基于生成对抗网络的多聚焦图像融合

Multi-Focus Image Fusion Based on Generative Adversarial Network

多聚焦图像融合能够融合同一场景下具有不同聚焦部分的一系列图像.为了克服多聚焦图像融合模糊特征提取中存在的不足,提出一种基于U-Net的生成对抗网络模型.首先,生成器采用U-Net和SSE对多聚焦图像的特征进行提取,并完成图像融合;其次,判别器采用卷积层对已知的融合结果和生成器生成的融合图像进行分辨;然后,损失函数采用生成器的对抗损失、映射损失、梯度损失、均方误差损失和判别器对抗损失对生成网络进行参数调节;最后,将生成器、判别器和损失函数组成生成对抗网络模型,并进行实验.Pascal VOC2012数据集作为生成对抗网络的训练集,包括近焦图像、远焦图像、映射图像和融合图像.实验结果证明,该生成对抗网络模型能够有效地提取多聚焦图像中的模糊特征,且融合图像在互信息、相位一致性和感知相似性等方面表现优异.

Multi-focus image fusion can fuse a series of images that have a different focus in the same scene.To overcome the disadvantage extraction of the blurring character in multi-focus images,a generative adversarial network model based on U-Net is proposed.Firstly,the generator uses U-Net and SSE to extract the feature of the multi-focus image,and fuses images.Then,the discriminator uses convolutional layers to distinguish the fused result between the existed and the generative.Furthermore,a loss function has the loss of adversarial in the generator,loss of mapping,loss of gradient,loss mean square error and the loss of adversarial in the dis-criminator.The train data of generative adversarial network uses the dataset of Pascal VOC2012 to generate and includes near-focus image,far-focus image,mapping image and all-in-focus image.The experimental result shows that the proposed generative adversarial network model can effectively extract the blurring feature in multi-focus images,and the fused image have good performances on mutual information,phase congruency and structural similarity.