To address the problems of low accuracy and slow detection speed in existing target detection algorithms used by rebar-binding robots for identifying binding points, an improved rebar mesh binding-point detection method named FNB-YOLOv5, was proposed based on enhancements to YOLOv5. First, a dataset of rebar grid intersections was created through image acquisition. Then, the lightweight FasterNet backbone was incorporated into the YOLOv5 network to enhance feature extraction while reducing network complexity. Next, the BiFormer attention mechanism was introduced into key network components to improve the accuracy of feature extraction. Considering that small-sized targets dominate the detection task, the NWD loss function was used for normalization to optimize the localization accuracy of binding-point detection. Finally, an improved F-global feature pyramid network (F-GFPN) feature fusion module was devised to enhance feature interaction and improve computational performance by incorporating skip connections and cross-scale connections. Comparative experiments with different models show that the proposed method achieves a precision of 99.82%, a recall of 99.10%, and a mean average precision of 98.64%, which represent an increase of 2.9 percentage points, 1.53 percentage points, and 1.63 percentage points over the original model, respectively. The frame per second (FPS) reaches 44.8, an increase of 4.1 compared with the original model, while the size of weight model is reduced by 2.93 MB. Experimental results demonstrate that the improved FNB-YOLOv5 model achieves higher accuracy and real-time performance on the rebar binding point dataset, providing technical support for the development of rebar binding robots in the construction industry.
