基于注意力和梯度残差密集块的图像融合

Image Fusion Using Coordinate Attention and Gradient Residual Dense Module

本文提出了一种基于坐标注意力机制和梯度残差密集块的融合算法,利用普通卷积提取源图像的浅层特征,使用梯度残差密集块模块提取源图像深层特征和细粒度细节特征;使用坐标注意力模块捕获特征图中空间位置之间的远程依赖关系重,将融合后的特征利用解码器重建融合图像。实验结果说明本文提出的融合算法具有较好的融合性能。

In this paper a fusion algorithm based on gradient residual dense block and attention mechanism is proposed.The shallow features of the source images are extracted by convolution,and the deep features and fine-grained details of the source images are extracted by gradient residual dense block module.The coordinate attention module is used to capture the remote dependence between the spatial positions in the feature maps,and the decoder reconstructs fused features to obtain the fused image.The results show the proposed fusion algorithm has better fusion performance.