Defect　detection　has　a　wide　range　of　applications　in　industry,　and　previous　work　has　tended　to　be　supervised　learning,　which　typically　requires　a　large　number　of　samples.　In　this　paper,　we　propose　an　unsupervised　learning　method　that　learns　knowledge　about　normal　images　by　distilling　knowledge　from　a　pre-trained　expert　network　on　ImageNet　to　a　learner　network　of　the　same　size.　For　a　given　input　image,　we　use　the　differences　in　the　features　of　the　different　layers　of　the　expert　network　and　learner　network　to　detect　and　localize　defects.　We　show　that　using　comprehensive　knowledge　makes　the　differences　between　the　two　networks　more　apparent　and　that　combining　the　differences　in　multi-level　features　can　make　the　networks　more　generalizable.　It＇s　worth　noting　that　we　don＇t　need　to　split　the　picture　into　patches　to　train,　and　we　don＇t　need　to　design　the　learner　network　additionally.　Our　general　framework　is　relatively　simple,　yet　has　a　good　detection　effect.　We　provide　very　competitive　results　on　the　MVTecAD　dataset　and　DAGM　dataset.