KhmerWriterID: Toward Robust Khmer Writer Verification Using Deep Learning: CNN, RNN, and Hybrid Approaches

Writer verification from online handwriting remains a challenging pattern-recognition problem and is particularly under-explored for complex Southeast Asian scripts such as Khmer. We address online, text-independent, word-level Khmer writer verification, formulated as a pairwise decision task: given a pair of word instances, the system determines whether they were written by the same writer. We introduce KhmerWriterID, a hybrid Siamese architecture that processes dual input streams: grayscale word images and pen-based stroke coordinates. Each branch couples a lightweight CNN for spatial feature extraction with a bidirectional GRU (BiGRU) encoder for stroke dynamics. The two modality embeddings are fused into a 128-D representation, and similarity is computed using scaled cosine similarity. We evaluate a Khmer online handwriting dataset with both image and coordinate modalities under strict writer-disjoint splits. A Khmer-aware preprocessing pipeline performs sample cleaning, coordinate normalization, scaling to [0, 1], and pen-state encoding. We report ROC-AUC, EER, and performance at FAR=1%. KhmerWriterID achieves 99.74% test accuracy and 99.68% F1, with validation AUC=99.92% and EER≈0.32%. At FAR=1%, it maintains Acc=99.27% and F1=99.09% with FRR≈0.07%. These results indicate that jointly modeling spatial word structure and online stroke dynamics enables robust Khmer writer verification, with potential applications in forensic document analysis, digital authentication, and script-specific handwriting research.