基于注意力机制的循环一致性生成对抗网络

Cycle-Consistent Generative Adversarial Networks based on attention mechanism

针对循环一致性生成对抗网络(Cycle-GAN)在图像风格转换任务上出现的纹理细节处理得不好、背景颜色保留较差等问题,并且缩小在配对图像数据集和非配对图像数据集上训练结果的差异,提出一种基于注意力机制的循环一致性生成对抗网络,在生成器网络中融入通道注意力机制(SE-Net),利用网络自主学习的方法得到每一个特征通道的重要程度,再分别赋予每个特征通道不一样的权重系数,以此来强调有重要特征的部分、抑制非重要特征的部分,使得不同特征和不同区域能够被生成器网络非均匀的处理。同时引入对比学习(CL),使网络能够学习到图像的更高层次的通用特征。实验结果表明,所提方法在horse2zebra数据集上取得了较好的结果。

Aiming at the problems of ineffective texture detail processing and ineffective background color retention in the image style conversion task of Cycle-Consistent Generative Adversarial Networks(Cycle-GAN),and reducing the difference between paired image dataset and unpaired image dataset,a Cycle-Consistent Generative Adversarial Networks based on the attention mechanism is proposed.Channel attention mechanism(SE Net)was incorporated into the generator network.The importance of each feature channel was automatically obtained through network learning,and then different weight coefficients were assigned to each channel,so as to strengthen the important features and restrain the unimportant ones,allowing the generation network to handle different features and different regions unevenly.At the same time,Contrastive Learning(CL)is introduced to enable the network to learn the higher level common features of images.The experimental results show that the proposed method has achieved good results on horse2zebra dataset.