基于级联注意力与密集特征融合的图像修复算法

Image Inpainting Algotithm Based on Cascading Attention and Intensive Feature Fusion

针对图像修复算法存在的语义不连贯、纹理不清晰等问题,提出一种基于生成对抗网络的新型级联密集生成网络CDGAN(Cascade Densely Generative Adversarial Network),采用encoder-decoder作为生成器主干,利用下采样提取图像特征;为使网络关注修复图像的高频纹理和颜色保真度等有效信息,引入级联的注意力模块,并加入密集特征融合模块扩大网络的整体感受野,充分学习图像特征,提高编码器提取特征的利用率,最后将处理后的图像特征进行上采样重建。在Celeb A和Places2数据集的测试结果表明,CDGAN在语义连贯性、纹理清晰度等方面都有所提升。

To address the problems of semantic incoherence and unclear texture in image inpainting algorithms,a new cascade dense generation network(CDGAN)based on generative adversarial networks was proposed,using encoder-decoder as the generator backbone,extracting image features using downsampling.The added attention module extract high frequency texture and color fidelity of the inpainted image.In addition,a dense feature fusion module was added to expand the overall receptive field of the network,fully learn the image features,improve the utilization of the encoder extracted features,and finally upsample the processed image features for reconstruction.The experimental results on the Celeb A and Places2 datasets show that CDGAN is improved in terms of semantic coherence and texture clarity.