Deep　convolutional　neural　networks　have　been　widely　used　for　saliency　detection.　However,　most　of　the　previous　works　focus　on　the　visible　light　image.　In　this　paper,　there　are　mainly　two　contributions.　First,　we　propose　a　new　architecture　named　Multi-level　Up-sampling　Network　(MLUNet)　for　infrared　(IR)　ship　object　saliency　detection.　Specifically,　the　architecture　of　MLUNet　is　an　Encoder-Decoder　like　network　embedded　with　subtraction　feature　filtering　module　(SFFM).　The　encoder　uses　the　DenseNet　like　architecture,　and　the　decoder　part　use　two　upsampling　methods,　which　are　deconvolution　and　sub-pixel　convolution.　SFFM　is　a　feature　subtraction　module　which　is　in　charge　of　feature　filtering.　In　our　proposed　MLUNet,　SFFM　is　embedded　after　each　convolution　and　deconvolution　block.　Secondly,　we　construct　an　IR　ship　object　image　dataset　for　saliency　detection.　This　dataset　includes　3845　IR　images　and　ground-truth　images　with　different　backgrounds　and　different　objects.　Experimental　results　show　that　our　method　outperforms　the　state-of-the-art　methods　in　terms　of　regional　evaluation　measures.