Deep learning-based bubble detection with Swin Transformer
Swin Transformerを用いたディープラーニングによる気泡検出
上澤 伸一郎
; 吉田 啓之

Uesawa, Shinichiro; Yoshida, Hiroyuki
本研究では、重なり合う気泡の中から個々の気泡を検出・分割するために、Shifted window Transformer (Swin Transformer)を用いた深層学習ベースの気泡検出器を開発した。検出器の性能を検証するため、学習画像数を変えて平均適合率(AP)を計算した。APは、訓練画像の数が50枚以下の場合は、トレーニング画像の数の増加とともに増加したが、50枚を超える場合は一定であった。50枚以上ではSwin Transformerと一般的なCNNであるResNetのAPはほぼ同じであったが、学習画像が少ない場合はSwin TransformerのAPがResNetのAPを上回った。また、ボイド率が増加すると、Swin TransformerのAPはResNetの場合と同様の減少を示したものの、学習画像が少ない場合はSwin TransformerのAPが全てのボイド率においてResNetのAPを上回った。さらに、合成気泡画像で学習した検出器で、気泡流可視化実験の重なった気泡や変形気泡の検出が可能であることを確認した。このように、Swin Transformerを用いた新しい気泡検出器は、ResNetを用いた検出器よりも少ない学習画像で高いAPを得られることが確認された。
We developed a deep learning-based bubble detector with a Shifted window Transformer (Swin Transformer) to detect and segment individual bubbles among overlapping bubbles. To verify the performance of the detector, we calculated its average precision (AP) with different number of training images. The mask AP increased with the increase in the number of training images when there were less than 50 images but remained constant when there were more than 50 images. It was observed that the AP for the Swin Transformer and ResNet were almost the same when there were more than 50 images; however, when few training images were used, the AP of the Swin Transformer were higher than that of the ResNet. Furthermore, with regard to the increase in void fraction, the AP of the Swin Transformer showed a decrease similar to that in the case of the ResNet; however, for few training images, the AP of the Swin Transformer was higher than that of the ResNet in all void fractions. Moreover, we confirmed the detector trained with synthetic bubble images was able to segment overlapping bubbles and deformed bubbles in a bubbly flow experiment. Thus, we verified that the new bubble detector with Swin Transformer provided higher AP than the detector with ResNet for fewer training images.