In view of the complexity of the current target detection algorithm net work and the high demand on the platform equipment;And the low detection accura cy of the lightweight network YOLOv3_Tiny on the varistor surface defects,which is prone to miss and wrong detection,an improved algorithm DAYOLOv3_Tiny based on YOLOv3_Tiny is proposed.DAYOLOv3_Tiny constructs the deep separable convolu tion block to replace the standard convolution,use the convolution operation to downsample,so that the detection network can reduce the number of network para meters and increase the feature extraction;The channel attention module and spa tial attention module are introduced into the network to enhance the learning of important feature information.The results on a self-made varistor surface defe c t data set show that the MAP value of DAYOLOv3_Tiny is 92.23%,which is 12.25% h igher than before.The size of the improved DAYOLOv3_Tiny model is 55.42% of tha t of YOLOv3_Tiny,which is only 18.9MB.Experiments show that DAYOLOv3_Tiny has a higher detection accuracy for varistor surface defects,and can effectively im prove the situation of missed and wrong detection.Besides,the network model is small,and it does not require high hardware platform,so it is easy to deploy i n the platform with limited performance.