无监督深度学习模型的多聚焦图像融合算法

深度学习技术应用到多聚焦图像融合领域时，其大多通过监督学习的方式来训练网络，但由于缺乏专用于多聚焦图像融合的监督训练的标记数据集，且制作专用的大规模标记训练集代价过高，所以现有方法多通过在聚焦图像中随机添加高斯模糊进行监督学习，这导致网络训练难度大，很难实现理想的融合效果。为解决以上问题，提出了一种易实现且融合效果好的多聚焦图像融合方法。通过在易获取的无标记数据集上以无监督学习方式训练引入了注意力机制的encoder-decoder网络模型，获得输入源图像的深层特征。再通过形态聚焦检测对获取的特征进行活动水平测量生成初始决策图。运用一致性验证方法对初始决策图优化，得到最终的决策图。融合图像质量在主观视觉和客观指标两方面上进行评定，经实验结果表明，融合图像清晰度高，保有细节丰富且失真度小。

Multi-focus Image Fusion Algorithm Based on Unsupervised Deep Learning Model

When deep learning technology is applied to the field of multi-focus image fusion, it mostly trains the network by supervised learning. However, due to the lack of labeled data sets of multi-focus image fusion for supervised training, and the cost of making a dedicated large-scale labeled training set is too high. The existing methods mostly simulate multi-focus image for supervised learning by adding Gaussian blur to the focused image, which makes network training difficult and difficult to achieve the ideal fusion effect. In society to solve the above problems, a multi-focus image fusion method with easy implementation and good fusion effect is proposed. Encoder-Decoder network model with attention mechanism is introduced by training in an unsupervised learning way on easily accessible unlabeled data set to obtain the rich features of the input source images. And then the initial decision map is generated by measuring the activity level of these features through multi-scale morphological focus detection. Eventually, the consistency verification method is used to adjust and optimize the initial decision map, and the final decision map is obtained. The quality of the fusion image is valued on both subjective vision and objective indicators. The experimental results demonstrate that the fusion image has high definition, rich details and low distortion.