基于注意力的生成对抗网络图像超分辨率重建  

Image Super-Resolution Reconstruction Based on Attention Generative Adversarial Networks

作　　者：张惠君 李桐 ZHANG Huijun;LI Tong(School of Information Engineering,Beijing Institute of Graphic Communication,Beijing 102600,China)

机构地区：[1]北京印刷学院信息工程学院,北京102600

出　　处：《北京印刷学院学报》2025年第3期56-62,共7页Journal of Beijing Institute of Graphic Communication

摘　　要：针对部分模型未能有效利用特征信息,存在模型训练不稳定、生成图片模糊、质量不高等问题,对SRGAN模型进行改进,提出一种融合注意力机制的WGAN图像超分辨率重建算法(CBAMWGAN)。在SRGAN的残差网络中融合注意力机制(CBAM)模块,使网络自适应调整各通道权重,关注图像中的重要区域,以更好地表达高频特征;同时去除生成器中的BN层和使用Leaky ReLU激活函数,进一步提升了模型的计算效率和生成图像质量;最后引入WGAN的思想,用Wasserstein距离代替判别器中的二分类交叉熵损失,提高了网络训练的稳定性。将训练好的模型在Set5、Set14、BSDS100三个数据集上进行测试,并将重建效果与Bicubic、SRCNN、VDSR、SRGAN进行对比。实验结果表明,CBAMWGAN模型无论在客观指标上,还是主观视觉效果上均优于对比模型。Aiming at the problems that some models fail to use the feature information effectively,the model training is unstable,the generated images are fuzzy and the quality is not high,an improved SRGAN model is proposed,and a WGAN image super-resolution reconstruction algorithm(CBAMWGAN)with attention mechanism is proposed.The CBAM module is integrated into the residual network of SRGAN,so that the network adaptively adjusts the weight of each channel and pays attention to the important areas in the image to better express the high-frequency features.At the same time,the removal of BN layer in the generator and the use of Leaky ReLU activation function further improve the computational efficiency of the model and the quality of the generated image.Finally,the WGAN idea is introduced and Wasserstein distance is used to replace the binary cross entropy loss in the discriminator,which improves the stability of network training.The trained model was tested on Set5,Set14 and BSDS100 datasets,and the reconstruction effect was compared with Bicubic,SRCNN,VDSR and SRGAN.The experimental results show that the CBAMWGAN model is superior to the comparison model in both objective index and subjective visual effect.

关 键 词：图像超分辨率重建 注意力机制 生成对抗网络 Wasserstein GAN 

分 类 号：TP273[自动化与计算机技术—检测技术与自动化装置]