CNN Classifier Parameter Optimization with Genetic Algorithms: A Case Study of Indonesian Batik Patterns

Abstract

Batik is one of Indonesia’s most famous and commonly encountered cultural heritage sites. Batik is a type of Indonesian specialty cloth where the pattern is drawn on the cloth using wax. The number of batik pattern varieties produced by Indonesia introduces challenges in recognizing and differentiating between batik patterns. Various previous works have been done to research how to use computer vision to recognize and differentiate between batik patterns. Some of the methods used in previous works involve convolutional neural networks; previous studies use pre-trained models or manually design the model, which may not be suitable for recognizing batik patterns. This study focuses on developing a convolutional neural network model optimized and designed automatically using genetic algorithms using more unique images than previous works. Genetic algorithms have been proven in previous works from other fields to produce better models compared to pre-trained and manually designed models. The experiment results show that a model design using genetic algorithms outperforms pre-trained models by a significant margin. The model automatically achieved an accuracy of 0.8654 with a parameter ∼1% of the VGG-19 model, whereas the VGG-19 model achieved an accuracy of 0.7596.