多通道CartoonGAN下的图像风格动漫化

为解决真实图像转换为动漫风格图像出现的参数量大、图像纹理和颜色损失的问题,提出了一种多通道卡通生成对抗网络(MC_CartoonGAN).首先,使用HSCNN+(advanced CNNs for the hyperspectral reconstruction task)和遗传算法重新构建多通道图像数据集,丰富图像信息.其次,利用DenseNet网络进行特征复用减少参数的内存占用率及缓解梯度消失的问题.最后,引入多通道颜色重建损失函数,在保证了生成图像内容完整的情况下,降低了生成图像的颜色损失.实验结果表明,提出的多通道卡通生成对抗网络将真实图像转换成动漫风格图像的质量更优.

Animation of image style in multi-channel CartoonGAN

In order to solve the problems of large number of parameters and loss of image texture and color when real images are transformed into animation style images, this paper proposed a multi-channel cartoon generative adversarial networks(MC_CartoonGAN). Firstly, it reconstructed the multi-channel image data set by using HSCNN+ and genetic algorithm, and obtained more abundant image information. Secondly, it used DenseNet for feature reuse to reduce the memory consumption of parameters and alleviate the problem of gradient disappearance. Finally, it reconstructed the loss function by introducing multi-channel color to reduce the loss of the generated image while ensuring the integrity of the generated image content. Experimental results show that the proposed multi-channel cartoon generative adversarial network(MC_ CartoonGAN) has better quality in converting real images into cartoon style images.