摘要
针对单一图片的超分辨率提升作业,超分辨率生成对抗网络(SRGAN)技术在实现高度逼真纹理方面表现卓越。此外,大多数先进的SISR方法因参数量过大而无法应用于资源受限的设备。为了应对这些挑战,本文提出了一种用于单图像超分辨率的轻量级生成对抗网络,称为CBSRGAN。具体而言,在提出的模型中,融合了三维注意力机制SimAM增加模型对图像细节特征的提取,同时降低了模型的参数。在模型中插入了通道和空间注意力模块CBAM,并且采取了混合损失函数测试该模型。实验结果表明,提出的CBSRGAN显著提高了轻量级SISR方法的感知图像质量。
For the super-resolution enhancement task of a single image,the super-resolution generative adversarial network(SRGAN)technology performs excellently in achieving highly realistic textures.In addition,most advanced SISR methods cannot be applied to resource limited devices due to the large number of parameters.To address these challenges,a lightweight generative adversarial network for single image super-resolution has been proposed,it is called CBSRGAN.Specifically,in the proposed model,integrated the three-dimensional attention mechanism simAM to increase the model's extraction of image detail features,while reducing the parameters of the model.The channel and spatial attention module CBAM was inserted into the model,and a hybrid loss function is adopted to test the model.The experimental results indicate that,the proposed CBSRGAN significantly improves the perceived image quality of lightweight SISR methods.
作者
陈诗怡
CHEN Shiyi(Harbin Normal University,Harbin Heilongjiang 150000)
出处
《软件》
2024年第12期42-44,共3页
Software
