Abstract: Objective: To improve the accuracy of surface defect detection of quick-frozen dumplings. Methods: A dataset covering five quick-frozen dumpling forms (normal, leak, half, broken and adhesion) was elaborated, and the GX-RetinaNet network model was proposed for surface defect detection and localization of quick-frozen dumplings. The model was improved based on the RetinaNet network. The backbone feature extraction network adopted the ResNeXt-50 model, which had strong feature extraction ability. The addition of the Convolutional Block Attention Module (CBAM) and the use of the Swish activation function could effectively suppress the influence of background noise. By adding the Path Aggregation Network (PAN) structure behind the Feature Pyramid Networks (FPN) structure to form a bidirectional feature fusion module, the fusion ability of target multi-scale feature information could be improved. Results: The online detection accuracy of the GX-RetinaNet network for surface defects of quick-frozen dumplings under industrial field conditions was better than that of several mainstream target detection networks. The mAP was 94.8 %, the Recall was 77.0 % and the F1-score was 84.9 %. Compared with the RetinaNet network, mAP, Recall, and F1 increased by nearly 2.6 %, 2.6 % and 2.4 %. Conclusions: The GX-RetinaNet network model could meet the requirements of surface defect detection accuracy of quick-frozen dumplings. This study provided a feasible method for the application of deep learning theory in the surface defect detection of frozen dumplings.