基于反向投影的倒N型轻量化网络图像超分辨率重建

Inverted N-Type Lightweight Network Based on Back Projection for Image Super-Resolution Reconstruction

针对基于深度学习的图像超分辨率重建算法通过增加神经网络的深度提高图像超分辨率重建性能,造成网络模型复杂度提高和参数量过大的问题,提出一种基于反向投影的倒N型轻量化网络图像超分辨率重建算法.首先使用初始卷积块提取低分辨率图像的浅层特征;然后通过基于反向投影的倒N型网络的2轮逐步的模型压缩和反倒N型网络的2轮逐步的模型恢复,提取低分辨率图像的深层特征;再将提取到的深层特征和浅层特征通过全局残差学习相结合并被上采样模块放大到期望输出的重建图像尺寸;最后使用重建模块进行超分辨率图像重建.对Set5,Set14,BSDS100和Urban100测试集进行实验的结果表明,所提算法比对比算法拥有更轻量化的网络结构,重建的超分辨率图像不仅有更高的峰值信噪比(PSNR)和结构相似度(SSIM),而且重建的超分辨率图像具有更好的视觉效果.

Aiming at the problem that the super-resolution reconstruction algorithms based on deep learning increase the complexity of the network model and cause a large number of parameters by increasing the depth of neural network to improve the super-resolution reconstruction performance,an inverted N-type lightweight network based on back projection for image super-resolution reconstruction is proposed.First,the initial convolutional block is used to extract the shallow features of low-resolution images.Secondly,the deep features of low-resolution images are extracted through two rounds of gradual model compression based on the inverted N-type network of back projection and two rounds of gradual model restoration of the inverted N-type network.Then,the extracted shallow features and deep features are combined by global residual learning and scaled up to the desired output reconstructed image size by using the upsampling module.Finally,the reconstruction module is used to reconstruct the super-resolution image.The results of experiments on the Set5,Set14,BSDS100 and Urban100 testing sets show that the proposed algorithm has a lighter network structure than other algorithms,and the reconstructed super-resolution images not only have higher peak signal-to-noise ratio(PSNR)and structural similarity(SSIM),but also have better visual effects.