%0 Journal Article
%T 基于掩码Transformer的图像修复网络<br>An Image in Painting Model Based on Mask Transformer
%A 康延亭
%A 王直杰
%J Computer Science and Application
%P 83-94
%@ 2161-881X
%D 2022
%I Hans Publishing
%R 10.12677/CSA.2022.121010
%X <div style=\"text-align:justify;\">
	<span style=\"font-size:14px;\">现有的基于深度学习的图像修复网络通常采用注意力机制以相似匹配的方式将完好区域信息填充到待修复区域来提升待修复区域的纹理细节。然而，现有的注意力机制的度量方式仅考虑特征纹理而缺少对语义的理解以至于会利用到一些语义不相似区域的信息。为了解决这一问题，本文提出一种基于掩码transformer的图像修复网络，该掩码transformer模块相较于基本的transformer层的区别主要包括两部分：1) 通过掩码将特征图分为有效区域和无效区域并提出掩码注意力机制有效的建模待修复区域和完好区域的相似性；2) 提出用查询集和相似度矩阵加权融合的方法为待修复区域精确填充信息。与传统的注意力机制相比，基于transformer的方法能够较为有效的提升修复的纹理效果。<br />
Existing deep learning-based image repair networks usually use an attention mechanism to fill intact area information into the area to be repaired in a similar matching manner to improve the texture details of the area to be repaired. However, the existing measurement method of attention mechanism only considers the feature texture and lacks the understanding of semantics, so that it will use the information of some semantically dissimilar regions. In order to solve this problem, this paper proposes an image restoration network based on mask transformer. The difference between the masked transformer module and the basic transformer layer mainly includes two parts: 1) The feature map is divided into valid regions and invalid regions by mask, and the mask attention mechanism is proposed to effectively model the similarity between the regions to be repaired and the intact regions; 2) A method of weighted fusion of query set and similarity matrix is proposed to accurately fill in information for the region to be repaired. Compared with the traditional attention mechanism, the transformer-based method can effectively improve the texture effect of repair.</span> 
</div>
%K 掩码，注意力机制，Transformer，查询集，相似度矩阵<br>Mask
%K Attention
%K Transformer
%K Query Set
%K Similarity Matrix
%U http://www.hanspub.org/journal/PaperInformation.aspx?PaperID=48215