Batik Sketch Coloring Using Generative Adversarial Network Pix2pix

Abstract

Batik, an Indonesian cultural heritage recognized by UNESCO, involves a complex and time-consuming coloring process. Digitalization offers a crucial solution for the preservation and development of batik art in the modern era. This research implements a Generative Adversarial Network (GAN), specifically the Pix2Pix model, to automate the transformation of batik sketches into colored images. The primary focus is a performance comparison between the U-Net generator architecture, which excels at preserving spatial details via skip-connections, and the ResNet architecture, which is capable of learning deeper and more complex features. This study utilized 1164 paired images, divided into 931 training, 117 validation, and 116 test data points. The models were trained with consistent hyperparameters, including an Adam optimizer and a combination of L1 and binary cross-entropy loss functions, with evaluations at 50 and 100 epochs. Quantitative evaluation was performed using Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Fréchet Inception Distance (FID) metrics. The results indicate that the model with the ResNet generator trained for 100 epochs achieved the most balanced and superior performance, with a PSNR of 8.11, SSIM of 0.39, and an FID of 120.72. Overall, the ResNet generator proved more capable of producing realistic and high-quality colored batik images, offering an innovative solution to enhance the efficiency of the coloring process while supporting cultural preservation.