AFP-IA: Adaptive Foreground Perception-Based Image Augmentation for Insulation Pull Rod Defects Detection

The complex background noise in raw insulation pull rod (IPR) images, coupled with the small and inconspicuous nature of many defect regions, often hinders the effective perception of defect features by deep learning networks. To address this issue, this study proposes an adaptive foreground perception-based data augmentation (AFP-IA) strategy for IPR defect detection. This strategy employs appropriate digital image processing techniques to separate the foreground and background of the raw IPR images under inspection. Furthermore, the efficacy of this image enhancement method is corroborated by its integration with contemporary mainstream detection networks in the domain of industrial defect detection networks. The experimental results demonstrate that the AFP-IA is an effectively method for enhancing the detection performance networks for IPR defects. Following training on the augmented dataset, the networks demonstrated an average mean average precision (mAP) at 0.5 and mAP at 0.5:0.95 of 0.923 and 0.626, respectively, indicating a significant improvement in detection accuracy. Furthermore, an intelligent IPR defect detection system based on translucent imaging was developed, and instance testing was performed using models trained on the aforementioned experimental data. This further validated the effectiveness of the proposed augmentation strategy in IPR defect detection. The study also investigated the impact of feature information quality within the receptive fields of different network layers on detection performance, providing deeper insights into the underlying mechanisms. Moreover, the successful integration of deep learning networks with digital image processing techniques was demonstrated through a practical application.