Study on Spark Image Detection for Abrasive Belt Grinding via Transfer Learning with YOLOv8

Aiming to solve the problems of low precision and poor efficiency caused by relying on manual experience during the manual polishing of blades, a multi-view spark image detection method based on YOLOv8 transfer learning is proposed. A multi-pose spark image dataset including front, side, and 45° angle views is constructed, and the cross-view detection task is achieved for the first time. The generalization ability of the model is enhanced through the following innovative strategies: (1) a cross-view transfer learning framework based on dynamic anchor box optimization is designed, and the parameters of the front spark detection model YOLOv8 are transferred to the side and 45°-angle detection tasks; (2) an attention-guided feature alignment module is introduced to alleviate the feature distribution shift caused by view differences; and (3) a curriculum learning strategy is adopted, where the datasets of different views are trained separately first and then sampled to reconstruct the dataset for further training, gradually increasing the weight of samples from complex views. The experimental results show that on the self-built multi-view dataset (containing 3000 annotated images), this method achieves an average detection accuracy of 98.7%, which is 14.2% higher than that of the original YOLOv8 model. The inference speed reaches 55 FPS on an NVIDIA RTX 4090, meeting the requirements of industrial online monitoring. The research results provide key technical support for the intelligent prediction of the material removal rate in the precision machining of blades and have the potential for rapid deployment in industrial scenarios.