Sumyung Gang

Continual learning (CL) is becoming increasingly important, not only for storage space because of the ever-increasing amount of data being generated, but also for associated copyright problems. In this study, we propose ground truth’ (GT’), which is a combination of ground truth (GT) and a prediction of the teacher model that distills the prediction results of the previously trained model, called the teacher model, by applying the knowledge distillation (KD) technique to an anchor-free object detection model. Among all the objects predicted by the teacher model, an object for which the prediction score is higher than the threshold value is distilled into the current trained model, called the student model. To avoid interference with new class learning, the IoU is obtained between every object of the GT and the predicted objects. Through the continual learning scenario, even if the reuse of past data is limited, if new data are sufficient, the proposed model minimizes catastrophic forgetting problems and enables learning for newly added classes. The proposed model was learned in PascalVOC 2007 + 2012 and tested in PascalVOC2007, with better results of 9.6% p mAP and 13.7% p F1ⁱ shown in the scenario 19 + 1. The result in scenario 15 + 5 showed better results than the compared algorithm, with 1.6% p mAP and 0.9% p F1ⁱ. The scenario 10 + 10 also outperformed the other alternatives, with 0.9% p mAP and 0.6% p F1ⁱ.

As the size of components mounted on printed circuit boards (PCBs) decreases, defect detection becomes more important. The first step in an inspection involves recognizing and inspecting characters printed on parts attached to the PCB. In addition, since industrial fields that produce PCBs can change very rapidly, the style of the collected data may vary between collection sites and collection periods. Therefore, flexible learning data that can respond to all fields and time periods are needed. In this paper, large amounts of character data on PCB components were obtained and analyzed in depth. In addition, we proposed a method of recognizing characters by constructing a dataset that was robust with various fonts and environmental changes using a large amount of data. Moreover, a coreset capable of evaluating an effective deep learning model and a base set using n-pick sampling capable of responding to a continuously increasing dataset were proposed. Existing original data and the EfficientNet B0 model showed an accuracy of 97.741%. However, the accuracy of our proposed model was increased to 98.274% for the coreset of 8000 images per class. In particular, the accuracy was 98.921% for the base set with only 1900 images per class.