Pix2Pix++: An Enhanced GANs Based Model for Portrait to Pencil Sketch Translation

The generation of sketches from portraits has been an active area of research for decades due to its applications in various sectors such as entertainment and education. However, due to numerous qualitative factors involved in sketching, programmatically generating sketches from images has been a challenging task. Recent advances in generative artificial intelligence (GAI) seem very promising in solving these challenges. In specific, generative adversarial networks (GANs) like conventional Pix2Pix, were specifically designed for accurate and efficient image-to-image translation tasks. In this research, inspired by GAI and original Pix2Pix, we proposed a GANs-based Pix2Pix++ technique for translating portrait images into sketches. We replaced U-Net with the UNet++ for generator in Pix2Pix++ GAN and trained the proposed model on a self-collected dataset for sketch generation. Furthermore, we evaluated the proposed model's performance both qualitatively and quantitively and performed a comparative analysis with existing techniques. Proposed Pix2Pix++ GAN got MS-SSIM median scores of 84%, PSNR of 21.17, and MSE median score of 495.55. Experimental results showed our proposed technique outperforms the existing techniques in terms of capturing minute details, tonal variations, realistic appearance, and generalization.