融合频率分离通道注意力机制的真实图像风格迁移算法

Photorealistic style transfer algorithm fusing frequency separation channel attention mechanism

目前的真实图像风格迁移算法普遍存在过于注重提升图像真实感而忽视算法的风格化强度问题.为了解决这一问题,选择PhotoWCT~2算法作为基准算法,并在此基础上引入并优化了通道注意力机制,提出了一种融合频率分离通道注意力机制的真实图像风格迁移算法.算法使用离散余弦变换进行特征分解,将分解得到的不同频率分量并行输入共享感知机内部.在每个编码器块的相应卷积层后引入频率分离通道注意力机制,对通道域中不同尺度的特征进行自适应筛选,筛选出高价值的纹理和颜色特征,进而增强了算法风格化强度.对比实验表明,所提算法在保证图像细节信息不丢失的同时,增强了算法的风格化强度,在定性视觉效果和定量评价指标上都获得了较好效果.

Existing photorealistic style transfer algorithms generally tend to focus too much on enhancing the realism of the generated images,while neglecting the strength of stylization.To address this problem,the PhotoWCT~2 algorithm is selected as the benchmark algorithm,and a real image style transfer algorithm with fusion of frequency separation channel attention mechanism is proposed.The algorithm utilizes the Discrete Cosine Transform for feature decomposition and parallelly inputs the decomposed different frequency components into the shared perceptron.Frequency separation channel attention mechanism is introduced after the corresponding convolutional layers in each encoder block to adaptively filter the features of different scales in the channel domain,select high-value texture and color features,and enhance the style intensity of the algorithm.Comparative experiments demonstrate that the proposed algorithm enhances the style intensity while preserving the image details,and achieves good qualitative visual effects and quantitative evaluation results.