Label recognition on metal surfaces in semiconductor industry by YOLO object detection model

Under stringent semiconductor manufacturing conditions, mistakenly installing an incorrect diamond dresser onto the machine can lead to unforeseen consequences. This study explores the application of optical character recognition (OCR) systems in recognizing laser-engraved labels on metal surfaces and evaluates the feasibility of using the You Only Look Once (YOLOv7) object detection model as an alternative method. In this study, the label data processing workflow was optimized by using an OCR system to identify text locations and predict characters. The equidistant segmentation method was employed to obtain individual bounding boxes for each character, converting them into a format recognizable by the YOLOv7 model. Simultaneously, image preprocessing techniques such as pixel adjustment and image segmentation were employed to reduce noise interference with the model and to minimize the impact of non-recognition objects on image complexity. Additionally, the adjustment of YOLOv7 data augmentation hyperparameters and the random generation of training images were used to address the effects of various environmental factors on recognition accuracy, thereby enhancing the efficiency of label recognition on metal surfaces. The experimental results indicate that the YOLOv7 model outperforms other OCR methods in terms of the average word accuracy rate for specification names and Lot-IDs.