A Comparison of VGG Architecture Convolutional Layers in Migrating Batik Style into Fractal Shape

The neural style transfer algorithm is created to produce high-quality artistic images. Many studies have been performed to advance its implementations. However, the application of neural style transfer in the Indonesian batik industry was not studied often. This paper performed a comparison between the VGG16 and the VGG19 neural network in migrating batik-style images into fractal shapes. This study also proposes some post-processing methods used to improve the quality of the generated images. Fractal images and batik images were utilized as the content and style images respectively for this experiment. Pre-trained VGG16 and VGG19 neural networks were used to perform style transfer in this study. Methods popular in image processing were utilized to augment the artistic value of the generated images. The process yields several results aside from the generated images, such as records of computation time, records of loss value changes, and sequences of style transfer iteration. A survey was also conducted to gather opinions regarding the produced images. The result of this experiment was analyzed with some hypothesis tests and visualized with graphs. This study finds that the VGG16 model mainly performs faster than the VGG19 model. Both models can produce batik-style transfer images with identical results. This study also finds that the convolutional layers “conv2_1” and “conv3_1” of both architectures are the best layers to be implemented in storing batik-style information. The proposed post-processing methods are found to be decent in bettering the artistic value of batik-style transfer images.