Integrating Generative Adversarial Networks and Clustering Techniques for Automated Screen Printing

In the textile screen printing industry, manual color separation as an age-long practice and the perfection of the image generation process are the main obstacles to enhancing high print fidelity, as there is a lack of professionals able to carry out color separation and deal with related implications in the industry. This research proposes an innovative approach, which features the application of Generative Adversarial Networks Pix2PixHD with Kmeans clustering method to enhance and accelerate the color separation process. The Pix2PixHD architecture is responsible for generating realistic outputs in high resolution, while the Kmeans allows for the separation of all similar colors in the image into different branches. This research investigated the process and outcomes of textile screen printing, focusing on the automation of fabric selection, pattern correction, and color separation techniques to enhance textile prints' quality. Two distinct approaches-one with augmentation and the other without-were employed to implement the proposed methodology during the preliminary phase of this study. The augmented Pix2PixHD+Kmeans model provided better results in terms of multiple performance measures. Based on our experimental result, the proposed Pix2PixHD+Kmeans model can increase Structural Similarity Index Measure (SSIM), IoU, and Pixel Accuracy (Acc) by 2.33%, 2.47%, and 1.18% correspondingly, and decrease Learned Perceptual Image Patch Similarity (LPIPS) by 1.76% (the lower is better). These results indicate that the augmented Pix2PixHD+Kmeans approach can reinvigorate traditional screen printing techniques and represent a more precise, efficient, and less manpower-dependent future for the industry.