注意机制生成对抗网络单通道盲图像分离

Single channel blind image separation based on attentional generative adversarial network

针对单通道极端欠定状况及传统盲源分离(blind source separation,BSS)方法难以克服源信号独立、非高斯分布等多项约束和先验知识缺乏等问题,基于生成对抗模型的网络架构,提出了一种基于注意机制的单通道盲图像分离方法,将基于流形排序的视觉注意机制嵌入到分离网络中,增强目标的关键信息,以生成思想迭代分离混合图像。实验结果表明,融合注意机制的生成对抗网络能仅从单源数据中学习,无需多种先验约束,比经典盲源分离方法有更高的分离精度,与已知分布的神经网络分离方法相比,能更有效地分离单通道混合图像。

Traditional blind source separation(BSS)algorithms have multiple constraints such as independent source signals,non-Gaussian distribution,and lack of prior knowledge.Therefore,a single-channel blind image separation algorithm was proposed based on attention mechanism GAN in this work,and it would have more hopeful prospects in blind image separation task.The algorithm embedded the visual attention mechanism based on manifold sorting into the separation network to enhance the key information of the target.The algorithm extracted the features,strengthened the edge guidance,and separated the mixed images with the generative idea.Experimental results show that the adversarial network incorporating attention mechanism learns from a single source of data,it is significantly better than the classic blind source separation algorithm and the neural network separation algorithm with known distribution.