Froth image analysis by use of transfer learning and convolutional neural networks

Significance 

Flotation is currently one of the most important separation technology methods in mineral processing. Basically, it is a modified gravity process in which finely ground ore is mixed with a liquid after which a chemical separation process is subsequent. The control of this process is intricate in that the crucial performance indicators of this process are immeasurable online. Fortunately, images of the froth itself are a good source of viable information regarding the operational state of the floatation system. Such information is crucial for the development of an online measuring system. A two-stage model technique is currently in use for such measurement purposes, however, a better approach would be to directly and automatically guide the feature extraction process based on the predictive power of the features in the model itself.

Recently, in a research paper published in Minerals Engineering journal, professor Chris Aldrich and his PhD student Yihao Fu at the Western Australian School of Mines- Curtin University in Australia demonstrated that supervised feature extraction can yield significantly better results than what could be achieved with features not directly extracted to achieve the same goal. Additionally, they showed that the aforementioned expectations could be achieved by utilizing convolutional neural networks that have been pretrained on image data from a different domain.

The research method employed entailed the use of convolutional neural network, AlexNet, pretrained on a database of images of common objects. The convolutional neural network was then used to extract features from flotation froth images. The research pair then considered two case studies where the features were subsequently used to predict the conditions/performance of the flotation systems. In the first case study, froth regimes in an industrial flotation plant were identified using AlexNet and in the second case study, arsenic concentration in the batch flotation of realgar-orpiment-quartz mixtures were predicted.

The authors observed that the conditions in an industrial platinum flotation plant could be predicted reliably by use of a convolutional neural network pretrained on image data from a completely different domain. They also noted that in the flotation tests with arsenic sulfides, the convolutional neural network obtained outperformed state-of-the-art approaches in multivariate image analysis and was capable of explaining approximately 77% of the variance in the concentration of the arsenic in the froth, despite the small set of samples available.

Chris Aldrich and Yihao Fu study presented a novel approach to feature extraction from froth flotation images based on transfer learning and a convolutional neural network, commonly referred to as AlexNet. It has been seen that the obtained predictions are significantly better than what could previously be attained with other algorithms. Altogether, the results suggest that deep learning neural networks, such as AlexNet, trained on complex data sets from other domains can serve as more reliable methods than previous state-of-the-art approaches to froth image analysis.

Reference

Yihao Fu, Chris Aldrich. Froth image analysis by use of transfer learning and convolutional neural networks. Minerals Engineering, volume 115 (2018) pages 68–78.

Go To Minerals Engineering

Check Also

Effects of third-order susceptibility in sum frequency generation spectra