DBP-YOLO: A Dynamic Bifurcated-Path Fusion Network for Weld Seam Defect Detection

In order to tackle the issues related to identifying defects on weld seam surfaces, such as the concentrated distribution of defect targets, interference from complex background textures, and the lack of feature differentiation in conditions with low signal-to-noise ratios, this research proposes the Dynamic Bifurcated-Path Fusion Network (DBP-YOLO). This innovative network, grounded in the YOLOv8n architecture, incorporates the C2f-BiFusion (C2f-BiF). This unit greatly improves the ability to represent defect features by employing channel compression techniques, depthwise separable convolutions, and mechanisms for feature gating. Additionally, a Dynamic Perception Pyramid Network (DFFN) is introduced, which establish a cross-level two-way interaction mechanism to achieve multi-scale adaptive feature fusion. A lightweight detection module, low-rank and sparse cross-domain recommendation(LSCD), has been designed to successfully extract local characteristics of small defects while keeping the consumption of computational resources to a minimum. Moreover, we introduce the Focal PIoU loss function, which dynamically adjusts the weights for difficult and easy samples, integrating a joint optimization approach for accurate bounding box regression. This enhances both the convergence efficiency and the accuracy of the model. Experimental findings indicate that DBP-YOLO surpasses the benchmark model (YOLOv8n) with a 3.7% improvement in the mAP@0.5 metric and a reduction in model parameters by $1.21\times 10 ^{6}$ , achieving a detection speed of 62.4 FPS alongside an mAP@0.5 of 83.5% on the Jetson Xavier NX platform. These results satisfy the requirements for real-time detection and offer solid technical support for the advancement of intelligent weld seam detection technologies, the source codes are at https://github.com/xuchengniu/DBP-YOLO