Defect　detection　in　material　micro-images　has　a　significant　impact　on　the　study　of　the　relationship　between　the　micro-structure　and　macro-properties,　however,　material　microdefects　are　usually　relatively　small　and　span　a　wide　range　of　scales,　which　increases　the　difficulty　of　defect　detection.　Meanwhile,　since　defects　exist　in　a　small　number,　overfitting　becomes　another　challenge.　In　this　paper,　based　on　the　Faster　R-CNN　algorithm,　automated　data　enhancement　is　used　to　solve　the　overfitting,　and　a　feature　pyramid　model　is　proposed　for　the　defects　multi-scale　problem,　and　finally,　the　feasibility　of　the　above　viewpoint　is　verified　by　experiments.　©　2024　IEEE.