基于多尺度空洞卷积网络的多聚焦图像融合算法

Multi-Scale Dilated Convolutional Neural Network Based Multi-Focus Image Fusion Algorithm

针对现有基于深度学习的图像融合算法中存在图像特征提取尺度单一、卷积核感受野小、不能有效突出显著特征等问题,提出了一种基于含注意力机制的多尺度空洞卷积网络的多聚焦图像融合算法。首先,构造一种多尺度空洞卷积模块,通过不同的扩张率改变卷积的感受野,从而提取源图像中的多尺度特征。此外,在多尺度空洞卷积模块中引入注意力机制,能自适应地选择显著性特征,进一步提高融合性能。所提融合网络包含特征提取、特征融合和图像重建等3个部分,其中特征提取部分主要由多尺度空洞卷积模块构成。相关实验结果表明,所提算法与现有基于深度学习的算法相比具有一定的竞争力。消融实验也验证了所提多尺度空洞卷积模块能强化网络的特征提取能力,提高图像融合质量。

According to the issues of single-scale image feature extraction,small receptive field,and cannot highlighting salient features in existing deep learning based image fusion algorithms,this paper proposes a multi-scale dilated convolution network with attention mechanism for multi-focus image fusion.First,a multi-scale dilated convolution block(MDB)is proposed.The MDB with different dilation rates can provide different receptive fields,and consequently it can extract the multi-scale features.Moreover,the attention mechanism is introduced into the MDB,which can adaptively select the salient features and improve the performance further.The proposed fusion network consists of three parts,including feature extraction,feature fusion,and image reconstruction.Specifically,the feature extraction part is composed of several MDBs.The experimental results demonstrate that the proposed method is competitive to some existing deep learning based methods.Some ablation studies also verify that the MDB can enhance the ability of feature extraction and improve the image fusion quality.