一种结冰显微图像中气泡的自动提取方法  

Automatic extraction method of air bubbles in icing microscopic images

作　　者：赵红梅 彭博[1] 周志宏[2] 易贤[3] ZHAO Hongmei;PENG Bo;ZHOU Zhihong;YI Xian(School of Computer Science,Southwest Petroleum University,Chengdu 610500,China;College of Architecture and Environment,Sichuan University,Chengdu 610000,China;Key Laboratory of Icing and Anti/De-icing,China Aerodynamics Research and Development Center,Mianyang Sichuan 621000,China)

机构地区：[1]西南石油大学计算机科学学院,成都610500 [2]四川大学建筑与环境学院,成都610000 [3]中国空气动力研究与发展中心结冰与防除冰重点实验室,四川绵阳621000

出　　处：《航空动力学报》2023年第11期2738-2746,共9页Journal of Aerospace Power

基　　金：国家自然科学基金(12072213);国家科技重大专项(J2019-Ⅲ-0010-0054)。

摘　　要：针对传统图像分割方法提取结冰显微图像中的气泡漏检率高和无法分离粘连气泡的问题,提出深度神经网络和传统分割算法相结合的方法。基于Attention U-Net网络,采用双分支融合预测策略对结冰显微图像中的气泡进行提取。针对部分气泡粘连问题,引入直方图均衡化和局部极小值,采用基于距离变换的分水岭算法,对结冰显微图像中粘连气泡进行二次分割。实验结果表明:通过双分支融合预测的Attention U-Net网络,对不同结冰显微图像中的气泡提取更精确,特别是对于较小气泡的检出率更高。测试图像的像素精度、平均像素精度、平均交并比和频权交并比分别达到0.9767、0.8916、0.8188和0.9575。基于距离变换的分水岭算法在粘连气泡分割中也展现了良好的性能,为后续统计气泡个数、面积等特征提供可量化的数据支撑。In view of the problems of high missed detection rate and inability to separate adhering bubbles in the extraction of icing microscopic images by traditional image segmentation methods,a method combining deep neural network and traditional segmentation algorithm was proposed.Based on the attention U-Net network,a two-branch fusion prediction strategy was adopted to extract the bubbles in the icing microscopic images.For some bubble adhesion problems,histogram equalization and local minima were introduced,and a watershed algorithm based on distance transformation was used to segment the adhesion bubbles in icing microscopic images twice.The experimental results showed that the twobranch prediction prediction Attention U-Net network was more accurate for the extraction of bubbles in different icing microscopic images,especially the detection rate for smaller bubbles was higher.The pixel accuracy,mean pixel accuracy,mean inter-section over union and frequency weighted intersection over union of the test images reached 0.9767,0.8916,0.8188 and 0.9575,respectively.The watershed algorithm based on distance transformation also showed good performance in the segmentation of sticky bubbles,providing quantifiable data support for the subsequent statistics of the number and area of bubbles.

关 键 词：飞机结冰 气泡提取 图像分割 分水岭算法 双分支融合预测模型 

分 类 号：V219[航空宇航科学与技术—航空宇航推进理论与工程]