A skull-face translation network used for generating faces from skulls

Reconstructing the antemortem appearance of unknown human skulls, referred to as craniofacial reconstruction, is an important research topic in forensic science and criminal investigation. In response to the high data requirements, tedious acquisition steps, high storage costs, high demands for network design and training, and low accuracy of three-dimensional craniofacial reconstruction methods, this paper proposes an automated and intelligent craniofacial translation network called Skull2Skin-GAN. The generator replaces ordinary convolutions with depthwise separable convolutions based on U-Net to reduce model volume. The discriminator adopts random non-permeable data augmentation to adapt to this network structure. The SCUT-FBP5500 facial beauty prediction Dataset is introduced to calculate fine feature-aware losses. Finally, the GFPGAN was introduced to perform super-resolution reconstruction on the generated faces, restoring high-quality faces with rich geometric shape, texture, color from low-quality human faces. Experimental results show that under small samples, this network can quickly and easily generate clear and high-fidelity facial images from skull images, and can identify identity.