Summary
2023
Session Number:C2L-1
Session:
Number:C2L-15
Writer Identification Method That Is Able to Focus on Local Writing Habits by Attention Mechanism
Ishimori Ami, Arai Shuichi,
pp.459-462
Publication Date:2023-09-21
Online ISSN:2188-5079
Summary:
Writer identification is a task to guess who wrote the letters using inherent writing habits. Conventionally, personal characteristics were defined in advance by experts. In recent years, new methods have emerged to identify authors by learning personality traits with deep neural networks. However, those methods had no constraints at all regarding the discovery and learning of individuality features, and it was unclear what part of the text image the discriminator was looking at to identify the writer. Since Japanese characters show writing habits in fine details, it is a big problem if the part that the　architecture is focusing on is unknown. Therefore, we propose to construct a writer identifier that can capture even local writing habits by introducing an attention mechanism. The attention mechanism we employed is Attention Branch Network. We experimented with where to insert the attention mechanism and placed it after the second convolution layer. As a result, the discrimination rate improved by 1.6%, from 78.8% to 80.4%.