SLAM　technology　is　an　important　basis　for　autonomous　navigation　of　robots,　and　loop　closure　detection　is　an　important　part　of　SLAM　technology.　Its　task　is　to　identify　whether　the　robot　has　moved　to　its　current　position,　which　has　a　good　role　in　drawing　map.　In　this　paper,　a　feature　extraction　model　is　obtained　by　training　the　custom　data　set　on　the　ResNet50　residual　network,　which　can　extract　features　with　better　robustness.　Then,　the　twin-network　and　terny-loss　function　are　introduced　to　improve　the　network　performance　through　weak　supervised　training.　Finally,　the　cosine　similarity　method　is　used　to　calculate　the　similarity[6].　If　the　similarity　exceeds　the　threshold,　it　is　considered　that　there　is　a　loop.　After　the　comparison　experiment　between　the　proposed　method　and　the　traditional　method　on　the　open　data　set,　it　is　found　that　the　ResNet50　network　model　improves　the　feature　extraction　ability　and　the　loop　detection　accuracy,　which　proves　the　feasibility　and　application　value　of　this　paper.　©　2021　IEEE.