YOLO-DFW: An improved detection method for printed circuit board surface defects based on YOLOv8

High-precision defect detection on printed circuit board (PCB) plays a crucial role in ensuring the productivity and safety of electronic products. However, traditional defect detection methods are difficult to meet the demand for detecting tiny targets and difficult samples in complex environments. To address this challenge, we proposed an improved detection algorithm, YOLO-DFW, based on YOLOv8 network. In our study, the DynamicConv (DC) improves the C2f module, which enhances the model’s ability to express different features through adaptive weighted convolution. We propose the Feature Focused Pyramid Network (FFPN) by reconstructing the original neck structure to enhance the model’s multi-scale feature fusion capability through cross-scale feature fusion, and the Context Enhancement Module (CEM) is introduced into FFPN to expand the model’s receptive field. What’s more, a new loss function named WMN-loss is proposed to make the model pay more attention to the difficult-to-classify and small bounding boxes by using a non-uniform loss allocation strategy. Through training and testing on the PCB surface defect dataset, our method achieves improvements of 2.2%, 2.4%, 2.6%, and 4.2% in precision, recall, mAP₅₀ (mean average precision), and mAP_50:95, respectively, compared with the baseline model. Extensive experiments demonstrate the superiority of our method for PCB surface defect detection. Furthermore, comparison experiments on DeepPCB and NUE-DET datasets verify the feasibility and generalization of the YOLO-DFW detection algorithm.