Engineering Applications of Artificial Intelligence 95 (2020) 103912 Contents lists available at ScienceDirect Engineering Applications of Artificial Intelligence journal homepage: www.elsevier.com/locate/engappai A deep learning framework for text-independent writer identification Malihe Javidi, Mahdi Jampour ∗ Quchan University of Technology, Quchan, Iran ARTICLE INFO Keywords: Enhanced ResNet Deep residual networks Text-independent Handwriting writer recognition ABSTRACT Handwriting Writer Identification (HWI) refers to the process of handwriting text image analysis to identify the authorship of the documents. It has yielded promising results in various applications, including digital forensics, criminal purposes, exploring the writer of historical documents, etc. The complexity of the text image, especially in images with various handwriting makes the writer identification difficult. In this work, we propose an end-to-end system that relies on a straightforward yet well-designed deep network and very efficient feature extraction, emphasizing feature engineering. Our system is an extended version of ResNet by conjugating deep residual networks and a new traditional yet high-quality handwriting descriptor towards handwriting analysis. Our descriptor analyzes the handwriting thickness as a preliminary and essential feature for human handwriting characteristics. Our approach can also provide text-independent writer identification that we do not need to have the same handwriting content for learning our model. The proposed approach is evaluated and achieved consistent results on four public and well-known datasets of IAM, Firemaker, CVL, and CERUG-EN. We empirically demonstrate that our conjugated network outperforms the original ResNet, and it can work well for real-world applications in which patches with few letters exist. 1. Introduction Handwriting is an individual human characteristic that represents the writer’s psychological state during the writing and can prove a person’s authenticity through its pattern analysis. It has many potential applications such as digital forensics, criminal purposes, exploring the writer of historical documents, etc. Handwriting writer identification refers to finding an author from his/her handwriting documents among numerous writers’ documents. This identifying is possible due to the handwriting style, which has multiple specific features that show the author’s personality. The features are including the shape of letters, spacing between letters, the slope of letters, cursive or separated writ- ing, rhythmic repetition of the elements, the pressure to the paper, the size of letters, the thickness of letters, etc. (Khan et al., 2019). Since a human can intelligently discover a bunch of above features, modern AI systems focus on proposing approaches that can imitatively identify human from handwriting images that is an offline handwriting writer identification system. In contrast, an online writer identification system analyzes a script during the writing. Therefore, it needs expensive peripheral devices like a digital pen to record vital information such as writing speed, chaotic behavior, etc. Of course, more information helps to achieve a better identification rate. Handwriting writer identification has numerous employment. While handwriting analysis is useful for guessing the author of the historic or non-historic documents based on the recognized scripts (De Stefano ∗ Corresponding author. E-mail address: jampour@qiet.ac.ir (M. Jampour). et al., 2018), it is also pretty beneficial for forensics purposes like writer identification or verification due to carry the personality of the authors. Handwriting writer identification can also be divided into two branches of text-depended and text-independent categories. Both approaches find the most similar handwriting (author label) from a pre-defined gallery compared to the input test handwriting sample. Nevertheless, the former refers to the methods that use the same text for discovering author properties, while the latter is not sensitive to the analyzed text content (Xiong et al., 2017). This means that the content of the test and train texts must not be needed to be the same for all authors. Therefore, only the second category can be used for applications like exploring writers of historical documents, although both are useful for forensic purposes. In contrast to the identification, writer verification refers to confirming the similarity of two scripts (items). In handwriting analysis, while we know who the author of a document is, we need to verify whether the claimed person is correct or not? Similar to other verification biometrics like signature verification as well as fingerprint verification. Of course, the identification is more complicated than verification because the authentication system needs to find the similarity of input sample from thousands of examples in the gallery, which means that the identification is a process of finding one from N while verification is confirming 1 by 1. Although, several approaches achieved excellent performance on different writer identification datasets; they are still far to address this problem in more challenging datasets or real-world applications. https://doi.org/10.1016/j.engappai.2020.103912 Received 27 February 2020; Received in revised form 26 July 2020; Accepted 23 August 2020 Available online 28 August 2020 0952-1976/© 2020 Elsevier Ltd. All rights reserved.