Recognition of flotation froth conditions with k-shot learning and convolutional neural networks

Abstract

In this study, previous work on k-shot learning in flotation froth image analysis using small sets of froth images, is extended. As before, image synthesis is used to augment these data sets, but in addition, fine-tuning of convolutional neural networks, as well as smaller data sets with as few as 10 samples per class are considered. Two convolutional neural network models, namely AlexNet and GoogLeNet, were compared and the latter demonstrated better performance generally. Both performed better than a traditional approach based on the use of image features derived from local binary patterns. Fine-tuning of the convolutional neural networks markedly improved their performance compared to pure transfer learning without fine-tuning. Moreover, a case study with platinum froth images showed that as few as only 10–30 real images per class are needed to achieve reasonably good classification performance, i.e. the ability to recognize four different classes of froths with an accuracy of 63%–84%, if AlexNet or GoogLeNet with fine-tuning is used. Using 100 real images per class with a fine-tuned GoogLeNet resulted in predictive accuracy approaching 92%. These results are comparable to those achievable with traditional approaches to feature extraction based on training data sets up to two orders of magnitude larger.