dc.description.abstract |
Coke is mainly used in steel industry as a fuel and a reducing agent for melting iron
in the blast furnace, since it generates intense heat but little smoke. The quality of
the coke material (like porosity, wall thickness, texture etc., as seen in a microscopic
image of coke) affects the performance of blast furnace impacting the profit/loss of
the industry. Therefore it is important to determine the structure and porosity of
coke on a large scale. Manual process of coke characterisation is costly and slow. Automation
of coke characterization, from microscopic images of cokes, is beneficial for
the steel industry. An attempt has been made to calculate porosity of coke from the
images, and produce semantic segmentation of the coke images into different types of
metallurgical textures like inert, incipient, circular, lenticular etc. A shallow convolutional
neural network (CNN) was trained with annotated coke images using cross
entropy loss (between the probability distributions of the predictions out of the CNN
and the target as per annotation, for different classes). A new contrastive loss function
has been written, that maximises entropy between the probability distribution
of a training sample with another sample belonging to a different class, in addition to
minimising entropy loss between the probability distributions of the predictions and
the target. This new loss function enables faster learning, and useful when quantity
of annotations for training a model, is less. A shallow CNN model obtained higher
accuracy in prediction of class for each pixel of the coke images, and the granularity
of semantic segmentation was reduced when trained using this novel loss function. |
en_US |