in the literature recently. In this work, an attempt is made to evaluate the importance of three factors, the skeleton information, information regarding the size of the text and information regarding the grey-scale intensity of the text, that might affect writer identification accuracy in applications that utilize Directional features. Towards that goal, four new Hinge Distribution features are suggested. More specifically, the Run Length Directional Hinge Distribution (RLDHD) considers all the available pixel information in the text and the Run Length Skeleton Directional Hinge Distribution (RLSDHD), a variation of the former that utilizes the skeleton information. Furthermore, the Weighted Skeleton Hinge Distribution (WSHD) method considers the text size using the Main Body Size estimating technique and the Quantized Skeleton Hinge Distribution (QSHD) that utilizes the grey-scale intensity of the text. For the evaluation, the Firemaker Data set, IAM data set, CVL data set and the ICDAR 2017 writer identification competition Data set were considered. Our results indicate that Skeleton information plays a vital role in writer identification. On the other hand, Main Body Size fluctuations do not seem to affect identification accuracy. Finally, considering the grey-scale intensity, results seem inconclusive, and further research might be necessary.