Image processing applications for integrated steel plants

Abstract

Manufacturing processes in an integrated steel plant involve handling, tracking and characterization of bulk raw material & products at different stages of operations. Computer vision is a promising and cost-effective technology which is finding application at an ever-increasing pace in many of the solutions used in the integrated steel plant. Imaging solutions are used for applications like material identification, size estimation, product tracking & quantification and quality inspection etc. in steel manufacturing process. This dissertation work applied three different flavors of image processing techniques to address three business cases across raw material and product handling in the steel plant. The entire project work can be arranged in three different phases as described below. In the first phase of the project, automatic estimation of the size distribution of coke particles using image analysis was attempted. This is done using the watershed algorithm for image segmentation to replace the existing technique of manual size determination using Sieve Analysis. The advantage of this is better accuracy, more reliability and productivity. In the second phase, we worked on the problem of Bar Counting. The bars are currently dispatched by their weights after being manufactured from the Bar Mill. The main problem with this is the pilferage taking place during transport. The objective of this project is to accurately count the bars using Image Processing algorithms. This time, we scaled up from the conventional algorithms to use machine learning techniques for object detection purposes. More than 97% accuracy is achieved until now. In the third and the last phase, we take up the problem of finding out the sizes of the stone fragments. Calculating the size distribution of rock particles is very crucial in getting to know whether the blasting done at the mining site is optimal or not. The segmentation task was very challenging as there was no particular shape or texture present for a specific stone fragment. Finally, this is done using the deep learning technique of mask R-CNN.