基于深层卷积网络的单幅图像超分辨率重建模型

Single Image Super-Resolution Restoration Model Using Deep Convolutional Networks

关键词:  超分辨率  深度学习  感知损失函数  卷积神经网络

Key words:  super-resolution  deep learning  perceptual losses  convolutional neural networks

摘要:

[目的]针对Mean squared error (MSE)作为损失函数在人眼感知方面存在局限性,以及基于卷积神经网络的图像超分辨率(Super-resolution,SR)算法生成的图像存在参数较多、计算量较大、训练时间较长、纹理模糊等问题,设计基于深层卷积神经网络的单幅图像超分辨率重建模型。[方法]使用ImageNet预先训练的大型卷积神经网络Visual geometry group (VGG)模型提取图像特征,利用该特征设计视觉感知损失函数进行训练学习,引入亚像素卷积层(Sub-pixel convolution)替换上采样层,缓解生成图像的棋盘效应。[结果]设计的模型对放大两倍的图像进行超分辨率修复,与其他4种超分辨率重建模型的Peak signal to noise ratio (PSNR)值接近,且生成图像的视觉效果更加清晰逼真,细节更加细腻。[结论]该模型可以实现输入不同大小的低分辨率图像而不必多次训练学习不同比例的放大模型,可以实现对不同放大倍数图像的训练和预测,在保持一定PSNR正确率的前提下,放大后的超分辨率图像能够恢复更多纹理细节和更佳视觉效果。

Abstract:

[Objective] On the limitations of human perception in MSE (Mean square error) as one of loss functions, and flaws remained in images generated by the super-resolution algorithm based on the convolution neural network, such as excessive parameters, large calculated amount, long training time and fuzzy texture and so on. This research aimed at designing a single image super-resolution reconstruction model based on a deep convolution of the neural network.[Methods] The deep convolution neural network optimized perceptual loss functions based on high-level features extracted from the pre-trained networks Visual Geometry Group (VGG), and a sub-pixel convolution layer was used to replace the upscale layer and effectively relieve to generate images' checkerboard artifacts.[Results] Our proposed methods doubled the magnification of the image super-resolution repair, which was close to the other four SRCNN SR algorithms and set a better photorealistic image SR.[Conclusion] Our proposed method performs a better visual improvement in our results which are easily noticeable.