In　direction-of-arrival　(DOA),　most　of　the　existing　popular　deep　learning-based　methods　consider　uniform　linear　array　(ULA)　systems.　Although　they　show　higher　performance　than　parametric　methods,　the　high　accuracy　of　the　estimation　still　needs　to　be　improved.　In　recent　years,　coprime　arrays　have　been　widely　studied　due　to　their　advantages　over　ULAs.　In　this　paper,　we　propose　a　deep　framework　for　dual-signal　DOA　with　coprime　array,　considering　it　as　a　multi-classification　problem.　The　entire　framework　consists　of　the　feature　network　and　the　classifier　network.　First,　we　propose　a　preprocessing　method　suitable　for　coprime　arrays,　where　the　extended　covariance　matrix　is　used　as　the　input　processing　object.　Next,　the　feature　network　extracts　feature　from　the　data.　We　introduce　a　lightweight　channel　attention　mechanism　to　help　improve　the　performance　of　the　deep　model.　Finally,　the　classifier　network　performs　DOA　estimation.　©　2024　SPIE.