Power Adapter Appearance Defect Detection Based on Task Feature Decoupling YOLOv8n

Abstract

In the realm of defect detection, there are distinctions in the emphasis placed on features between the classification and localization components of the task. The classification task emphasizes semantic information from the global context, while the localization task prioritizes spatial details such as edges. Directly coupling these two subtask features can hinder model convergence and degrade performance in the appearance defect detection of power adapters. To address this issue, we proposed the Task-feature Decoupled Feature Pyramid Network (TDFPN) module based on YOLOv8n. This module enhances semantic information and fuses corresponding features to improve detection performance in both localization and classification tasks. Additionally, we introduced the EMA module to suppress redundant information, enhance the model's attention towards defects, and improve the precision rates of detection. Furthermore, we replaced CIoU with an Inner-SIoU loss function that combines Inner-IoU based on auxiliary bounding boxes with SIoU, considering the matching direction. This replacement accelerated model convergence and improved the recall rates of detection. During training, transfer learning is employed by utilizing pre-trained weights from the YOLOv8n backbone, along with frozen training, to enhance efficiency. The experimental findings indicated that our proposed approach outperforms the original YOLOv8n model, demonstrating a 3.12% enhancement in mAP@0.5 and a 14.41% improvement in mRecall@0.5.