摘要
现有的Anammox颗粒污泥性能评估方法中,生化反应等效测定的方法较为精确但耗时长,基于图像颜色建模的方法足够快但不够精确.因此本研究基于实例分割任务,提出一种新的结合颗粒污泥视觉特征(如颗粒污泥颜色、尺寸、粗糙度等)进行性能评估的模型,使其能够在快速评估的同时保证较高精度.首先设计采样方案,收集颗粒污泥图像后进行图像标注,进而构建颗粒污泥图像数据集.针对图像场景中样本不均衡和小目标信息难提取的难点,本研究对无锚实例分割方法BlendMask进行改进,在主干网络上结合可变形卷积提取小目标信息,在头部网络上提出自适应样本选择策略平衡样本,并重新设计了后处理、损失函数与数据增强等模块,使改进后的BlendMask方法能快速精确提取图像场景中的颗粒污泥个体.然后获取颗粒污泥的视觉特征,设计视觉特征与性能的关联模型,判别输入图像的性能类别.相比之前的研究,本文提出的方法能够快速精确地获得颗粒污泥的性能评估.最后比较了所提出改进的BlendMask方法和先进方法在颗粒污泥图像分割和性能评估方面的表现,实验结果表明,改进的BlendMask和性能评估模型精度分别提升了3.93%和1.49%,方法取得了最先进的性能.
The method of measuring the biochemical reaction equivalence of Anammox granular sludge is accurate but time-consuming,and the method based on image color modeling is fast but inaccurate enough.Therefore,the study proposes a new model combining the visual features of granular sludge(such as color,size,and roughness)for performance evaluation based on the instance segmentation task,which can guarantee high accuracy while performing a fast evaluation.A sampling scheme is first designed to collect the granular sludge images,and then image annotation is performed to obtain the granular sludge image dataset.This study proposes an improved BlendMask anchor-free instance segmentation method,which combines deformable convolution on the backbone network to extract small target information,adopts an adaptive sample selection strategy to balance samples on the head network,and redesigns the modules of postprocessing,loss function,and data augmentation to quickly and accurately extract the granular sludge individuals in image scenes.Then,the association model of visual features and performance is designed to obtain the visual features of granular sludge and discriminate the performance class of the input image.The proposed method can obtain a faster and more accurate evaluation of the performance of granular sludge.Finally,the effectiveness of the proposed method is verified.Compared with the previous methods,the improved BlendMask and the performance evaluation model have improved the accuracy of granular sludge image segmentation and performance evaluation by 3.93%and 1.49%,respectively,achieving the most advanced performance.
作者
田腾飞
彭鑫
钟伟民
厉巍
章旭
TIAN TengFei;PENG Xin;ZHONG WeiMin;LI Wei;&ZHANG Xu(Key Laboratory of Smart Manufacturing in Energy Chemical Process,Ministry of Education,Chinese Academy of Sciences,Shanghai 200237 China;School of Information Science and Engineering,East China University of Science and Technology,Shanghai 200237,China;National Engineering Research Center of Industrial Wastewater Detoxication and Resource Recovery,Shanghai 200237,China;School of Resource and Environmental Engineering,East China University of Science and Technology,Shanghai 200237,China)
出处
《中国科学：技术科学》
EI
CSCD
北大核心
2023年第7期1150-1162,共13页
Scientia Sinica(Technologica)
基金
国家自然科学基金重大项目课题(编号:61890930-3)
国家自然科学基金(批准号:62173145)
国家杰出青年科学基金(批准号:61925305)
上海市浦江人才计划(编号:21PJ1402200)
上海人工智能实验室资助。
关键词
无锚实例分割
性能评估
深度学习
厌氧氨氧化颗粒污泥
anchor-free instance segmentation
performance evaluation
deep learning
Anammox granular sludge
