In this thesis, we propose a framework to solve the blind face restoration that
<br>recover a high-quality face image from unknown degradations. Previous methods
<br>have shown that the Vector Quantization (VQ) codebook can be powerful prior
<br>to solve the blind face restoration.
<br>However, it is still challenging to predict code vectors from low-quality im-
<br>ages. To solve this problem, we propose a multi-scale transformer consisting of
<br>multi-scale cross-attention (MSCA) blocks. The multi-scale transformer com-
<br>pensates for lost information of high-level features by globally fusing low-level
<br>and high-level features with different spatial resolutions.
<br>Also, there is a trade-off problem between pixel-wise fidelity and visual qual-
<br>ity of the results. To improve the fidelity of the results, we employ shifted win-
<br>dow cross-attention modules at multiple scales. The shifted window method can
<br>not calculate inter-window attention to model the abundant facial global con-
<br>text. To solve this problem, we propose a shifted window token cross-attention
<br>module SW-TCAFM with a global class token to model the global context of
<br>face. The global class token models the global context by aggregating informa-
<br>tion across all windows and passing it to the next step. In addition, we propose a
<br>semantic token regularization loss that makes each global class token represents
<br>a specific face component by utilizing the face parsing map prior.
<br>Our framework achieves superior performance in both quality and fidelity
<br>compared to state-of-the-art methods. In our experiments, we show that the
<br>PSNR and FID results of our framework are better than 3.21% and 2.92%,
<br>respectively, compared to state-of-the-art method.
