Liu Jie, Wu Yang
School of Intelligent Equipment Engineering, Wuxi Taihu University, Wuxi, China.
PLoS One. 2025 Jun 24;20(6):e0325784. doi: 10.1371/journal.pone.0325784. eCollection 2025.
Addressing the issue of insufficient key feature extraction leading to low recognition rates in existing deep learning-based flow pattern identification methods, this paper proposes a novel flow pattern image recognition model, Enhanced DenseNet with transfer learning (ED-DenseNet). The model enhances the deep feature extraction capability by introducing a multi-branch structure, incorporating an ECA attention mechanism into Dense Blocks and dilated convolutions into Transition Layers to achieve multi-scale feature extraction and refined channel information processing. Considering the limited scale of the experimental dataset, pretrained DenseNet121 weights on ImageNet were transferred to ED-DenseNet using transfer learning. On a gas-liquid two-phase flow image dataset containing Annular, Bubbly, Churn, Dispersed, and Slug flow patterns, ED-DenseNet achieved an overall recognition accuracy of 97.82%, outperforming state-of-the-art models such as Flow-Hilbert-CNN, especially in complex and transitional flow scenarios. Additionally, the model's generalization and robustness were further validated on a nitrogen condensation two-phase flow dataset, demonstrating superior adaptability compared to other methods.
针对现有基于深度学习的流型识别方法中关键特征提取不足导致识别率低的问题，本文提出了一种新颖的流型图像识别模型，即具有迁移学习的增强密集连接网络（ED-DenseNet）。该模型通过引入多分支结构来增强深度特征提取能力，在密集块中融入ECA注意力机制，并在过渡层中引入空洞卷积，以实现多尺度特征提取和精细化通道信息处理。考虑到实验数据集规模有限，利用迁移学习将在ImageNet上预训练的DenseNet121权重转移到ED-DenseNet。在包含环状流、泡状流、搅拌流、分散流和弹状流型的气液两相流图像数据集上，ED-DenseNet的整体识别准确率达到97.82%，优于诸如Flow-Hilbert-CNN等当前最先进的模型，尤其是在复杂和过渡流场景中。此外，该模型的泛化能力和鲁棒性在氮气冷凝两相流数据集上得到进一步验证，表明其与其他方法相比具有卓越的适应性。
