With　the　development　of　deep　learning,　this　technology　has　unique　advantages　in　feature　extraction　and　feature　nonlinear　mapping,　and　it　has　become　the　primary　solution　to　computer　vision　problems.　Research　on　similarity　image　retrieval　has　turned　machine　learning　into　deep　learning.　Generally,　similarity　image　retrieval　systems　based　on　deep　learning　first　train　a　model　for　image　classification,　and　then　extract　bottleneck　layer　features　as　image　representation　features　for　storage　and　retrieval.　The　problems　with　such　a　design　are:　the　bottleneck　layer　features　lose　image　detail　information,　and　the　features　of　similar　images　are　not　compact　enough.　This　article　aims　to　improve　the　method　for　the　above　problems.　Adding　metric　learning　to　constrain　the　image　features　to　solve　the　problem　of　inadequate　compact　image　features;　using　local　feature　fusion　features　instead　of　bottleneck　layer　features　as　image　representation　features　can　express　more　image　content.　Finally,　the　effectiveness　of　the　improvement　proposed　in　this　paper　is　verified　by　experiments.