基于LapSRN的图像超分辨率重建算法

Image Super-resolution Reconstruction Based on LapSRN

针对基于拉普拉斯金字塔结构的图像超分辨率重建算法(LapSRN)的损失函数只采用了robust Charbonnier损失会偏向于提高PSNR不能更好地重建出高频信息,以及采用传统的卷积核在参数相同的情况下拥有较小的感受野对全局信息掌握不够的问题,提出一种结构损失和robust Charbonnier损失结合的新的鲁棒性更强的损失,和用三种膨胀率不同的空洞卷积核替代传统卷积核的一种新的网络结构。经过实验证明,改进的LapSRN在没有增加模型复杂度的同时提高了超分辨率重建图像的重建效果。

Since the existing algorithm LapSRN only using the robust Charbonnier loss and traditional convolution kernel,it can cause model cannot extract more high-frequency information and have relatively small receptive filed. In this paper, we use structural loss combined with robust Charbonnier loss and use dilated convolution kernel to replace the traditional convolution kernel to improve the performance of our model. Experimental results show that our model can extract more high-frequency information and the quality of reconstructed image has been improved, meanwhile it doesn’t increase the complexity of model.