基于像素校正的编解码多聚焦图像融合网络

Multi-Focus Image Fusion Network Based onP ixel Correction Module

聚焦区域边缘的精确提取是多聚焦图像融合的研究难点,现有的融合方法普遍存在融合图像边界模糊、关键信息丢失等问题。为解决上述问题,提出了一种基于编解码网络的多聚焦图像融合方法。首先,利用公开数据集构建一个带有精细标注的模拟多聚焦图像训练数据集,然后,在编解码网络中增加像素矫正模块与结构相似性损失函数,并用模拟数据集训练编解码网络;最后根据源图像对与模型预测的聚焦区域精确得分图构建最终的融合图像。实验结果表明,上述方法能够精确提取聚焦区域边缘信息,获得较高的互信息熵、平均梯度和结构相似度,且具有较好的视觉效果。

Accurate extraction of the focus region in multi-focus image fusion is a difficult problem. The problem of fusion result blurred in edge and key information lost have existed in many research methods. Therefore, a multifocus image fusion method based on the encoder-decoder network is proposed. First, the existing public dataset was used to create a multi-focus simulate image dataset. Then, the pixel correction module and structural similarity loss function were added to the encoder-decoder network, and the encoder-decoder network was trained with the simulated dataset. With the score map predicted from the model and a pair of source images, the fusion result could be determined. The results show that this method can accurately extract the edge information of the focus area, and the obtained mutual information entropy, average gradient, and structural similarity are high and have a better visual result.