Preconcentration greatly reduces the ore mass input to the processing plant while upgrading the grade. One of the commonly used methods of preconcentration is optical sorting. These methods rely on optical sensors which classify the material stream into “accept-reject” streams. Currently, it is common to use near infrared (NIR) or colour cameras to classify streams based on reflectance or colour thresholds set by experts. The imaging is done through line-scan sensors where the material is scanned on a belt or a chute. Then, a separation apparatus like a diverter gate or air jet is controlled to separate the streams according to the classification made by the algorithm. Separation efficiency is dependent on the classification algorithm.
Existing classification algorithms rely on low-level feature discrimination like observing individual pixel values for reflectance or colour hues. This limits discrimination between minerals with low contrast in colour. Human experts perform mineral segmentation using texture, colour distribution and shape, but it is difficult to translate these visual rules to mathematical thresholds. Deep learning methods like Convolutional Neural Networks (CNNs) provide high-level feature discrimination similar to a human expert. Instead of relying on pre-defined features, the CNNs learn complex features from the dataset. CNNs can construct a hierarchy of features based on pixel clusters to learn texture, shapes, and colour spectrums of the mineral grain surface to provide better separation.
In this study, a state-of-the-art deep learning instance segmentation method is used to outline boundaries and classify grains from the background (mineral segmentation). The algorithm is trained on images collected from a flotation feed of a gold deposit, where the gold is mechanically locked inside pyrite. This method can process on-line video inputs at 5 frames per second to provide a live grade. The proposed method is inexpensive as it only uses a camera and a desktop computer.