基于生成对抗网络的图像超分辨率改进模型

Improved Image Super Resolution Model Based on Generative Adversarial Network

针对图像去噪处理后出现的细节模糊等问题,提出一种改进生成对抗网络的图像超分辨率模型。以Super Resolution Generative Adversarial Networks(SRGan)模型为基础,为充分提高其特征提取能力和重建能力,使用空间注意力机制改造残差模块,并在生成器网络中使用了空洞卷积层。在对抗损失、感知损失和L1损失函数的作用下,生成器和判别器子网络交替训练,逐步提高各自性能。训练数据集中的低分辨率图像添加了高斯噪声,使其更好地模拟真实场景。使用Set5、Set14和Urban100数据集进行测试。结果表明,超分辨率模型的重建效果较好,SSIM和PSNR两项指标均有明显提升。

To address the problems such as edge blurring after image reconstruction process,an image super resolution model with generative adversarial network is proposed.Based on the Super Resolution Generative Adversarial Networks(SRGan)model,in order to fully improve its feature extraction and reconstruction capabilities,the residual module is modified using the spatial attention mechanism,and dilated convolutional layers are used in the generator network.The generator and discriminator are trained alternately under the effect of the adversarial loss,perceptual loss and L1 loss functions to gradually improve their respective performance.Gaussian noise is added to the low-resolution images in the training dataset to make it better simulate the real scene.Tests are conducted using Set5,Set14 and Urban100 datasets,and the results show that the super-resolution model has a better reconstruction effect,and the two metrics of SSIM and PSNR are significantly improved.