Aiming at the high missed detection rate of casting defects, a casting defect detection method based on deep learning model fusion is proposed. Firstly, the Faster RCNN network is improved, the feature pyramid structure is used to improve the feature extraction network module, multi-scale feature fusion is realized, and the feature extraction of casting defects is completed. Then, the ROI pooling layer in the network is improved based on ROI Align, and the IOU score is introduced into the judgment process of NMS algorithm. And the improved network is integrated with Cascade RCNN and YOLOv3. Finally, an experiment study was carried out to verify that the fusion model can effectively reduce the missed detection rate of casting defects. The experiment results show that the defect recall rates in the Faster RCNN model and the network model proposed in this paper are increased by 1. 73% and 4. 08% , respectively after the pooling improvement of the region of interest. Using the method of model fusion, in the condition without considering the classification accuracy, the defect recognition rate of the entire model reaches 95. 71% . Compared with single model, while guaranteeing the detection accuracy of casting defects, the method also improves the defect detection recall rate and meets the requirements of industrial applications.