Compared　with　traditional　natural　images,　remote　sensing　images　(RSIs)　typically　have　high　resolution.　The　objects　in　the　images　are　densely　distributed,　with　heterogeneous　orientation　and　large　scale　variation,　even　among　objects　of　the　same　class.　In　recent　years,　object　detection　algorithms　have　made　great　strides　in　general　images,　but　they　are　still　difficult　to　meet　the　challenges　that　exist　in　RSIs.　Therefore,　we　propose　a　foreground　feature　embedding　network　(FFE-Net)　for　object　detection　in　RSIs.　To　better　grasp　the　object　features　in　RSIs,　we　design　a　foreground　feature　embedding　module　(FFEM)　to　learn　the　foreground　features　of　the　object.　This　is　achieved　by　introducing　an　additional　semantic　segmentation　branch　and　embedding　the　features　in　the　classification　and　regression　branches.　Simultaneously,　we　propose　a　modified　Gaussian　function　with　focal　loss　(MGFFL)　as　a　way　to　eliminate　the　extra　background　noise　from　soft　labels,　making　the　learned　foreground　features　more　robust.　Our　experimental　results　on　two　publicly　available　remote　sensing　image　datasets,　DOTA-v1.0　and　HRSC2016,　validate　the　effectiveness　of　FFE-Net.　©　2024,　The　Author(s),　under　exclusive　license　to　Springer　Nature　Singapore　Pte　Ltd.