QCF-YOLO: A Lightweight Model of Surface Defect Detection for Quick-Connect Fittings

In response to the current enterprise needs for edge deployment of defect detection models within the “Cloud–Edge–Device Collaborative Architecture” and the challenge of low detection accuracy for minor defects in quick-connect fittings (QCFs), this article proposes an intelligent lightweight detection model for QCF defects based on YOLOv8 (QCF-YOLO). First, we replace the standard convolution and CSP bottleneck with two convolutions-fast (C2f) modules in the backbone network with GhostConv and C3Ghost modules, and reducing the number of channels in the Neck network. This modification effectively decreases the model’s parameter size, facilitating deployment on embedded devices. Second, to counteract accuracy loss from reducing parameter size, we incorporate DySample in the feature fusion network. This operator effectively restores detailed information in low-resolution maps, boosting feature expression capability. Additionally, to enhance the detection accuracy of minor defects, we introduce a P2 small target detection head in the detection component, which can capture more details, to improve detection accuracy for minor defects. Finally, we have established a defect detection experimental platform for QCFs to evaluate the feasibility of model deployment. Compared to other mainstream detection models, the proposed QCF-YOLO model demonstrates advantages in high precision, low parameter size, and rapid detection speed, achieving a mean average precision (mAP) of up to 95.5%, a model parameter size of only 1.187 M, and a frame rate of up to 69.8 frames/s. This model effectively meets the real-time detection requirements of enterprises and is deployed on the NVIDIA Jetson TX2 development board to prepare for the construction of a cloud-side collaborative architecture.