Length Independent Writer Identification Based on the Fusion of Deep and Hand-Crafted Descriptors

• Published in: IEEE access 2019, Vol.7, p.91772-91784
• Main Authors: Sulaiman, Alaa, Omar, Khairuddin, Nasrudin, Mohammad F., Arram, Anas
• Format: Article
• Language: English
• Subjects: Context; Datasets; deep descriptor; Deep learning; Encoding; Feature extraction; feature fusion; feature length independent; Frequency-domain analysis; hand-crafted feature; Handwriting; Languages; Learning; Mathematical analysis; Matrix algebra; Matrix methods; Shape; Writer identification; Writing
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2019.2927286

Writer's identification from a handwritten text is one of the most challenging machines learning problems because of the variable handwritten sources, various languages, the similarity between writer's pattern, context variation, and implicit characteristics of handwriting styles. In this paper, a combination of the deep and hand-crafted descriptor is utilized to learn patterns from the handwritten images. First, to do so, the local patches are extracted from the handwritten images. Then, these patches are simultaneously fed to deep and hand-crafted descriptors to generate the local descriptions. The extracted local features are then assembled to make the whole description matrix. Finally, by applying the vector of locally aggregated descriptors (VLAD) encoding on the description matrix, a 1-D feature vector is extracted to represent the writer's pattern. It is worthwhile to mention that the generated description does not rely on any language model or context information. Thus, the proposed approach is language and content independent. In addition, the proposed method does not have any restriction on the input length, hence, the writer's sample can be a passage, paragraph, line, sentence, or even a word. The obtained results on three public benchmark datasets of IAM, CVL, and Khatt indicate that the proposed method has a high-accuracy rate in writing identification task. Furthermore, the performance of the proposed method on CVL dataset using both German and English samples demonstrates that the proposed approach has a high capability in learning a writer's pattern from both languages at the same time.