Research on intelligent recognition method of mechanical parts with high feature similarity in industrial field environment

Precise and swift recognition of mechanical parts in complex industrial settings is crucial for the intelligent assembly and handling of materials. The study examines the intelligent identification of various parts that exhibit significant resemblance in features across different lighting and occlusion scenarios. This study proposes an enhanced You Only Look Once version 8 nano (YOLOv8n)-based lightweight network to boost detection precision, accuracy, and speed. Integrating a bidirectional feature pyramid network (BiFPN) feature fusion module into the YOLOv8 head improves the ability to extract features on multiple scales. The incorporation of FasterNet’s streamlined architecture into the core network simplifies the model for industrial use. Furthermore, the integration of an attention mechanism and the Wise-IoU loss function enhances the nuanced extraction of features, thereby improving the precision and speed of detection. The experimental results show that the improved model outperforms the standard in identifying highly similar mechanical parts in complex industrial scenarios, resulting in greater precision and efficiency. In contrast to the initial YOLOv8n, this model shows notable enhancements: a 1.5% increase in mAP@0.5, a 0.3% rise in detection precision, and a 41.7% reduction in parameters. This method offers a proficient way to identify industrial parts.