Multistage Convolutional Neural Network With Deformable Attention for Word-Level Offline Text-Independent Writer Identification
Manabu Okawa · IEEE Access · 2025
Offline writer identification is an essential component of biometric and forensic security fields, focusing on identifying the authors of handwritten documents in physical form. Historically, this approach has focused on text-dependent identification using page-level document images containing multiple paragraphs and sentences. However, text-independent identification using a limited number of word-level text images still presents challenges in performance, especially when relying on traditional handcrafted features. To address these issues, this study improves the generalization and interpretative capabilities in word-level offline text-independent writer identification by introducing a novel multistage convolutional neural network designed as an end-to-end model. We then improve the model's ability to interpret and classify handwriting effectively by incorporating the insights of handwriting analysts through the use of a deformable attention module. The experimental validation on four public datasets, namely, IAM, CVL, Firemaker, and CERUGEN, confirmed the effectiveness of the proposed method.