Abstract: The structure of simple generated countermeasure network is optimized, whichis used to obtain visual saliency map more effectively through antagonistic case training, so as to reduce false positive and improve saliency. The network model still follows the traditional generation countermeasure network structure. In the first stage, it is composed of a generator built with residual structure. The weight of the model is calculated by the back-propagation of the binary cross entropy loss (BCE) of the down sampling version of saliency map, and the more effective saliency map is obtained by training. The prediction results are classified by the trained discriminator network between saliency map and truth graph. The experiment shows that the improved ability to generate predictive saliency map of generators in the countermeasure network can improve the performance of the whole network, and it is also ahead of other saliency map prediction models.