Combination of Pre-Trained CNN Model and Machine Learning Algorithm on Pekalongan Batik Motif Classification

Abstract

Pekalongan is a region in Indonesia well-known for its batik production. The Pekalongan batik is rich in varieties of motifs, such as the Jlamprang, Liong, Terang Bulan, and Tujuh Rupa. The difficulty of distinguishing Pekalongan batik motifs for ordinary people causes the need for a model that can help recognize these motifs automatically based on input from digital images. This research aims to classify the Pekalongan batik motifs using a pre-trained Convolutional Neural Network (CNN), the Inception V3, and machine learning, the K-Nearest Neighbors (K-NN) algorithm. First, we extract the features from the digital image using the Inception V3 model, resulting in m x 2048 features, where m is the number of images. The extracted features generated from the Inception V3 model will be used as the dataset for the motif classification. We build models to classify the features using the K-Nearest Neighbors (KNN) with a K value of 5. In the classification process, we employ two distance metrics, the Euclidean and Manhattan distance, and analyze their performance using the 10-fold and 20-fold crossvalidation. The results of this study are the highest overall performace of accuracy (0.987), precision (0.987), and recall (0.987) produced by the Euclidean model.