Rapid Recognition and Semantic Segmentation of Molten Steel Distribution Images on a Steel Ladle Surface based on the Improved SqueezeNAS Network

Techniques of real-time image recognition of slag-steel distribution on the surface of a steel ladle under bottom argon-blowing (BAB) conditions are vital for monitoring and adjustment of BAB parameters and improving refining efficiency at industrial sites. An improved SqueezeNAS, a rapid semantic segmentation network, was proposed for the segmentation of molten steel distribution images on the surface of a steel ladle under BAB. The improved SqueezeNAS changed the inverted residual basic unit for search, reduced the search space, and introduced a hybrid dilated convolution (HDC) into the search space to avoid the grid effect of dilated convolution. By taking pictures of the molten steel distribution on the slag-steel surface under steel ladle BAB at a steel mill in real-time, a distribution image segmentation dataset of molten steel on the surface of a steel ladle in industrial scenarios was established, and the images were preprocessed to obtain target images. Additionally, the target images were segmented using the improved SqueezeNAS, being further recognized and classified by deep learning. The mean intersection over union (mIoU) of the dataset images of the steel ladle under BAB was 98.17%, indicating superior segmentation performance compared with U-Net, SegNet, SqueezeNAS network models. The average blowing time shortened 2 min or more and the increment of nitrogen was less after the improved SqueezeNAS segmenting used in the BAB monitor system for the same grade steels.