超燃冲压发动机燃烧流场纹影图像无监督分割方法

超燃冲压发动机燃烧室流场纹影图像常存在大量噪声信号，如何高效、准确提取燃烧流场图像的主要波系结构成为当前亟需探索的问题。以超燃冲压发动机燃烧室内冷流到氢燃料点火阶段流场为研究对象，基于深度神经网络方法，发展一种燃烧室内流场的关键波系结构快速识别方法。首先，采用基于图论的超像素分割方法对纹影图像进行聚类分割，为语义信息明显相同区域分配伪标签；其次，设计了一种全卷积特征提取神经网络，并使用残差结构对各个通道进行加权，提取纹影图像高级语义特征；最后，使用交叉熵目标函数优化网络模型，并通过阈值滤波操作去除噪声像素点，提升语义分割效果。结果表明：与K-means及自适应高斯阈值方法相比，本文提出方法在准确率、召回率、F1分数和交并比指标性能明显是最优的，能够准确完成燃烧流场纹影图像语义分割任务，可以更加清晰地反应流场内的主要波系和剪切层结构

Unsupervised Segmentation Method for Schlieren Image of Scramjet Combustion Flow Field

There are often a large number of noise signals in the schlieren image of flow field in scramjet combustion chamber. How to efficiently and accurately extract the main wave system structure of the combustion flow field image has become a problem that needs to be explored. Taking the flow field from the cold flow in the combustion chamber of scramjet engine to the stage of hydrogen fuel ignition as the research object, based on the deep neural network method, a rapid identification method of the key wave system structure of the flow field in the combustion chamber is developed. First, the super-pixel segmentation method based on graph theory is used to cluster and segment the schlieren image, and pseudo-labels are assigned to the regions with obvious semantic information; Secondly, in order to extract the high-level semantic features of the schlieren image, a fully convolutional feature extraction network was designed, and the semantic channels were weighted by the residual structure; finally, the cross-entropy objective function is used to optimize the network model, and the threshold filtering operation is used to remove noise pixels to improve the semantic segmentation effect. The results show that: compared with K-means and adaptive Gaussian threshold method, the method proposed in this paper is obviously the best in accuracy, recall, F1 score and intersection ratio index performance, and can accurately complete the semantics of the combustion flow field schlieren image The segmentation task can more clearly reflect the main wave system and shear layer structure in the flow field.