Iron Ore Pellet Size Analysis: A Machine Learning-Based Image Processing Approach

In this article, an ensembled convolutional neural network (CNN)-based algorithm is proposed for iron ore pellet size analysis. A new customized CNN is ensembled along with VGG16, MobileNet, and ResNet50. The algorithm uses images captured from the inside area of a pelletizer disk to directly estimate the pellet size class instead of employing a circle fitting method. An image data set is created for different size classes (VERY SMALL, SMALL, GOOD, BIG, and VERY BIG) by using a novel cropping and resizing method, where the mean sizes are 7–19 mm, at a step of 3 mm. Actual industrial images captured using an industrial camera during pellet production are used as the seed images for the data set creation and training, validation, and testing of the proposed model. Through extensive experimentation, it is shown that the proposed classification algorithm can achieve about 96%–99% accuracy, whereas if the custom-designed CNN is not used, the accuracy obtained is about 77%–95% for different classes. In addition to this, the proposed trained network is validated with pellet images of an intermediate mean size to show that the trained model is not overfitted.