基于注意力机制的改进双判别器图像修复算法

Improved dual discriminator image inpainting algorithm based on attention mechanism

为解决基于生成式对抗网络的图像修复模型存在的修复结果效果差和内容、细节等特征信息还原不准确这一问题,提出一种融合通道、像素注意力机制的多损失生成对抗网络算法.首先,该算法利用通道注意力块获取高关联的通道特征图;然后,通过像素注意力块对高关联通道特征图上所有像素进行打分,从而获取与缺损区域关联性更高的图像未缺损区域信息;最后,通过引入Vgg16特征提取模型向生成器的优化函数中引入内容、风格损失项,以多损失融合的方式提高图像的修复效果.在目前广泛使用的CelebA数据集和SVHN数据集上验证模型的修复效果,本算法在主客观指标上均优于DCGAN算法、CE算法和DD算法.

To solve the problems of poor inpainting results and inaccurate inpainting of the feature information such as content and details in the image inpainting model based on the generative adversarial network,a multi-loss generative adversarial network algorithm that integrates channels and pixel attention mechanisms is proposed.First,the algorithm uses the channel attention block to obtain highly correlated channel feature maps,then,the algorithm uses the pixel attention block to score all the pixels on the highly correlated channel feature maps,so as to obtain the information of the non-defective area of the image with higher correlation with the defect area,finally,it introduces the content and style loss items into the optimization function of the generator through the introduction of the Vgg16 feature extraction model,which improves the image inpainting effect by multi-loss fusion.Besides,it verifies the inpainting effect of the model on the widely used CelebA data set and SVHN data set.In terms of the algorithm in this thesis,it is superior to the algorithm of DCGAN,CE and DD in subjective and objective indicators.