PaperlayerView exact revision on arXiv

Source-linked AI summary

DeepWriterID: An End-to-end Online Text-independent Writer Identification System

Weixin Yang, Lianwen Jin, Manfei Liu

arXiv:1508.04945v2cs.CVcs.LGstat.ML

TL;DR

Writer identification faces insufficient data and difficulty designing features across handwriting conditions. DeepWriterID combines DropSegment, path-signature feature maps, and a deep CNN; on NLPR it achieves new state-of-the-art accuracies for Chinese and English text, using pen-position information in the pen-down state.

  • Problem

    Writer identification remains challenging under insufficient data, differing source lengths, and multiple handwriting languages.

    Source: §1 · p. 2

  • Method

    DeepWriterID combines DropSegment segment-removal augmentation, path-signature feature maps, and a deep CNN for online text-independent writer identification.

    Source: §1 · p. 2, ANALYSIS · p. 2, §3 · p. 3

  • Results

    DeepWriterID achieved new state-of-the-art writer identification rates of 95.72% for Chinese text and 98.51% for English text.

    Source: INVESTIGATION OF DROPSEGMENT METHOD · p. 6, §6 · p. 6

  • Takeaways & Limitations

    DropSegment supports training CNNs when writer-identification data are inadequate, and the reported system reaches new state-of-the-art accuracies on two NLPR subsets.

    Source: §6 · p. 6

    The method does not consider rejection of unknown writers, an issue left for further study.

    Source: §6 · p. 6

Abstract

from arXiv · show

Owing to the rapid growth of touchscreen mobile terminals and pen-based interfaces, handwriting-based writer identification systems are attracting increasing attention for personal authentication, digital forensics, and other applications. However, most studies on writer identification have not been satisfying because of the insufficiency of data and difficulty of designing good features under various conditions of handwritings. Hence, we introduce an end-to-end system, namely DeepWriterID, employed a deep convolutional neural network (CNN) to address these problems. A key feature of DeepWriterID is a new method we are proposing, called DropSegment. It designs to achieve data augmentation and improve the generalized applicability of CNN. For sufficient feature representation, we further introduce path signature feature maps to improve performance. Experiments were conducted on the NLPR handwriting database. Even though we only use pen-position information in the pen-down state of the given handwriting samples, we achieved new state-of-the-art identification rates of 95.72% for Chinese text and 98.51% for English text.

1. INTRODUCTION

Writer identification compares handwriting with samples of unknown authorship and serves applications including forensics and authentication. DeepWriterID addresses unresolved data and feature-design challenges with DropSegment, path-signature features, and a deep CNN.

1. INTRODUCTION

  • Writer identification ranks candidate writers by similarity between their handwriting and a sample of unknown authorship.Applications include forensic evidence, mobile banking, and network access authentication.Source: §1 · p. 1
  • Online handwriting can include position, velocity, pressure, and altitude, while text-independent systems are robust to content but require substantial data for generalization.Source-material length and language also change the requirements for generalized applicability.Source: §1 · pp. 1–2
  • DeepWriterID combines DropSegment, path-signature features, and a deep CNN to address writer identification under data, source-length, and language variation.DropSegment removes some character segments while retaining identity information; the path-signature feature represents trajectories.Source: §1 · p. 2

2. DROPSEGMENT METHOD

DropSegment addresses scarce data, generalization, segmentation, and ensemble-storage challenges by randomly removing character segments while preserving writer-discriminant information. The remaining segments can form many augmented samples, and the method is presented as more robust than dropping whole strokes.

MOTIVATION

  • The method targets scarce handwriting data, which is tedious to collect, and the data needs of text-independent identification and deep neural networks.These constraints are especially relevant when real-world users would need to provide paragraphs or pages.Source: MOTIVATION · p. 2
  • DropSegment is motivated by limited generalization across diverse handwriting structures, segmentation errors across materials and languages, and the storage cost of ensemble models.The authors seek successful ensemble-like results while maintaining constant storage size.Source: MOTIVATION · p. 2

ANALYSIS

  • DropSegment randomly removes some segments from characters to augment training data while retaining identity information.The method defines segments within strokes, for example by using stroke corner points as segmentation points.Source: ANALYSIS · p. 2, MOTIVATION · p. 2
  • By probabilistically omitting segments, DropSegment can tolerate relatively crude segmentation and avoid relying on variable stroke structure that may harm generalization.The authors also describe elastically generating test samples and averaging predictions, which is expected to improve performance without extra storage.Source: ANALYSIS · p. 3
  • Compared with DropStroke, DropSegment better accommodates concatenated strokes, languages with few strokes, and generation of more possible samples.When all segments of a stroke are dropped, DropSegment is equivalent to DropStroke.Source: ANALYSIS · p. 3

3. GENERATION OF PATH-SIGNATURE FEATURE MAPS

DeepWriterID introduces path-signature feature maps to extract further information for handwriting identification. Combining signature levels enriches the path representation while limiting feature-computation time.

  • Path signatures represent finite-length pen trajectories through iterated integrals; levels 0, 1, and 2 denote the input, path displacement, and path curvature, respectively.The k-th level has dimension 2k.Source: §3 · p. 3
  • Combining different integral-iteration levels enriches path representation while keeping feature-computation time to a minimum.Source: §3 · p. 3
  • At truncation level n, each sample point along a segment generates a set of signature values whose dimension determines the number of feature maps.Source: §3 · p. 3
  • Beyond the second iterated-integral level, feature maps show appreciable differences among writers, supporting the hypothesis that higher levels may contribute positively to handwriting identification.Source: §3 · p. 4

4. ARCHITECTURE AND CONFIGUATION OF DEEP CNN

The system uses a deep convolutional neural network designed to exploit the spatial sparsity of handwriting. Signature feature maps serve as input channels to the network.

  • The DCNN has five convolutional layers, each followed by max-pooling, and two fully connected layers with 480 and 512 units.It uses 3×3 filters in the first convolutional layer, 2×2 filters in the others, and a 2×2 max-pooling window.Source: §4 · p. 4
  • The network exploits handwriting’s spatial sparsity: foreground trajectory pixels are scarce relative to zero-valued background pixels, allowing background computation to be avoided.Source: §4 · p. 4
  • The input is a 54×54 bitmap centered on a 96×96 grid, and the number of input channels equals the number of signature feature maps.Source: §4 · p. 4

5. EXPERIMENTAL RESULTS

Experiments evaluate DeepWriterID on Chinese and English NLPR subsets, testing path-signature features and DropSegment. The system surpasses previously reported NLPR identification rates using only pen-position information.

DATABASE

  • DB I comprises 187 writers with two free-content Chinese pages for training and one fixed-content page for testing; DB II uses the same page split for 134 English writers.Experiments deliberately used pen-position information in the pen-down state, although the NLPR samples include additional signals.Source: DATABASE · p. 4

DATA PREPROCESSING

  • Preprocessing divides strokes into segments at local bending maxima, then groups segments into approximately square pseudo-characters without requiring meaningful character boundaries.The paper sets the input height-to-width ratio to 1 and reports similar results for ratios from 1 to 1.3.Source: DATA PREPROCESSING · pp. 4–5

DROPSEGMENT IMPLEMENTATION

  • Training randomly removes segments from characters and recombines remaining fragments into longer segments or new strokes, supporting augmentation while retaining identity information.The implementation limits removals to at most half of a character’s segments per iteration to avoid excessive information loss.Source: DROPSEGMENT IMPLEMENTATION · p. 5

INVESTIGATION OF PATH-SIGNATURE FEATURE

  • Path-signature features substantially improved the Bitmap baseline on Chinese DB I, while gains became more negligible at higher truncation levels as feature dimensions increased exponentially.Higher truncated levels provide more subtle handwriting information; Figure 4 reports page-level Top-1 test error rates.Source: INVESTIGATION OF PATH-SIGNATURE FEATURE · p. 5, Figure 4 · p. 5

INVESTIGATION OF DROPSEGMENT METHOD

  • 95.72% for Chinese text and 98.51% for English text surpassed prior NLPR bests of 91.50% and 93.60%, respectively, with relative error reductions of 49.6% and 76.5%, using pen-position alone.The comparison is against prior results that mostly used richer online information, including pressure, altitude, azimuth, velocity, and pen-position in pen-down and pen-up states.Source: Figure 5 · p. 6, INVESTIGATION OF DROPSEGMENT METHOD · p. 6
  • All identification rates on Chinese DB I improved with DropSegment; the authors averaged 20 test-time predictions to combine its variable samples without retraining.The table reports average writer identification rates for methods with and without DropSegment on Chinese text.Source: INVESTIGATION OF DROPSEGMENT METHOD · p. 5, Table 1 · p. 6

6. CONCLUSIONS

DeepWriterID combines DropSegment, path-signature features, and a deep CNN, achieving state-of-the-art identification accuracies on two NLPR subsets. The authors identify unknown-writer rejection and new-writer registration as issues for future study.

6. CONCLUSIONS

  • DeepWriterID combines DropSegment, path-signature features, and a deep CNN, and achieves state-of-the-art writer identification accuracies on two NLPR subsets.The conclusion describes path-signature features as useful and effective, and DropSegment as enabling CNN training with inadequate data.Source: §6 · p. 6
  • DropSegment can provide ensemble performance without training or storing additional network models.Source: §6 · p. 6
  • The system does not consider rejection of unknown writers; confidence-based rejection and registration of new writers remain future work.Source: §6 · p. 6
Loading 1508.04945v2…