基于改进生成对抗网络的单图超分辨率重建

SINGLE IMAGE SUPER-RESOLUTION RECONSTRUCTION BASED ON IMPROVED GENERATIVE ADVERSARIAL NETWORK

针对目前通过超分辨率技术重建后图像的质量不高、纹理细节模糊、网络训练不稳定等问题,提出一种基于改进生成对抗网络的单图超分辨率重建算法。该算法以生成对抗网络为基础,采用多尺度卷积层和GELU(Gaussian Error Linear Units)激活函数对生成网络中的残差块进行优化,提高网络泛化能力;利用Wasserstein距离和Huber损失对损失函数进行优化,增强网络训练的稳定性;减少判别网络中的批规范化层,优化网络结构。实验结果表明:在Set5等数据集上,该算法重建后的图像在客观评价指标和主观视觉效果上均优于其他经典算法。

In order to solve the problems of low image quality, blurry texture details and unstable network training after reconstruction by super-resolution, a single image super-resolution reconstruction algorithm based on improved generation adversarial network is proposed. The algorithm was based on the generative adversarial network. Multiscale convolution and GELU activation function were used to optimize the residual blocks in the generate network and to improve the generalization ability of the network. The Wasserstein distance and Huber loss function were used to optimize the loss function to enhance the stability of network training. The batch normalization layer in the discrimination network was reduced to optimize the network structure. The experimental results show that the image reconstructed by this algorithm is superior to those images reconstructed by other classical algorithms in objective evaluation indexes and subjective visual effects on Set5 and other data sets.