A froth image segmentation method via generative adversarial networks with multi-scale self-attention mechanism

Froth image segmentation is the most popular method for bubble size measurement. However, in the froth flotation, froth images are with high-complexity and labeling the froth images with true bubble edges requires high human-labors and professional experiences. This causes that current convolutional neural network-based segmentation methods cannot segment well the bubbles accurately. Therefore, a generative adversarial network with multi-scale self-attention mechanism is proposed for froth image segmentation. First, to reduce the high requirements of true labeled samples, generative adversarial network was used in the framework to make the proposed segmentation model run in a self-supervised learning manner. Next, to strengthen the learning ability of the generator to detect bubble boundaries with different sizes, the receptive field is enlarged through the convolutional down-sampling operation and self-attention gating units are added to different layers. Finally, according to the output image of the generator, true labeled image and original input image, the discriminator outputs a confidence map to improve the parameters of generator and determines final edges. Many experiments were implemented to validate the effectiveness of the proposed method. Compared with other current convolutional neural network-based segmentation methods, the accuracy of the proposed model has been increased at least by 5%.