Automatic off-line text-independent writer identification from handwriting
Abderrazak Chahi · theses.fr (ABES) · 2021
Handwriting-based writer identification has experienced a resurgence in recent years and continues to attract a great deal of interest and attention in the field of biometrics and pattern recognition. It is a challenging task considering the large within-writer and between-writer style variability. The automatic offline writer identification systems consider handwriting as scanned image containing certain recurring patterns that need to be analyzed. The motivation for this work stems from the need to improve behavioral biometric tasks that have been mainly used for writer identification to enhance security and forensic applications in today's world. The interest is to develop near real-time, effective, and robust approaches for writer identification by leveraging theoretical and technical advances in image analysis and artificial intelligence. This dissertation contributes to the numerous challenges encountered in all the main steps of an automatic system for offline writer identification, including image pre-processing and segmentation, feature extraction, and classification methods.Our first contribution investigates writer identification based on texture features. We propose four texture-based approaches to improve the task of writer identification: (1) The first approach, namely the Block Wise Local Binary Count (BWLBC)-based system, characterizes the variability of writing style within small blocks by capturing the pixels' distribution corresponding to writing ink in binary components ; (2) In the second approach, the Local Binary Patterns (LBP), Local Ternary Patterns (LTP), and Local Phase Quantization (LPQ) hand-crafted descriptors are applied to small regions of interest in the writing, called zones, to extract related texture features. They are performed in an efficient way using a new learning framework ; (3) The task of writer identification is improved thanks to a well-defined approach based on the Cross multi-scale Locally encoded Gradient Patterns (CLGP) descriptor to better represent salient local writing structures. It extracts transform features from connected components and encodes the obtained texture codes in multiple scales over the Histograms of Oriented Gradients (HOG) ; (4) The fourth approach computes local intensity gradients of the writing within non-overlapping blocks using the Local gradient full- Scale Transform Patterns (LSTP) method. This feature gives the overall system the ability to extract more relevant information to characterize the writing better. Convolutional Neural Networks (CNN) are also investigated to further improve the identification performance. Two computationally efficient and high-quality deep CNN-based approaches named DeepWINet and WriterINet are proposed. Extensive experiments are conducted on ten challenging handwritten benchmarks in different languages (English, Arabic, French, German, Chinese, Dutch, Greek, and hybrid). All the proposed approaches achieve competitive, or the highest SOTA performance in the benchmarks studied. We also participated in the ICFHR2020 competition to award the best approach for image retrieval for historical handwritten fragments. We proposed an effective deep learning-based approach based on multi-path CNN streams trained with different input data. The overall approach achieved excellent results and won first place in one of two tracks of the contest.