Classification of flotation froth images using neural networks with explanations

Computer vision systems have been widely used for analyzing flotation froth images. Nowadays, neural networks are considered the state-of art method in that field. There are many studies describing high accuracy of neural networks used for classifying froth images. However, few of them consider the latest network architectures and the problem of explainability for this domain remained unexplored. In our study we used a publicly available dataset of froth images to compare the accuracy of a few modern neural network architectures, including ResNet, EfficientNet, MobileNet V3 and Swin. Next, we used various explainable artificial intelligence (XAI) methods to explain the predictions of the networks. We found out that most of the models achieved a high classification accuracy, exceeding 97% on average, including some lightweight ones like MobileNet V3 Large. Some XAI methods like occlusion and GradCAM proved to be useful for identifying important regions while others like Integrated Gradients and GradientSHAP were less helpful.