Abstract: During flotation, the features of the froth image are highly correlated with the concentrate grade and the corresponding working conditions. The static features such as color and size of the bubbles and the dynamic features such as velocity have obvious differences between different working conditions. The extraction of these features is typically relied on the outcomes of image segmentation at the froth edge, making the segmentation of froth image the basis for studying its visual information. Meanwhile, the absence of scientifically reliable training data with label and the necessity to manually construct dataset and label make the study difficult in the mineral flotation. To solve this problem, this paper constructs a tungsten concentrate froth image dataset, and proposes a data augmentation network based on Conditional Generative Adversarial Nets (cGAN) and a U-Net++-based edge segmentation network. The performance of this algorithm is also evaluated and contrasted with other algorithms in this paper. On the results of semantic segmentation, a phase-correlation-based velocity extraction method is finally suggested.