Off-line writer identification using handcrafted features versus ConvNets

The writer identification task by using manuscripts became a very important research topic in forensics analysis of documents. It happens because the writing can be considered as an identifying characteristic of a person. There are different databases of writers through the use of manuscripts and containing different alphabets. In this paper, the CVL database will be used to address the writer identification problem with the focus of using manuscripts taken from different languages (i.e. English and German), but using the same alphabet (Latin). To that end, the texture generation is created starting from the original documents and textural features are extracted by using some well-known texture descriptors presented in the literature: Local Binary Pattern (LBP), and Local Phase Quantization (LPQ). Furthermore, a Convolutional Neural Networks is used to perform the same classification task also starting from the original documents after the texture generation is performed. In both approaches, a zoning scheme which divides the image in smaller parts is used, and one classifier is created for each zone. After that, the outputs of classifiers from different zones are combined using some well-known (i.e. max rule, sum rule, and product rule) in order to get the final decision.