Exploring ReLU Activation Functions in CNN for Handwritten Sundanese Script Recognition

Recognizing the handwritten Sundanese script holds significant importance in the preservation of Indonesian regional scripts. While traditional Optical Character Recognition (OCR) methods have been utilized, deep learning approaches have shown promise in enhancing handwritten character recognition. Convolutional Neural Networks (CNNs) have emerged as a prominent choice due to their efficacy in computer vision tasks. Various CNN architectures have achieved state-of-the-art results in handwritten character recognition. However, selecting activation functions remains a crucial aspect influencing CNN performance. Despite extensive research on activation functions in handwritten character recognition, there exists a gap in understanding which activation functions are optimal for training models specifically tailored to handwritten Sundanese script recognition. This study aims to address this gap by exploring the efficacy of Rectified Linear Unit (ReLU) activation functions, including Leaky ReLU, Randomized Leaky ReLU (RLReLU), and Optimized Leaky ReLU (OLReLU). In training Sundanese script OCR models, the utilization of OLReLU surpasses other activation functions. While OLReLU marginally lags behind RLReLU quantitatively on validation data, OLReLU maintains its qualitative superiority among other functions on test data, achieving the highest accuracy of 0.9875. This comparative analysis underscores the efficacy of OLReLU in the context of handwritten Sundanese script recognition, offering valuable insights for the development of robust OCR systems tailored to this regional script.