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

Deep learning constitutes a significant recent advance in machine learning and has been particularly successful in applications related to image processing, where it can already surpass human accuracy in some cases. In this paper, the use of a convolutional neural network, AlexNet, pretrained on a database of images of common objects was used as is to extract features from flotation froth images. These features could subsequently be used to predict the conditions or performance of the flotation systems. Two case studies are considered. In the first, froth regimes in an industrial flotation plant could be identified significantly more reliably with the features generated by AlexNet than with previous state-of-the-art approaches, such as wavelets, grey level co-occurrence matrices or local binary patterns. In the second case study, the arsenic concentration in the batch flotation of realgar-orpiment-quartz mixtures could be predicted more accurately than was possible with features extracted by wavelets, grey level co-occurrence matrices, local binary patterns or by use of colour. These results suggest that feature extraction with convolutional neural networks 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.