Dissertations - M Tech (CS)

Permanent URI for this collectionhttp://164.52.219.250:4000/handle/10263/2147

These Dissertations were submitted in partial fulfilment of the requirements for the award of M TECH (Computer Science) Degree of Indian Statistical Institute

Browse

Filters

20221

Settings

Author: search.filters.author.Biswas, LunaSubject: search.filters.subject.Image processing

Search Results

Now showing 1 - 1 of 1
  • No Thumbnail Available
    Item
    Coke characterization: Segmentation of pores and constituents from microscopic images of Coke
    (Indian Statistical Institute, Kolkata, 2022-07) Biswas, Luna
    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.