MBDANet: a multi-scale balanced dynamic alignment detection network based on knowledge distillation for small objects

Small object defect detection is a pivotal technique for the industrial transition from post-facto rework to proactive prevention. However, existing detection methods still face challenges, such as the trade-off dilemma among computational complexity, model size, and detection accuracy. To address these issues, this study proposes a multi-scale balanced dynamic alignment detection network (MBDANet) based on the improved YOLOv8n architecture. The process begins with a robust feature downsampling module that downsamples the backbone network to better extract small-object features from both deep and shallow layers. Following this, a multi-scale feature fusion and balanced pyramid network fuses these features. To further enhance the model's attention to small objects, we design a task dynamic alignment detection head. Finally, we incorporate a bridging cross-task protocol inconsistency for distillation method, leveraging a higher-precision teacher model to boost the detection accuracy. Experiments on the Printed Circuit Board (PCB) Defect Dataset, DeepPCB Defect Dataset and Steel Surface Defect Dataset show that MBDANet outperforms the baseline YOLOv8n algorithm, with a 2.7%, 3.7% and 4.5% increase in mean Average Precision, a 25% reduction in model size, a 27% reduction in parameter number, and a 15% reduction in computational complexity. In summary, MBDANet can effectively improve the detection performance of small object defects, balance speed and accuracy, and provide solid technical support for the stable operation of industrial equipment in manufacturing scenarios.