The　classification　of　marine　moving　targets　is　an　important　research　topic.　There　are　many　methods　to　apply　deep　learning　to　target　classification,　such　as　convolution　neural　networks　(CNNs).　However,　CNNs　have　complex　network　structures　and　large　numbers　of　parameters,　and　cannot　pay　attention　to　the　spatial　information　of　the　target.　This　paper　proposes　a　new　method　for　the　classification　of　marine　moving　targets　based　on　an　adaptive　multi-feature　extraction　module　and　a　capsule　network　(DA_CapsNet).　The　attention　mechanism　captures　the　dependencies　between　feature　channels　and　enhances　the　weights　of　salient　features.　Different　from　the　Squeeze-and-Excitation　network　(SENet),　SENet_avgmax　can　weight　the　feature　maps　comprehensively.　Experimental　results　show　that　the　proposed　method　achieves　better　classification　performance　compared　to　the　state-of-the-art　methods.　　©　2022　IEEE.