Optimasi Convolutional Neural Network melalui Fungsi Aktivasi dan Inisialisasi Kernel untuk Pengenalan Tulisan Huruf Hijaiyah

Nasty, Khairuddin (2025) Optimasi Convolutional Neural Network melalui Fungsi Aktivasi dan Inisialisasi Kernel untuk Pengenalan Tulisan Huruf Hijaiyah. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025201041-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only until 1 April 2027. Download (2MB) | Request a copy

Abstract

Penelitian ini mengoptimalkan model Convolutional Neural Network (CNN) untuk pengenalan tulisan tangan huruf Hijaiyah dengan menganalisis kombinasi fungsi aktivasi (ReLU, Leaky ReLU, Sigmoid, Tanh) dan metode inisialisasi kernel (He normal, He uniform, LeCun normal, LeCun uniform, Glorot normal, Glorot uniform). Dataset yang digunakan adalah Hossam Magdy Balaha Dataset (HMBD) yang dimodifikasi—dengan penambahan tanda baca fathah, kasrah, dan dhammah—untuk mengevaluasi 24 kombinasi parameter. Hasil eksperimen menunjukkan bahwa kombinasi He normal-ReLU mencapai performa terbaik dengan akurasi 93,84%, presisi 93,96%, recall 93,77%, dan F1-score 93,71%. Analisis konvergensi mengungkapkan kombinasi ini stabil setelah epoch ke-10, dengan validation loss di bawah 0,5, serta fluktuasi akurasi kurang dari ±1%. Kesalahan klasifikasi tertinggi terjadi pada pasangan dengan kemiripan visual, yaitu Ain fathah-Haa fathah (14,63%) yang diidentifikasi melalui confusion matrix. Konversi model ke TensorFlow Lite berhasil mengurangi ukuran dari 29,4 MB menjadi 9,8 MB (66,67%) tanpa penurunan performa, dengan akurasi tetap 93,84%. Temuan ini membuktikan bahwa optimasi inisialisasi kernel dan fungsi aktivasi secara signifikan meningkatkan akurasi dan efisiensi model, sekaligus memberikan panduan implementasi CNN untuk aplikasi mobile edukasi berbasis tulisan tangan non-Latin.
==================================================================================================================================
This study optimized a Convolutional Neural Network (CNN) model for recognizing handwritten Hijaiyah characters by analyzing combinations of activation functions (ReLU, Leaky ReLU, Sigmoid, Tanh) and kernel initialization methods (He normal, He uniform, LeCun normal, LeCun uniform, Glorot normal, Glorot uniform). The modified Hossam Magdy Balaha Dataset (HMBD)—augmented with diacritical marks (fathah, kasrah, dhammah)—was used to evaluate 24 parameter combinations. Experimental results demonstrated that the He normal-ReLU combination achieved the best performance, with 93.84% accuracy, 93.96% precision, 93.77% recall, and 93.71% F1-score. Convergence analysis revealed this combination stabilized after the 10th epoch, with validation loss below 0.5 and accuracy fluctuations within ±1%. The highest misclassification rate (14.63%) occurred between visually similar pairs, specifically Ain fathah-Haa fathah, as identified through confusion matrices. Model conversion to TensorFlow Lite successfully reduced its size from 29.4 MB to 9.8 MB (66.67%) without performance degradation, maintaining 93.84% accuracy. These findings prove that optimizing kernel initialization and activation functions significantly enhances model accuracy and efficiency, while providing a practical guideline for implementing CNNs in mobile-based educational applications for non-Latin handwriting recognition.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Convolutional Neural Network, Fungsi Aktivasi, Inisialisasi Kernel, Pengenalan Huruf Hijaiyah, Konvergensi Model, TensorFlow Lite. Activation Functions, Convolutional Neural Network, Hijaiyah Character Recognition, Kernel Initialization, Model Convergence, TensorFlow Lite.
Subjects:	T Technology > T Technology (General) > T57.5 Data Processing T Technology > T Technology (General) > T57.8 Nonlinear programming. Support vector machine. Wavelets. Hidden Markov models.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Khairuddin Nasty
Date Deposited:	03 Feb 2025 02:23
Last Modified:	03 Feb 2025 02:23
URI:	http://repository.its.ac.id/id/eprint/117839

Actions (login required)

View Item