The Model Regularized CLAHE-CNN for the Traditional Script Character of Kaganga Bengkulu Recognition

Itis necessary to preserve the Kaganga script to be in demand again, especially by the younger generation today. Using and developing current technology, namely image processing techniques and deep learning, can be one way to reintroduce and preserve the culture of Kaganga script recognition. The purpose of this study is to implement the CNN method in recognizing the Kaganga script, observe the results of the implementation of the $\mathbf{L} 2$ regularization technique in reducing the overfitting process in CNN, and increase the accuracy of Kaganga script recognition using CLAHE and L2 regularization technique in CNN. The collected datasets are divided into test, valid, and training data. To acquire primary data for the Kaganga script, 50 respondents were involved. Each respondent was requested to write a Kaganga script on the provided paper. Organizations engaged in cultural heritage research are gathering data for this project: Yayasan Budaya Nusantara Digital (Jakarta), Yayasan Insani Mandiri Santani (Bengkulu), and Yayasan Sejahtera Bersama (Bengkulu). This study will evaluate using an accuracy measurement model, precision, and recall, which is included in the confusion matrix measurement model. In this research, we conducted four experiments, including CNN with L1, CNN with L2, CNN with dropout and the CLAHE-CNN with best regularization. From the experimental results, CLAHE-CNN with $L 2$ will be found that accuracy during the training stage gets a value of $100 \%$, accuracy during the validation stage gets a value of $85 \%$ and accuracy during the testing stage gets a value of $\mathbf{8 0. 7 5 \%}$