Gender Classification in Korean Handwriting Using EfficientNetV2S: Performance Analysis Across Word Length and Age Groups

Existing studies on handwriting-based gender classification have utilized traditional convolutional neural network models, ensemble learning techniques, and deep convolutional networks like DenseNet201.However, these methods have encountered computational efficiency and performance optimization challenges.This paper proposes a new approach using the EfficientNetV2S model to address these issues.The proposed method involves preprocessing handwritten data and applying the EfficientNetV2S model, SAM optimizer, and a learning scheduler to classify gender based on handwriting.In the preprocessing steps, images are converted to grayscale, binarized using Otsu's method to extract word objects, and then normalized to a specified size according to the number of characters.In this study, gender classification accuracy was compared based on character count using a dataset of individuals in their 20s and based on age groups using a dataset of three-character words.The results show that, for the 20s dataset, the highest accuracy of 79.88% was achieved with three-character words.Additionally, in gender classification experiments across age groups from teenagers to 50s, accuracy improved up to the 30s group, reaching 97.29%, but then showed a slight decline in the 40s and 50s groups.The method proposed in this paper shows that gender can be classified with better accuracy than in previous studies and can be used as supplementary data in author attribution analysis.