基于卷积神经网络的超分辨率重建

Super-resolution reconstruction based on convolutional neural network

分析超分辨率重建中基于卷积神经网络与基于稀疏表示方法的联系,讨论网络卷积核的作用,以及不同网络参数对重建效果的影响,设计一个权衡重建质量和结构复杂度的超分辨率重建卷积神经网络模型。实现低分辨率图像到高分辨率图像的映射,整个网络训练过程是端到端学习,具有全局优化的特点,避免基于稀疏表示方法的复杂特征提取和数据重组的操作,实验结果表明,该方法重建的图像质量在视觉效果和参数评价指标上有较大提高。

A single image super resolution reconstruction model was designed which balanced the quality of reconstructed image and the network complexity.The relationship of the network and the algorithm based on sparse representation was analyzed,and the roles of network convolution kernel were discussed.The influences on reconstructed image were analyzed using network parameter.The low resolution image was mapped to the high resolution image,and the whole network was trained end-to-end,global optimization features were taken advantaged to avoid the complex feature extraction and data reorganization in the sparse representation algorithm.Experimental results show that the adopted method increases the evaluation parameters and improves the visual quality of the edges and the details significantly.