CBA-GAN: Cartoonization style transformation based on the convolutional attention module

Feng Zhang, Huihuang Zhao, Yuhua Li, Yichun Wu, Xianfang Sun

Abstract: Highlights•A novel CBA-GAN model is proposed in which unpaired real photos and cartoon images are used as training sets. The method uses an attention mechanism to increase expressive power and pay attention to important features to suppress unnecessary features.•An improved boxed model based on the block attention module is proposed. In the generator, in order to distinguish the importance of image features, a U-shaped network based on a lightweight convolutional attention module is designed.•To solve the problem of image pixel overflow during the model training process, a Tanh activation function is added to the generator to normalize the pixel value to [−1,1]. To keep the sharp edges in the discriminator network, an edge discriminator is designed to distinguish the edges of the image.