As　an　important　application　field　for　object　detection　and　instance　segmentation,　industrial　inspection　has　attracted　more　and　more　attention　from　the　manufacturing　industry　in　recent　years.　Improving　the　performance　of　existing　algorithms　in　the　field　of　industrial　inspection　has　many　advantages　such　as　reducing　costs　and　increasing　safety.　In　this　paper,　we　select　the　four　most　representative　types　of　watch　parts　and　use　them　to　build　a　micro-part　dataset　for　training.　Given　a　test　image　with　small　parts,　the　learned　model　can　detect　and　segment　all　parts　effectively.　In　addition,　by　optimization　of　excess　bounding　boxes,　we　modify　the　NMS　and　propose　a　novel　loss　function　to　solve　the　problem　on　the　number　of　bounding　boxes　that　arise　during　the　procedure　of　detection　and　segmentation.　Experimental　results　on　the　micro-part　dataset　demonstrate　that　our　method　can　reduce　the　number　of　error　bounding　box　and　improve　the　performance　on　detection　and　segmentation　in　contrast　to　Mask　R-CNN.　©　2019,　Springer　Nature　Switzerland　AG.