Insulation Pull Rod Defect Detection With Data Augmentation Algorithm Focused on Target Region Integrity

Based on deep learning, object detection networks provide an accurate and sensitive method for ensuring the factory quality of insulation pull rods (IPRs). However, the stability of network training and the performance of the trained model heavily depend on large-scale, high-quality sample data. In practice, obtaining defect samples for IPR is challenging, and the available data is often scarce. To effectively address these issues, this article proposes a data augmentation algorithm focused on target region integrity to bridge the significant semantic information gap caused by the Mosaic augmentation applied to the raw IPR defect images. Comparative experiments are conducted on datasets constructed by this augmentation algorithm and Mosaic augmentation, using mainstream industrial defect detection networks in the field. The experimental results demonstrate that the proposed algorithm enhances the network's ability to effectively perceive IPR defects. Furthermore, to overcome the challenges of low detection efficiency and poor timeliness of intelligent analysis under real-world conditions, an intelligent defect detection system for IPR based on translucent imaging is developed. By integrating the proposed algorithm, the system is tested in practice, further validating the approach. The results show that the system offers significant advantages in rapidly and accurately detecting IPR defects, thus meeting the quality inspection needs in actual production processes. This article provides a new perspective for the development of highly sensitive detection methods for IPR defects.