像素级卷积神经网络多聚焦图像融合算法

Multi-focus image fusion algorithm based on pixel-level convolutional neural network

提出了一种用于多聚焦图像融合的卷积神经网络(CNN)。与现有的基于CNN的图像融合方法将源图像分解成几个小块,然后使用一个分类器来估计图像块是否聚焦相比,本文方法直接将整个图像转换成一个决策图。像素级回归策略可以充分利用互补信息,解决了聚焦/散焦区域周围模糊程度估计的困难。此外,在图像融合领域,应用环形残差网络(RResNet)模块来提取更多聚焦区域的语义信息。同时,利用结构相似度(SSIM)估计生成的融合图像与参考图像之间的结构相似性以提高融合图像的质量,同时采用边缘保留损失函数来保留源图像中更多的梯度信息。实验结果表明:该方法在主观视觉效果和客观评价方面均优于其他融合算法。

In this paper,a novel convolutional neural network(CNN) for multi-focus image fusion is proposed. Compared with existing image fusion methods based on CNN which decompose the source image into several patches and adopt a classifier to estimate whether the patch is focused or defocused,the method in this paper directly converts the whole image into a decision diagram. The pixel-level regression strategy can make use of the complementary information and address the difficulty of estimating blur level around the focused/defocused region. Furthermore,the ringed residual network(RResNet)block is utilized to extract more semantic information from the focused region in image fusion field. In the meanwhile,the structural similarity index(SSIM)loss is utilized to estimate the structural similarity between the generated fusion image and the ground-truth reference to improve the quality of the fused images,and the edge preservation loss function is applied to preserve more gradient information from source image. Experimental results demonstrate that the proposed method is superior to other fusion algorithms in subjective visual effect and objective assessment.