Forensic sketch-to-photo transformation with improved Generative Adversarial Network (GAN)

Forensic sketch-to-photo transformation with improved Generative Adversarial Network (GAN)

연구 분야: Artificial Intelligence

학회: 2022 5th International Conference on Multimedia, Signal Processing and Communication Technologies (IMPACT)

Many of the limitations of Convolutional Neural Networks started to become apparent with the development of "Capsule neural networks", and more research is now being done on Capsule Networks since it is believed that it will be the state-of-the-art for image classification tasks in the future. In this paper, the discriminator inside a GAN for face sketch to photo translation employs capsule neural network layers rather than the traditional convolutional layers. Incorporating the Capsule design into the Discriminator will improve the classification loss and speed up the convergence compared to using traditional convolutions. With 188 hand-drawn sketch-photo pairings available in the CUHK dataset and constructed 1000 sketch-photo pairs of CelebA HQ dataset, our proposed method is evaluated. The results of the experiments demonstrate that the proposed approach can generate images that are visually extremely near to the real photos and have rich texture information and features. The translation of face sketch to photo can greatly help the law enforcement agencies to track and arrest the criminals.