无监督编解码模型的多聚焦图像融合算法

Multi-focus image fusion algorithm based on unsupervised codec model

目前在多聚焦领域,大部分基于监督学习的深度模型都需要制作带标签的大规模数据集来训练网络,而制作数据集则需要花费很大的成本。为此,提出一种基于无监督学习的深度模型来实现准确和有效的多聚焦图像融合。通过无监督的方式在公共数据集上训练引入双重注意力机制的编码-解码模型,提取源图像的深层特征;利用改进的拉普拉斯能量和对深层特征进行聚焦检测得到决策图;根据决策图得到融合图像。实验结果表明,所提方法与14种先进算法相比,在主观视觉方面保有更多的图像细节,在7个客观指标中,有6个指标实现了最优结果。

Currently in the field of multi-focus image fusion,most deep models based on supervised learning require the production of large-scale datasets with labels to train the network,which is costly to be produced.For this reason,an unsupervised learning-based deep model was proposed to achieve accurate and effective multi-focused image fusion.An encoding-decoding model introducing a dual attention mechanism was trained on a public dataset by unsupervised learning strategy to extract deep features of the source images.The decision map was obtained using the improved Laplace energy and the focused detection of the deep features.The fused image was obtained based on the decision map.Experimental results show that the proposed method not only retains more image details in subjective vision,but achieves optimal results in 6 out of 7 objective metrics compared with 14 advanced algorithms.