Froth velocity extraction method from froth image for flotation with large throughput based on composite deep learning network

Abstract

Froth velocity is a crucial variable in froth flotation and its accurate detection is vital to process control. The velocity extraction is a challenging work under large throughput conditions due to the complex motion of the froth. Therefore, in this paper, a novel method called composite deep learning network (CDLN) is proposed, which is mainly composed by a cascade of pre-trained supervised network and an unsupervised network. The supervised network is pre-trained on a benchmark dataset and then is used to extract an initial flow field of the froth motion. The unsupervised network is trained on industrial froth videos with occlusion modeling to get a calibration flow field. The two flow fields are then fused to obtain the final flow field. The test results show that the proposed method can extract the froth velocity more accurately and faster than other popular methods through industrial froth videos under large throughput.