GENERATIVE CONVOLUTIONAL NEURAL NETWORK LEARNING WITH SIMPLE DENOISING ON GENERATING BATIK IMAGES

Abstract

Batik is a traditional textile from Indonesia, and automatic batik pattern generation can help enrich its diverse motifs. This research explores the use of a generative convolutional neural network, a generative random field model that learns its own filters directly from training data without relying on pre-trained features. The model parameters are estimated by minimizing the difference between the synthesized images and the training images. To improve image quality, k-means clustering and block-matching and three-dimensional (BM3D) filtering denoising are incorporated to reduce noise and enhance the Fréchet Inception Distance score. Two experiments were conducted to evaluate the model. The first tested its ability to generate batik patterns from single images using two datasets: clean ITB-mBatik images and noisier internet-sourced images. The second generated blended patterns from pairs of internet images with similar color and texture. The results show that applying a simple denoising step improves the Fréchet Inception Distance score for the clean ITB-mBatik images, while showing little to no benefit for the noisier internet-sourced images. Overall, the generative convolutional neural network can produce attractive blended batik images, especially when the two original training images are carefully chosen so that their patterns and colors complement each other.