关键词: 深度学习, 循环生成对抗网络, 空洞卷积, 多尺度特征聚合模块, 像素自注意力模块

Abstract: Aiming at the problem that the current convolutional neural networks usually  obtained multi-scale image features on the conditio
n of reducing receptive fields, and it was difficult to capture the important relationship between channels.  Combined with the features of cycle-generative adversarial networks structure, we proposed a new cycle-generative adversarial networks with multi-scale and self-attention mechanism. Firstly, VGG16 module was used to form U-Net structure in the generator to enhance the extraction of image feature information. At the same time, the down-sampling  and up-sampling  in the network were improved to improve the feature resolution and obtain more detailed information. Secondly, a multi-scale feature fusion block was designed. The multiple parallel dilated convolutions with different sampling rates were used to integrate the spatial information on different scales, and capture image information in multiple proportions while maintaining  a large receptive field of the image. Finally, in or
der to capture the feature dependencies in the spatial dimension and channel dimension, the pixel self-attention module was designed to model the semantic dependencies in the spatial dimension and channel dimension, so as to enhance the representation ability of image features and improve the quality of the generated sketch images.

Key words:  deep learning, cycle-generative adversarial networks, dilated convolution, multi-scale feature fusion block, pixel self-attention module