Skull-to-skin generation method based on multi-level bottleneck attention module

Restoring the face from an unknown skull is an important research direction in archaeology, forensics and criminal investigation. The existing computer-aided 3D restoration process is cumbersome and time-consuming. In view of the distortion, twisting and non-smoothness of the existing model in the generation of skull-to-skin (face without texture and hair, etc.) images, this paper proposes a skull-to-skin image generation method combining a generative adversarial network and a multi-level bottleneck attention module. Specifically, the generator consists of six layers of AdaResBlock and a bottleneck attention module, which guides the generator to focus on more important areas from the two dimensions of channel and space, and adjusts the normalization method according to the feature adaptiveness. At the same time, in order to solve the problem of the large size of the generator model, the blueprint separable convolution is introduced to reduce its volume. In addition, the discriminator is divided into two parts. The first few layers are used for encoding, eliminating the separate encoder module in the traditional network, making the model more compact; the latter layers adopt a multi-scale discrimination strategy to classify and discriminate images from different levels to enhance their accuracy. Experimental results show that in the task of skull-to-skin image generation, the skin images generated by this method have higher quality than other existing methods, and have achieved the highest scores in both visual quality and image quality. The restoration effect is more realistic, and the image quantitative evaluation indicators PSNR and SSIM are improved by an average of 1.115 and 0.017, and LPIPS is reduced by an average of 0.026. The average facial similarity is 0.855.