Accurate Detection of Handwriting Features Using Bounding Box Optimization

Abstract

Handwriting is a unique human expression that provides insights into personality traits and behavioral tendencies. Automatic detection of handwriting features remains challenging due to variations in size, shape, writing style, and inconsistencies caused by environmental, physical, and psychological factors. These challenges are even greater in multi-class datasets. This study employed 150 handwriting images annotated with feature classes, including inter-word spacing (narrow-SKS, medium-SKN) and baseline variations (up-BN, down-BT, wavy-BG), representing key traits of written expression. The research highlights the role of parameter optimization through hyperparameter tuning, covering epochs, batch size, learning rate, momentum, image resolution, and weight decay. Such optimization is crucial to enhance detection accuracy and improve the robustness of handwriting feature recognition models. Among the ten experimental configurations evaluated, the proposed model in the fifth run achieved a recall of 0.98, demonstrating strong sensitivity in detecting handwriting features. However, this high recall was accompanied by relatively low precision, as reflected in the F1-Score of 0.55, indicating the presence of false positives. This trade-off highlights both the effectiveness and limitations of the current approach and underlines the importance of hyperparameter optimization, bounding box configuration, and dataset structuring in improving handwriting feature detection outcomes.