Froth image based monitoring of platinum group metals flotation with vision transformers and convolutional neural networks

The use of computer vision systems to monitor flotation froths hass become widely established over the last few decades. Features extracted from images of the froth can be used as predictors of the operational state of the plant or key performance indicators of the system. The quality of the froth features is key to the success of these models and over the last decade, convolutional neural networks have been shown to provide markedly better results than traditional approaches based on engineered features. More recently, vision transformers have emerged as similar or better alternatives to convolutional neural networks on multiple benchmarks. Although their real power lies in the fact that they can be trained to extract ad hoc features from images, they are also highly competent feature extractors in their pretrained form. It is as such that the performance of vision transformers in froth image analysis was compared with that of convolutional neural networks and traditional engineered features. More specifically, a pretrained vision transformer, as well as a more advanced version, the Swin transformer, is compared with AlexNet and GoogLeNet, as well as features extracted with grey level co-occurrence matrices and textons in three different unsupervised process monitoring frameworks, based on principal components, and two variants of isolation forests. Two case studies from the platinum group metal industries were considered. In the first case study, the performance of the feature extraction methods could be ranked as follows: GLCM approximate to GoogleNet approximate to ViT < textons approximate to AlexNet approximate to SViT. In the second case study, GoogleNet, ViT and Swin performed similarly well. Of the multivariate statistical process monitoring frameworks, the one based on principal component analysis performed better than both variants of the isolation forests. However, visualisation of the image features with a tSNE algorithm suggests that nonlinear monitoring frameworks with vision transformer features in particular, may ultimately be more effective than the ones considered in this investigation.