Abstract: Various types of cracks are prone to be produced in the production process of automotive cylinder head forgings, which affect the product quality. An improved algorithm model YOLOv5-MG based on the YOLOv5 network is proposed to address the current problem of low accuracy and efficiency in detecting defects in cylinder head forgings. Firstly, a sample dataset of cylinder head forgings defects is produced by building an image acquisition platform. Then, the accuracy and efficiency of the model are improved by replacing the YOLOv5 Head architecture of the SPP-YOLO network with the Decoupled Head structure; the fusion of shallow feature information with deeper feature information is enhanced by replacing the original path aggregation network (PANet) of the YOLOv5 network with the bidirectional feature pyramid network (BFPN); and the localization accuracy of the algorithm is improved by introducing the computation method of complete intersection over union (CIoU). 99.81% mean accuracy, 0.99 F1 score and 17.2 B floating point operations (FLOPs) are achieved by the improved algorithm on the test set.