基于非对称卷积块和多尺度融合判别的图像转换

Image-to-image Translation with Asymmetric Convolution Blocks and Multi-scal Fusion Discrimination

在图像转换过程中往往会出现图像模糊,细节丢失的问题,针对这些问题,文章设计了一种新的图像转换模型,用来减少图像模糊、丢失问题。文章在生成器中引入非对称卷积块,增强特征提取,通过上采样增加细节信息,判别器中使用一种新的多尺度融合方法,增加对前层信息的判别,提高图像转换的质量,通过实验表明,在不同的图像转换方法中,本文在定量和定性上都有明显优势。

In the process of image conversion, image blur and loss of details are the ever-present problems. Aiming for solving these problems, this paper designs a new image conversion model to reduce image blur and loss. In this paper, the asymmetric convolution block is introduced into the generator to enhance the feature extraction, and the detail information is added through the upsampling. A new multi-scale fusion method is proposed in the discriminator to increase the discrimination of the previous layer information and improve the quality of image conversion. Experiments show that compared with different image conversion methods, this paper has obvious advantages in both quantitative and qualitative aspects.