Abstract:In allusion to these issues that concentrate grade modeling of froth image processing exists involving shortage of effective froth image samples, low accuracy in model detection, poor generalization ability and robustness in the process of mineral flotation, it is proposed that the model of generative adversarial network based on self-attention mechanism and variational autoencoder. Among them, the generator employs the variational autoencoder consisting of an encoder and decoder. And the coding layer introduces the self-attention mechanism, making the convolution operation can better capture the long-distance dependence to acquire the overall information and generate high-quality images. The checker embedded in the classifier not only has the function of discriminating true from false, but also accomplishes the goal of inspection. Experimental results indicate that the model has quite strong generalization ability and robustness, and effectively increases the accuracy of froth image identification.