关键词: 梯度策略, 多目标生成对抗网络, 帕累托最优解, 图像处理

Abstract:

To address the problems that the gradient-based multi-objective algorithm cannot be applicable for multi-objective Generative Adversarial Nets（GANs） with high dimension and multiple tasks. Then, considering a fact that solutions with cross-of-validation can be sub-optimal and it is hard to search global solutions in optimizing multi-objective GANs, this paper presents a multi-objective GANs Pareto optimality algorithm based on gradient strategy. The proposed algorithm uses hardware parameter sharing method and decomposes multi-objective optimization into multiple binary-objective optimizations. Then, it computes and determines all the weighted parameters and searches Pareto optimality along the gradient direction, which can yield exactly Pareto optimality. Theoretically, it has been proved that the proposed method can result in one Pareto optimality with detailed demonstration. Practically, the proposed algorithm has been applied into common sub-fields of image processing to compare the source and proposed algorithm in the same setting. The experimental results show that the proposed algorithm has outperformed than source algorithms when the number of tasks is over 2.

Key words: gradient strategy, multi-objective generative adversarial nets, Pareto optimality, image processing