A test model of YOLOv3-Mv2 was proposed to detect the scratches and defects of castings. Firstly

a smaller skeleton network MobileNetv2 instead of the original feature extraction network structure Darknet53 was chosen to reduce the amounts of computational parameters of the network and to accelerate detection. Secondly

a new detection scale based on the fusion of deep and shallow features was added to enhance the detection capability of the small defect target.Then the parameters of the BN layer were recalculated

sharing connected area data with the reel layer to speed up the forward inference of the model. Finally

the Complete IoU function was introduced to improve positioning accuracy. The results indicate that the average accuracy（mAP） value and the real-time performance of YOLOv3-Mv2 were improved by 5.42% and 23 f/s than those of the original YOLOv3 algorithm

respectively.