Identification of Cells in Swirling Jets by Machine Learning

Abstract

The present paper deals with identification and detection of cells. To identify the cells is useful for the investigation of the non-steady phenomenon and can consider its reason and measure. In the previous method, although it is done to detect the cells by using FFT, it does not detect the generation process and identification in real time. This paper investigated how can recognize the cells by learning a lot of the images by convolutional neural network (CNN) and evaluated where CNN pays attention to by Gradient-weighted Class Activation Mapping (Grad-CAM). Then it investigated to detect the cells by Single Shot Multibox Detector (SSD) by learning the own cell in itself. In this case, CNN seemed like learning the cells. In contrast, it estimated its based on unrelated information by Grad-CAM. It caused by the bias of the input data, for example, the rate of magnification of images, the range of the output value for computational fluid dynamics and the lack of the dataset and the conditions of CFD. Although CNN detected the cells by using SSD, detection accuracy is low. As a result, in order to get high detection accuracy, it is required for the quantitative and qualitative improvement of the dataset.