Name
Abstract | ||
|---|---|---|
Although great success has been achieved in online handwritten Chinese text recognition (OLHCTR), most existing methods based on over-segmentation or long short-term memory are inefficient and not parallelizable. Moreover, n-gram language models and beam search algorithm were commonly adopted by many existing systems as a part of post-processing, resulting in extremely low speed and large footprint. To this end, we propose a fast, accurate and compact approach for OLHCTR. The proposed method consists of a global and local relationship network (GLRNet) and a Transformer-based language model (TransLM). A novel feature extraction mechanism, which alternately learns global and local dependencies of input trajectories, is proposed in GLRNet for the recognition of online texts. Based on the output of GLRNet, TransLM captures contextual information through Transformer encoder and further improves the recognition accuracy. The recognition and language modelling are always treated as two separate parts. However, the two components of our methods are jointly optimized, which ensures the optimal performance of the whole model. Furthermore, the non-recurrence design improves the parallelization and efficiency of our method, and the parameterized TransLM avoids the large footprint to store the probabilities of n-grams. The experiments on CASIA-OLHWDB2.0-2.2 and ICDAR2013 competition dataset show that our method achieves state-of-the-art performances with the fastest speed and the smallest footprint. Especially in the situation with language model, our method exhibits 2 times to 130 times acceleration compared with existing methods. |
Year | DOI | Venue |
|---|---|---|
2021 | 10.1007/978-3-030-86334-0_11 | DOCUMENT ANALYSIS AND RECOGNITION, ICDAR 2021, PT III |
Keywords | DocType | Volume |
Online handwritten Chinese text recognition, Language model, Transformer | Conference | 12823 |
ISSN | Citations | PageRank |
0302-9743 | 0 | 0.34 |
References | Authors | |
0 | 8 | |
Name | Order | Citations | PageRank |
|---|---|---|---|
| Dezhi Peng | 1 | 1 | 3.05 |
| Canyu Xie | 2 | 0 | 0.68 |
| Hongliang Li | 3 | 0 | 1.01 |
| Lianwen Jin | 4 | 1337 | 113.14 |
| zecheng xie | 5 | 96 | 7.55 |
| Kai Ding | 6 | 138 | 16.64 |
| Yichao Huang | 7 | 0 | 1.69 |
| Yaqiang Wu | 8 | 5 | 2.75 |