محمد حسين احمدي

Nowadays, with the widespread use of cameras across various domains and the significant increase in the availability of facial images, their processing has gained considerable importance. These process can be either done automatically by a machine or manually by a human agent. In both scenarios, the image quality and its complexity are crucial factors that can significantly impact the final result. The variation in pose is an important factor contributing to the complexity of face images. To tackle this challenge, given the significant progress of deep learning models in recent years, many approaches have been proposed for face frontalization. However, these approaches employ heavy architectures and involve complicated training procedures. Moreover, they struggle to reconstruct the frontal-view image with high quality. Another drawback is that, despite the availability of multiple images as frames in a video— which inherently contain more information— these methods reconstruct the frontal-view image by processing only a single input image. In this study, we aim to address all of the above challenges by proposing an approach that has fewer training challenges and is capable of reconstructing the frontal-view image with photorealistic quality. For this purpose, we first address the unavailability of an appropriate dataset for processing sequences of images by presenting an innovative solution to provide a diverse dataset with high quality. Furthermore, we proposed a novel single-input method for face frontalization. In this method, we incorporate the GAN inversion technique with a transfer learning approach to leverage their advantages for reconstructing the frontal-view image with significantly high quality while maintaining an efficient number of trainable parameters. Following this, by employing and modifying this method, we have proposed a new approach for processing sequences of images. This method is capable of leveraging both the independent and collective insights of frames to reconstruct the frontal-view image. Lastly, through a quantitative and qualitative analysis of our methods and comparing them with previous approaches, we demonstrate their effective capability to reconstruct frontal-view images with photorealistic quality. Lastly, through a comprehensive quantitative and qualitative analysis of our methods and a comparative eva‎luation with previous approaches, we demonstrate their superiority, reflected in a 0.15 improvement in the LPIPS metric and a 0.12 improvement in the distance of identity features.

شبكه ي مولد رقابتي , وارون گن , روبه رو سازي چهره , ترجمه ي تصوير به تصوير , يادگيري عميق

Generative adversarial networks , GAN Inversion , Face frontalization , Image-to-image translation , Deep Learning

Mohmmad Hossein Ahmadi

Mohammad Reza Mohammadi