微型电脑应用

基于生成对抗网络(GAN)的图像超分辨率重建存在训练不稳定、参数冗余、图片纹理细节不够清晰等问题。针对以上问题，提出一种融合注意力的WGAN图像超分辨率重建算法。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，以更好的表达高频特征；为了降低网络参数和提高重建图片的精度，去除SRGAN网络中冗余的BN层，同时将Charboonier损失函数定义为内容损失以减少噪声对图片的影响；引入WGAN的思想，将Wasserstein距离定义为对抗损失，以解决SRGAN模型训练不稳定问题。通过在Set5、Set14、BSD100三个公开数据集进行测试，实验结果表明，该算法相比较SRCNN、VDSR、SRGAN等其他SR模型，算法鲁棒性强，重建后图像的PSNR与SSIM值更高，图片纹理细节、主观视觉效果均有所提升。

Image super-resolution reconstruction based on generative adversarial network(GAN) has the problems of unstable training, redundant parameters, and unclear image texture details. To solve the above problems, an attention-integrated WGAN image super-resolution reconstruction algorithm is proposed. The convolutional block attention module(CBAM) is fused into the residual network of SRGAN to adaptively adjust the weight of each channel to better express the high frequency characteristics. In order to reduce the network parameters and improve the accuracy of reconstructed images, the redundant BN layer in SRGAN network is removed, and the Charboonier loss function is defined as content loss to reduce the influence of noise on images. The idea of WGAN is introduced, and Wasserstein distance is defined as anti-loss to solve the instability problem of SRGAN model. Through the test on Set5, Set14, BSD100 three public datasets, the results show that compared with SRCNN, VDSR, SRGAN and other SR models, the algorithm has strong robustness, the PSNR and SSIM of the reconstructed image are higher, the image texture details and subjective visual effects are improved.