基于Res2Net-RetinaNet的活性污泥指示性微生物目标检测

Res2Net-RetinaNet-based Target Detection of Activated Sludge Indicator Microorganisms

活性污泥中原生动物、后生动物等指示性微生物是污水处理运行调控的重要指标。针对活性污泥微生物不同种类之间,小目标类微生物体型较小、微生物个体颜色背景和图像颜色背景相似的现象,提出基于Res2Net-RetinaNet的活性污泥指示性微生物检测方法。Res2Net-RetinaNet模型采用精度更高的新维度残差块Res2Net模块捕获原有特征的丰富信息。在主干网络输出的第1层引入通道和空间注意力机制CBAM,进一步帮助浅层特征信息在网络中流动。最后,在特征融合模块中引入深度超参数化卷积(Do-Conv),在不增加计算量的前提下持续加快模型的收敛。将所提方法应用于某污水厂采集数据中进行实验,结果表明:所提方法与Fast R-CNN、SSD、YOLOv3、YOLOv4、FCOS、CenterNet及RetinaNet等目标检测模型相比,检测精度最高(92.8%),相对于原始RetinaNet目标检测算法精度提升4.97%。

Indicative organisms such as protozoa and post-zoa in activated sludge are important indicators for both operation and regulation of wastewater treatment.Considering small target class’microorganisms small in size and the color background of individual microorganisms and the image color background similar among different species of activated sludge microorganisms,a Res2Net-RetinaNet-based detection method for activated sludge indicator microorganisms was proposed,including having new dimensional residual block Res2Net module with higher accuracy adopted in this model to capture rich information of the original features.In addition,both channel and spatial attention mechanism CBAM introduced into the first layer of the backbone network output further helps the shallow feature information flow in the network;and then,the deep hyperparametric convolution(Do-Conv)introduced into the feature fusion module continuously accelerates the convergence of the model without increasing the computational effort.The experimental results of applying the proposed method to data collected from a sewage plant in Shenyang show that,the detection accuracy of the proposed method can be up to 92.8%compared with the current target detection models(Fast R-CNN,SSD,YOLOv3,YOLOv4,FCOS,CenterNet and RetinaNet)and the accuracy can be improved by 4.97%compared with the original RetinaNet target detection algorithm.