The Comparative Study of Adding Edge Information to Pix2pix Architecture for Face Image Generation

The image-to-image translation is part of image generation, which aims to transform the source image into the target image. The benefit of the image-to-image translation method in the forensic field is to recognize the face of the perpetrator of a crime through facial sketches. Pix2pix is a deep learning architecture for image-to-image translation that implements the Conditional Generative Adversarial Network (CGAN). Pix2pix can generate images well, although the resulting image is blurry in many cases. Modification of the Pix2pix architecture through the addition of edge information has been reported to improve the sharpness of the resulting image generation. This study aims to compare the performance of the modified Pix2pix architecture using various edge extraction methods, namely Laplacian, Sobel, and Prewitt. The comparison of the performance of the Similarity Structural Index (SSIM) on the Chinese University of Hong Kong (CUHK) faces student dataset shows that the modification of the Pix2pix architecture with the addition of Prewitt edge information produces the highest average value of 81.4%.