[3Xin4] Poster session 1

[3Xin4-73] The feasibility test of writer identification of handwriting by Alex Net transfer learning.

〇Shigeru Sugawara¹ (1.National Research Institute of Police Science)

Keywords:Authorship identification of handwriting., transfer learning

In handwriting identification, it is considered desirable to use both human-performed identifications and identification by image processing and other methods. In this study, an attempt was made to identify the writer of a character image using transition learning of a learned model. Among the learned models capable of image recognition, the AlexNet with the shortest computation time was used. The sample used was 401 handwriting images of the character 'Go', one of the characters with the highest number of strokes, from the handwriting database; there were five character images per writer, and the image size was 400 x 400 pixels. Matlab 2022b was used for deep learning; Alex Net was used, and the maximum number of epochs was set to 60, with 20 iterations per epoch. As a result, the model could accurately estimate the author for 2004 of the 2005 letters (= 401 respondents x 5 times). For the remaining one character, the accurate answer had the second highest probability. The results suggest the possibility of writer estimation by trained models using transition learning. In the future, we would like to carry out validation using validation data, validation when the number of characters is increased, and validation of models created with other characters with other characters.