Multi-scale neural network for accurate determination of the ash content of coal flotation concentrate using froth images

Flotation concentrate quality is strongly correlated with its froth. Therefore, froth images can be used to determine coal concentrate quality. Earlier studies on using coal flotation froth images to determine concentrate ash content ignore the multiscale characteristics of the images. This paper proposes a multiscale neural network (MSNet) to tackle this challenge. The MSNet model can process images with a wide range of resolutions and extract the image features from multiple scales for quality prediction. The model is validated using laboratorial and industrial datasets. Results show that the MSNet correlates with the laboratorial and industrial samples with R2 values of 0.9430 and 0.8288, respectively. The accuracy was further improved by transfer learning technique, resulting in R2 values of 0.9572 and 0.8647 for laboratorial and industrial samples, respectively. These results exhibit the good adaptability of the proposed MSNet. They also indicate that our MSNet is promising in real-world applications, promoting cleaner and more efficient coal production.