Difficulty-Aware Interleaved Distillation for Robust Cross-Surface Writer Identification

Abstract

Writer identification in realistic settings remains challenging due to pronounced intra-writer variability caused by changes in writing surfaces, instruments, posture, and acquisition devices. These condition-induced variations introduce significant distribution shifts in handwriting appearance and stroke dynamics, often degrading the performance of conventional deep learning models. To address this problem, we propose a difficulty-aware interleaved distillation framework for robust cross-surface writer identification. The proposed method follows a teacher–student architecture, where a transformer-based teacher captures global contextual handwriting characteristics, while a lightweight convolutional student focuses on fine-grained, writer-specific stroke patterns. Knowledge transfer is governed by a confidence-guided interleaved distillation strategy that dynamically alternates between supervised learning and teacher-guided optimization. Robustness under severe domain shifts is further enhanced through difficulty-aware data augmentation, embedding-level alignment between teacher and student representations, and an on-policy confidence-based regularization objective that stabilizes learning. Extensive experiments conducted on an extended dataset comprising 3520 text-line samples from 160 writers, evaluated across seven carefully designed cross-condition protocols, demonstrate consistent and significant improvements over strong baselines and recent state-of-the-art methods, achieving up to 88.28% top-1 accuracy under challenging cross-surface and cross-device scenarios. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2027.