Analisis Komparasi Convolutional Neural Network dan Learning Vector Quantization dalam Klasifikasi Khat Arab Digital

Abstract

Arabic khat is a form of writing that possesses complex visual characteristics, such as variations in letter shapes, stroke thickness, texture, and stylistic differences. This complexity creates challenges in manually recognizing different types of khat. This study aims to analyze and compare the performance of Convolutional Neural Network (CNN) and Learning Vector Quantization (LVQ) methods in classifying five types of Arabic khat digital images, namely Diwani, Farsi, Naskh, Ruqaa, and Tuluth. The dataset was obtained from the Kaggle.com platform. CNN architecture consists of an input layer of 100×100×1, followed by two convolutional layers with 32 and 64 filters of size 3×3, each followed by ReLU activation and max pooling with stride 2. The network then includes a fully connected layer with 64 neurons, a final fully connected layer corresponding to the number of classes, a softmax layer, and a classification layer. CNN training was conducted using 5-fold cross-validation, applying data augmentation in each fold. For the LVQ method, Local Binary Pattern (LBP) was used for feature extraction from 100×100 images with parameters: radius 1, 8 neighbors, cell size [48 48], and L2 normalization. The extracted features were used for training with an initialization of 25 prototypes from 5 classes. The process also employed 5-fold cross-validation. From 40 testing samples, the CNN model achieved an accuracy of 87.5%, while the LVQ model achieved an accuracy of 85%. The CNN algorithm demonstrated better performance in handling the complex visual patterns of Arabic khat. Meanwhile, LVQ showed advantages in architectural simplicity and computational efficiency. This research is expected to contribute to the development of Arabic khat image classification systems and serve as a reference in selecting optimal methods for Arabic khat recognition.