为解决图像修复中存在的伪影以及细节表达不一致等问题,提出一种基于注意力机制的对抗型自编码器(adversarial auto-encoder based on attention mechanism,AAEA)图像修复模型.在通用编码器模型基础上,通过在生成器跳跃连接处引入通道注意力构造通道相似性融合模块(channel similarity fusion module,CSFM)的方式使通道之间的特征关系更丰富.在解码器网络中,通过空间注意力与位置编码结合的方式构造位置融合模块(location fusion model,LFM),增强边界位置信息表达.消融实验结果表明,引入的CSFM、LFM模块均能有效提升模型性能,阈值为1.253 时的准确率达到 0.9808.AAEA模型能够更好地处理复杂的图像修复任务,有效地改正错乱纹理,并在掩膜边缘区域获得清晰的修复结果,对于壁画修复以及计算机视觉等图像修复领域的发展具有重要意义.
An Image Restoration Model of Adversarial Auto-encoder Based on Attention Mechanism
In order to solve the problems of artifacts and inconsistent details in image inpainting,an adversarial auto-encoder based on attention mechanism(AAEA)image restoration model was proposed.Based on the universal encoder model,the channel similarity fusion module(CSFM)was constructed by introducing channel attention at the generator jump connection,which enriched the feature relationship between channels.In the decoder network,a location fusion model(LFM)was constructed by combining spatial attention with location coding to enhance the expression of boundary location information.The results of ablation experiments showed that after introducing CSFM and LFM,the performance of the model is effectively improved,and the accuracy reached 0.9808 with the threshold of 1.253.The AAEA model could better deal with complex image restoration tasks,effectively correct disordered textures,and obtain clear restoration results in the edge region of the mask,which is of great significance for the development of mural restoration and computer vision and other image restoration fields.
万方数据
维普
