An unsupervised method for extracting semantic features of flotation froth images

Abstract

• A new network structure maps the froth images data space to the latent semantic space. • The rich semantic knowledge of froth images can be automatically discovered. • The semantics can be visually interpreted by manipulating the froth image features. • The feature extraction method uses the unlabelled froth images effectively. • The extracted physical features perform well in flotation circuit monitoring. The froth feature extraction plays an important role in flotation circuit monitoring. Given the low efficiency of hand-crafted features and the poor interpretability of high-dimensional features extracted by convolutional neural networks, an unsupervised method for extracting human-understandable semantic features of flotation froth images is proposed in this paper. First, based on the combination of generative adversarial networks and autoencoder, we design a new network structure that maps the froth images data space to the latent semantic space. Then through network training with the historical froth image data, a matrix with rich latent semantics is constructed. Finally, the semantic features of froth images can be automatically extracted by decomposing the constructed matrix. As demonstrated in the industrial experiment, the extracted semantic features can not only be visually interpreted but also can be effectively used in flotation condition recognition and grade prediction. This is the first report that generative adversarial networks can be used to extract froth image features and improve the semantic interpretability of features.