With　advancements　in　image　processing　and　the　proliferation　of　editing　software,　image　splicing　forgery　has　become　increasingly　facile　to　execute　yet　harder　to　detect,　thereby　impacting　societal　security.　Effective　detection　and　localization　methods　are　urgently　needed.　Existing　methods,　while　somewhat　effective,　often　over-rely　on　semantic　features,　overlook　shallow　features,　and　struggle　to　adapt　to　varying　tampered　region　sizes.　To　address　these　issues,　we　propose　a　two-stream　image　splicing　forgery　localization　network,　named　DMU-Net.　The　network　first　introduces　a　noise　stream　as　a　supplementary　feature　alongside　the　RGB　stream　to　provide　a　richer　feature　representation.　Subsequently,　we　improve　the　Atrous　Spatial　Pyramid　Pooling　module　by　incorporating　an　attention　mechanism　that　enables　the　model　to　obtain　feature　maps　of　different　scales,　effectively　use　context　information,　and　enhance　the　ability　to　capture　tampered　regions　of　various　sizes.　Finally,　we　employ　a　dual　attention　mechanism　to　fuse　features　from　both　the　encoder　and　decoder　stages;　this　approach　effectively　leverages　shallow　features　for　fusing　coarse-grained　and　fine-grained　features,　thus　enhancing　the　model’s　ability　to　capture　features　across　different　dimensions.　Extensive　experimental　results　show　that　the　proposed　method　outperforms　the　state-of-the-art　image　splicing　localization　methods　in　terms　of　detection　accuracy　and　robustness.　©　The　Author(s),　under　exclusive　licence　to　Springer-Verlag　GmbH　Germany,　part　of　Springer　Nature　2025.