疑似気泡画像自動生成によるFaster R-CNNベースの気泡検出アルカリ水電解への応用

抄録

Faster R-CNN method to automatically detect hydrogen bubbles within alkaline water electrolyzer was developed. This method includes an algorithm to automatically draw a pseudo-hydrogen bubble image based on actual hydrogen bubble images for training the CNN. This approach resolves the challenge of training models with a large amount of annotation dataset to achieve high-precision inference with CNN. With this method, it is possible to collect the amount of data needed to train the model by simply cutting out 12 bubbles. The method required approximately 0.1 s/image for bubble detection and F1 score exceeds 0.841 for all test images.