Surface Vortex Image Segmentation in KR Desulfurization Based on Improved SegNet Model

Abstract

In the process of hot metal KR desulfurization stirring, the visible length of the agitator blade exposed in the vortex on the liquid surface can reflect the mixing and dispersion effect of desulfurizer. Therefore, the Surface vortex information o is of great significance to the desulfurization control parameters setting. In this paper, an improved SegNet model is proposed. Based on the SegNet network model, the encoder of the network is replaced with a lightweight network structure, and the deep separable convolution is used to replace the conventional convolution, so that the model is simplified and the number of parameters is greatly reduced. The activation function in the network is modified to Mish function to strengthen the feature extraction ability of the network. The experimental results show that, on the established KR desulfurization liquid surface vortex image data set, the processing speed of the trained model reaches 49ms per frame. The segmentation mean pixel accuracy of KR liquid surface vortex image reaches 98.97%. And the mean Intersection over Union(mIoU) reaches 98.23%. Compared with the SegNet network model, the segmentation accuracy and rapidness of the proposed model are improved to a certain extent.