MDFA-DNANet: multi-scale dilated fusion attention network for infrared small target detection

The MDFA-DNANet proposed builds upon the accomplishments of Dense Nested Attention (DNA) networks, aimed at solving the problem of losing deep infrared small targets as the network depth increases. In order to enhance the feature representation of deep infrared small targets, We propose a novel Multi-scale Dilated Fusion Attention (MDFA) network, which has a multi-scale dilated convolution that is capable of extract features at different spatial scales using different dilation rates. The framework dynamically integrates, through feature fusion, both local geometric details and global semantic information. We propose a channel and spatial attention mechanism to enhance feature representation. Strengthen useful feature channels and suppress unimportant channels to Optimize quality of feature representation. The Multi-scale Dilated Fusion Attention Network demonstrated excellent performance and validated its effectiveness through extensive evaluation on the NUDT-SIRST dataset.