Image-to-Image Orthodontics Transfer Employing Gray Level CO-Occurrence Matrix Loss

Abstract

Orthodontics transfer is a new, challenging image-to-image transfer task from malpositioned-teeth images to neat-teeth images. More specifically, it belongs to the image-to-image location transfer, which aims to rearrange the chaotic foreground objects into an orderly layout. In this paper, we conducted further research on image-to-image orthodontics transfer task. Firstly, we studied the similarities and differences between malpositioned-teeth images and their corresponding neat-teeth images and found texture feature similarities between them. Then we analyzed the problems of directly applying the LTGAN method to the orthodontics transfer task and proposed an approach based on the boundary label transfer bridge. Finally, our model’s performance is further improved by employing gray level co-occurrence matrix loss. Moreover, we have augmented the OrthoD datasets to support our method and potential attempts to deal with orthodontics transfer task. The added data is available at https://drive.google.com/drive/folders/1bzoxzi_608SzCVgaABlPAjZVqp6pp7L8?usp=sharing.