残差网络在工程竹材图像超分辨率中的应用

Application of Residual Network in Super-resolution of Engineering Bamboo Images

针对现有的工程竹材散斑图像重建方法性能上的不足,在基于注意力密集残差网络的基础上,构建了用于工程竹材散斑图像超分辨率重建的生成对抗网络。使用带有注意力模型的残差网络作为主网络,去除所有的BN层,将单层网络逐层密接,使用L_(1)损失函数训练网络,将以上网络作为生成模型,使用相对均值判别器作为网络的判别模型,生成器使用结合感知损失、内容损失和对抗损失的综合损失函数,判别器使用相对均值判别器的损失函数,使网络能够还原出质量更高的图像。

In view of the performance shortcomings of the existing engineering bamboo speckle image reconstruction methods,based on the attention intensive residual network,a generation countermeasure network for super-resolution reconstruction of engineering bamboo speckle image was constructed.Using the residual network with attention model as the main network,all BN layer removed,the single-layer network connected layer by layer,and L_(1).loss function used to train the network,the above network taken as the generation model,the relative mean discriminator used as the discrimination model of the network,the comprehensive loss function combining perception loss,content loss and confrontation loss used by the generator,and the loss function of the relative mean discriminator used by the discriminator,the network could restore higher-quality images.