The　accuracy　of　existing　underwater　sound　source　localization　algorithms　is　unsatisfactory,　and　most　of　them　cannot　achieve　cross-domain　localization.　To　solve　these　problems,　a　cross-domain　underwater　sound　source　localization　algorithm　based　on　a　binaural　matrix　and　mutual　information　constraint　loss　is　proposed.　In　this　algorithm,　a　new　binaural　matrix　feature　is　first　extracted　based　on　binaural　cues,　which　is　less　susceptible　to　environmental　interference　and　can　obtain　reliable　direction　information　from　received　signals.　Then,　a　constrained　loss　based　on　mutual　information　is　designed　to　constrain　the　proposed　neural　network　to　accurately　learn　the　shared　representations　of　different　domains.　This　ensures　that　the　high-dimensional　representations　used　for　localization　have　more　explicit　orientation　directionality.　Finally,　a　cross-domain　underwater　sound　source　localization　network　is　constructed　to　achieve　accurate　cross-domain　localization.　Experimental　results　indicate　that　the　algorithm　proposed　in　this　study　has　a　higher　localization　accuracy　than　comparative　algorithms,　both　in　the　same　domain　and　in　different　domains.　©　1976-2012　IEEE.