Multi-focus image fusion algorithm based on adaptive connection and hybrid convolution attention

Abstract

In the field of multi-focus image fusion, in order to integrate the previous and subsequent information, most existing models simply overlay shallow and deep features directly at the encoding stage, but do not fully consider the importance of each feature layer, thus limiting the performance of image fusion. A multi-focus image fusion algorithm based on adaptive connection and hybrid convolution attention is proposed. The model uses a codec structure. Firstly, in the coding stage, the hybrid convolution attention module is used to enhance the feature extraction capability of the model. The design of adaptive connection helps the model to better capture the context information of the source image, and can adaptively calculate the weight of the feature map, and superimpose the feature map according to the weight. Secondly, deformable convolution module is used in decoding phase to enhance the modeling ability of complex edge details in the focused region and improve the feature reconstruction ability of the network. The experimental results on Lytro, MFI-WHU and HBU-CVMDSP datasets show that the proposed fusion algorithm achieves better fusion effect in both subjective and objective evaluation compared with other 10 fusion algorithms.