<p>Object detection tasks extensively employ neural networks to increase the efficiency of manufacturing processes in various industries. Developing extensive image datasets necessary for training these networks usually requires experts to perform manual labeling, which is labor-intensive and time-consuming. To address these challenges, this study introduces a semi-auto-labeling framework for high-frequency induction brazing (HFIB) using a semi-supervised rotational object detection (SSROD) network, score filtering and an expert review system (ERS). The SSROD utilizes both sparsely labeled and abundant unlabeled images, facilitating mutual training between student and teacher models to enhance the robustness and accuracy of the neural networks. Additionally, rotational bounding boxes are employed to reduce background noise and improve feature extraction. A score-filtering method automatically generates labels for unlabeled images based on high-confidence predictions from the trained SSROD, which enhances label quality in subsequent training phases. ERS further refines these auto-labeled images by correcting misclassifications, enhancing both labeling accuracy and object detection performance in neural networks. Systematic analysis demonstrates that the proposed framework significantly enhances the efficiency and accuracy of both labeling and object detection in HFIB images, supporting accurate real-time estimations. Additionally, this approach aids in developing large datasets, crucial for AI-driven automation in real-world industrial applications.</p>