基于条件边界平衡生成对抗网络的河流表面流速估测

Estimation of river surface flow velocity based on conditional boundary equilibrium generative adversarial network

针对不同流速类类间差异小而造成的分类困难问题,提出条件边界平衡生成对抗网络和多特征融合的卷积分类网络,分别进行流速图像的生成和分类.为了达到数据增强效果,引入标签机制和验证模块实现相应类别图像数据的拟合与生成;为了加强图像不同纹理特征信息对流速估测的影响,引入多特征融合机制对所有真实样本和生成伪样本进行特征提取和流速识别,实现对差异性较小的图像的分类.将该方法应用于实际的河流表面流速估测,结果表明,在图像生成模块中,引入的标签信息和验证机制在一定程度上能强制引导模型的数据生成方向;在图像识别模块中,引入的多特征融合机制使所提出方法相较于其他方法,在差异性较小的水流图像的识别上更具鲁棒性.

Aiming at the difficulty in image classification due to the high similarity between different flow speeds,a conditional boundary equilibrium generative adversarial network and a convolutional classification network based on the multi-feature fusion were proposed to realize the generation and the classification of flow velocity images,respectively.A labeling mechanism and a verification module were introduced to realize the fitting and generation of corresponding category images,in order to achieve data enhancement.To enhance the impact of different texture features on velocity estimation,a multi-feature fusion layer was introduced to realize the feature extraction and the flow velocity recognition so as to realize the classification for images with small differences.The proposed method was applied to the actual river surface velocity estimation.Results demonstrate that the added tag information and the verification module can guide the data generation of corresponding class to a certain extent in the image generation module.Compared with other methods,the multi-feature fusion mechanism makes the proposed classifier more robust in identifying flow velocity images with small differences.