Chinese Inertial GAN for Writing Signal Generation and Recognition

Yifeng Wang, Yi Zhao

Keywords: Inertial Sensor, Signal Generation, Generative Adversarial Network, Chinese Character, Writing Recognition

Abstract: Disabled people constitute a significant part of the global population, deserving of inclusive consideration and empathetic support. However, the current human-computer interaction based on keyboards may not meet the requirements of disabled people. The small size, ease of wearing, and low cost of inertial sensors make inertial sensor-based writing recognition a promising human-computer interaction option for disabled people. However, accurate recognition relies on massive inertial signal samples, which are hard to collect for the Chinese context due to the vast number of characters. Therefore, we design a Chinese inertial generative adversarial network (CI-GAN) containing Chinese glyph encoding (CGE), forced optimal transport (FOT), and semantic relevance alignment (SRA) to acquire unlimited high-quality training samples. Unlike existing vectorization focusing on the meaning of Chinese characters, CGE represents the shape and stroke features, providing glyph guidance for GAN to generate writing signals. FOT constrains feature consistency between generated and real signals through the designed forced feature matching mechanism, meanwhile addressing GANs' mode collapse and mixing issues by introducing Wasserstein distance. SRA captures the semantic relevance between various Chinese glyphs and injects this information into the GAN to establish batch-level constraints and set higher standards of generated signal quality. By utilizing the massive training samples provided by CI-GAN, the performance of six widely used classifiers is improved from 6.7\% to 98.4\%, indicating that CI-GAN constructs a flexible and efficient data platform for Chinese inertial writing recognition. Furthermore, we release the first Chinese writing recognition dataset based on inertial sensors in GitHub.

Primary Area: Generative models

Submission Number: 10532