【摘要】 传统的图像超分辨率重建方法由于其计算局限性,无法对大批量或者模糊因子不同的图像做最优处理,也无法得出高分辨率图像。近年来随着深度学习神经网络越来越多被学者关注和青睐,其中卷积神经网络被成功应用于图像超分辨率重建。但是传统的图像超分辨率卷积神经网络,无论在训练速度,泛化能力,还是生成图像质量等方面仍存在问题。针对上述问题,对图像超分辨率重建的原理进行研究,对SRCNN模型在多种训练通道下的超分辨率效果进行了实验,并提出了基于多层特征提取层的图像超分辨率重建模型,采用新的优化方法,验证了多种包含不同层数体征提取层的卷积神经网络模型。实验证明该方法在一定程度上优于SRCNN方法,能够有效加快网络整体的训练速度。更多还原

【Abstract】 Due to computational limitations, the traditional image super-resolution reconstruction method cannot optimally process images of different sizes or different blur factors, or obtain high-resolution images. As deep learning has been focused by more and more people, the convolutional neural network(CNN) has been applied to the image super-resolution reconstruction successfully in recent years. However, the traditional image super-resolution convolutional neural network still has problems in terms of training speed, generalization ability and image quality. Aiming at the above problems, we study the principles of image super-resolution reconstruction, tests the super-resolution effect of SRCNN model under various training channels, and based on the test results, propose an image super-resolution model based on multi-layer feature extraction layer. The results shows that the proposed method is better than SRCNN to some extent, which can improve the training speed of the whole network.更多还原