Estimation of Slag Removal Path using CNN-based Path Probability of Ladle Image Blocks

De-slagging is a task of removing slag on the surface of molten metals, such as steel, in a ladle. In this paper, we propose a method of slag removal path estimation using CNN (Convolution Neural Network) to automate de-slagging task using a robotic machine. From a sequence on images captured from the top of the ladle, we first extract the 2-dimensional trajectory of the slag removal motion of an experienced human operator. Then several image blocks are obtained at sample points along the removal trajectory to train a neural network. The output of the network consists of four labels which represent the probability of four different removal directions of an input image block. To test the trained neural network, we uniformly divide a test ladle image to a fixed-size block with a given stride value. All image blocks are tested and the probability of the four directions are determined and recorded by the trained network. By multiplying the slag probability with the removal direction probability, joint probability of slag removal direction (JPSRD) is introduced. Finally, a slag removal path is estimated by applying the backward tracing method from the endpoint of the ladle so that the estimated path yields the highest JPSRD. A curve fitting is then applied to make smooth slag removal path. The path decision accuracy of an image block is about 90%. We also compare the estimated a slag removal path with that of the experienced operator.