PCP-YOLO: an approach integrating non-deep feature enhancement module and polarized self-attention for small object detection of multiscale defects

Abstract

The detection of small objects within multiscale defects amidst complex background interference presents a formidable challenge in industrial defect detection. To address this issue and achieve precise and expeditious identification in industrial defect detection, this study proposes PCP-YOLO, a novel network that incorporates a non-deep feature extraction module and a polarized filtering feature fusion module for small object defect detection. Initially, YOLOv8 is employed as the foundational model. Subsequently, a lightweight, non-deep feature extraction module, PotentNet, is designed and integrated into the backbone network. In the neck network, a feature fusion module incorporating polarized self-attention, C2f_ParallelPolarized, has been developed. Finally, CARAFE is utilized to substitute the original upsampling module in the neck network. The efficacy of this approach has been rigorously evaluated using three datasets: the publicly available NEU-DET and PKU-PCB datasets, and the real-world industrial dataset GC10-DET. The mAP@0.5 values achieved are 79.4%, 96.1%, and 77.6%, significantly outperforming other detection methods. The method also has a fast inference speed. These results demonstrate that PCP-YOLO exhibits substantial potential for rapid and accurate defect detection.