多通道特征增强风格迁移算法

A Style Transfer Algorithm of Multi-channel Feature Enhancement

针对风格迁移算法在迁移过程中图像内容特征映射出错,无法保证特征映射完整性,使生成图像出现内容失真的问题,提出多通道特征增强风格迁移算法。该算法在CycleGAN(Cycle Generative Adversarial Networks)风格迁移网络基础上,引入多通道特征增强机制,增强每个通道对图像特征的非线性表达能力,不仅可保持通道的独立,还能提取更加丰富的内容特征,在确保特征映射一致的同时,提高风格迁移质量。经仿真实验表明,本文所提算法与CycleGAN相比:在风景方面,冬-夏季节场景数据集在评价指标IS和FID上分别提高6.2%和25.7%;在静物方面,苹果-橘子水果静物数据集在评价指标IS和FID上分别提高9.3%和24.4%。

For the problem of image content feature mapping error in the process of style transfer algorithm,which cannot guarantee the integrity of feature mapping and can cause content distortion in the generated image,multi-channel feature enhanced style transfer algorithm(MCGAN)is proposed.Based on CycleGAN(Cycle Generative Adversarial Networks)style migration networks,this algorithm introduces multi-channel feature enhancement mechanism to enhance the nonlinear expression ability of each channel to image features.This method can keep the channel independent and extract richer content features,not only ensuring the consistency of feature mapping but also improving the quality of style migration.While ensuring the consistency of feature mapping,the quality of style migration is improved.The experimental results show that,compared with CycleGAN,in terms of scenery,MCGAN improved the evaluation indexes IS and FID by 6.2%and 25.7%in winter and summer scenes datasets;in terms of still life,the evaluation indexes IS and FID are improved by 9.3%and 24.4%in the apple and orange fruit still life datasets.