细粒度全局感知多聚焦图像融合网络

Fine-grained Global Perception Multi-focus Image Fusion Network

多聚焦图像融合是图像融合的一个重要分支,在显微成像等方面具有广泛的应用.针对多聚焦融合中存在的纹理细节不清晰、聚焦区域误判等问题,本文从空间及通道信息全局关注的角度出发,结合Swin Transformer中的移动窗口自注意力机制和深度可分离卷积设计了一个全局信息编码-解码网络,采用综合损失函数进行图像重构任务的无监督学习;从特征邻域信息重要性的角度出发,引入了改进的拉普拉斯能量和函数在特征域进行图像聚焦属性的判别,增强图像聚焦区域判别的细粒度效果.与7种经典图像融合算法比较,本文算法在定性和定量分析中均取得了先进的融合性能表现,对原始图像的聚焦区域信息具有更高的保真效果.

Multi-focus image fusion is an prominent branch of image fusion,which is widely used in microscopic imaging.Aiming at the problems of unclear texture details and misjudgment of focus areas in multi-focus fusion,this paper designs a global information encoding-decoding network from the perspective of global attention of spatial and channel information,combined with the shifted window self-attention mechanism in Swin Transformer and deep separable convolution.The comprehensive loss function is used to perform unsupervised learning of image reconstruction tasks.From the perspective of the importance of feature neighborhood information,an improved Laplacian energy sum function is introduced to discriminate the image focusing-properties in the feature domain,and the fine-grained effect of image focusing region discrimination is enhanced.Compared with seven classical image fusion algorithms,the proposed algorithm achieves advanced fusion performance in both qualitative and quantitative analysis and has a higher fidelity effect on the focus area information of the original image.