Multi-Focus Image Fusion with Shuffle Coordinate Attention Guidance

To address the problem of degradation of multi-focus image fusion performance due to insufficient decision map accuracy, an algorithm is proposed to improve the decision map accuracy by multi-scale dense residual network combined with attention mechanism. Firstly, a multi-scale dilated convolution module is designed to extract information at various scales from the multi-focused images, enabling a wider range of perceptual field feature extraction. Then the feature extraction capability is enhanced using a dense residual network to preserve the middle layer feature information. Additionally, the attention mechanism of shuffle coordinate is employed at the end of the encoding sub-network, focusing on the global context information to enhance the capability of highlighting important feature information in multi-focused images. Finally, the decoding sub-network is constructed by full convolutional layers to reconstruct the features, and the decision map is generated directly by the symbolic function binary mapping to guide the image fusion. The fusion experiments show that the method significantly enhances various indices, including standard deviation, spatial frequency, and mutual information, compared to several classical comparison algorithms.