基于无线传感网络的污水微生物非刚体运动在线识别

Accurate extraction of the microbes with non-rigid motion in sanitary sewage and automatic identification of the microbes under complex background is a hot topic in the field of environmental protection, especially for sewage treatment. In recent years, image analysis technology has been widely used in sewage microbial identification. However, in the cases of light changes, dynamic background, especially for movement deformation, the recognition performance will fall sharply. In addition, it is also difficult to realize real-time and on-line detection. Therefore, this project will focus on exploring novel strategy based on wireless video sensor network, from target detection, dynamic routing, motion deformation model and image classifier design, for the fast identification of microbes with non-rigid motion. A new hierarchical background subtraction algorithm will be proposed in this project, where a MRF-MAP framework can explicitly model the spatial-temporal context to achieve more accurate foreground detection. A data relevance dynamic routing protocol will also be designed to ensure the more important image data to be transmitted accurately and timely, which will defilitely prolong the network lifetime. We will establish motion manifold models to describe different types of microbial non-rigid motion and design a novel deep learning classifier accompanied with deep belief neural network trained by multiply clusters and greedy algorithm, which will help us obtain higher recognition accuracy and less recognition time. This work will be helpful for wastewater treatment and energy saving optimization control, and will promote the microbial detection technology development. Therefore, it has wonderful scientific research and application value.

精确提取生活污水中非刚体运动微生物，并在复杂环境背景下实现微生物自动识别，是当前环境保护，尤其是污水处理领域的前沿课题。尽管图像分析技术在污水微生物种群自动识别中受到广泛关注并在近年获得快速发展，但当光线变化、动态背景，特别是运动形变时识别性能急剧下降，且难以实现在线实时检测。本项目拟探索微生物非刚体运动的无线网络识别方法，在运动目标检测优化、无线动态路由协议、非刚体运动建模和分类识别等方面进行深入研究，运用马尔科夫决策原理构建一种新时空域分层码本背景模型，精确提取复杂背景下运动微生物；设计图像数据相关动态路由协议，保证重要数据准确及时传输，延长网络寿命；构造微生物非刚体运动流形模型，并引入深度学习模型，结合立体特征簇和贪婪训练算法改进深度信念网络分类器，提高识别精度和效率。本项目研究结果可为污水处理节能优化控制提供依据，并对微生物检测技术发展起到促进作用，具有良好理论研究和应用价值。

基于无线传感网络的污水微生物非刚体运动在线识别的研究难点在于如何充分利用资源受限的无线传感器以及显微视觉提供的特定信息与先验知识，对复杂环境背景下种类繁多、形态多样的污水微生物进行自动辨识与统计，以及辨识的准确率和实时性。针对以上问题，本项目首先研究了非刚体运动微生物复杂背景模型，设计了基于新时空域自适应分层码本和3D自组织映射神经网络两种运动目标自动检测算法，有效解决了光线变化、动态背景、运动形变和目标遮挡等干扰问题；其次，本项目研究了聚类分簇与反应式随机路由原理，分别提出了基于萤火虫元启发式算法的动态分簇路由协议，以及分级考虑能耗与传输时延的无线机会路由协议，在保证重要数据准确及时传输的同时，有效降低了网络能耗，延长了网络寿命；再次，本项目采集收集整理了15种常见指示性污水微生物图像与视频数据，制成图谱库，设计了基于边缘算子与边界连接模型的目标轮廓提取算法，获取了清晰完整且连续的非刚体运动微生物形状特征。最后，利用仿生群智能计算与机器学习技术，分别设计了基于多阈值粒子群优化分割的微生物目标自动识别计数方法，以及多层神经网络的目标自动识别计数方法，实验结果验证了方法的准确性、鲁棒性与实时性。本项目研究结果对微生物自动检测技术和无线传感网络技术的推广应用，以及污水处理的节能优化控制起到一定的促进作用。