Abstract:

With the advancement of generative adversarial network research, the computational amount of the network model increases sharply, its own training instability still exists, and the quality of the generated image also needs to be improved. To solve the problems, a lightweight generative adversarial network is proposed, which introduces multi-scale gradient structure to solve the problem of unstable training. By combining the ideas of self-attention mechanism and dynamic convolution, the cyclic module and image enhancement module are used to improve the learning ability of the model under the premise of keeping fewer parameters. The verification experimental results show that the inception score is 2.75 and the FID is 70.1 on CelebA data set, the inception score is 2.61 and the FID is 73.2 on LUSN data set, which is better than that of the classical models such as SAGAN and DCGAN, and verifies the feasibility and performance of the proposed algorithm.

Key words: multi-scale gradient, dynamic convolution, cyclic block, half-attention mechanism, sparse attention, convolutional neural networks, deep learning, image generation, generative adversarial net