Few-shot　anomaly　detection　needs　to　solve　the　problem　of　anomaly　detection　when　the　training　samples　are　scarce.　The　previous　anomaly　detection　methods　showed　incompatibility　in　the　case　of　lack　of　samples,　and　the　current　few-shot　anomaly　detection　methods　were　not　satisfactory.　So,　we　hope　to　solve　this　problem　from　the　perspective　of　few-shot　learning.　We　utilize　a　pre-trained　model　for　feature　extraction　and　construct　multiple　sub-prototype　networks　in　multi-scale　features　to　compute　anomaly　maps　corresponding　to　each　scale.　The　final　anomaly　map　can　be　used　for　anomaly　detection.　Our　method　does　not　need　to　be　trained　for　each　category　and　can　be　plug-and-play　when　there　are　a　small　number　of　normal　class　samples　as　the　support　set.　Experiments　show　that　our　method　achieves　excellent　performance　on　MNIST,　CIFAR10,　and　MVTecAD　datasets.