基于特定深度内部学习网络提高"不理想"图像分辨率

现有基于深度学习的图像分辨率提升方法中需耗费大量时间训练,且受限于特定的训练数据等问题。本文引入“零射击”超分辨率概念,可以重复利用已有照片、噪声图像、生物数据以及获取过程未知或非理想的其他图像的相关内部信息,以提高其分辨率。在测试阶段训练一个小的图像特定卷积神经网络,仅需对从输入图像本身提取的示例进行训练；然后通过单个图像内部信息再现,进一步利用图像内部相关信息,以增强图像分辨率。实验结果表明,本文算法可以加快训练速度且不需要标准训练集,图像分辨率优于现有基于卷积神经网络的超分辨率方法,以及已有的无监督超分辨率方法。

Improving Image Resolution Based on Deep Internal Learning

Existing image learning enhancement methods based on deep learning require a lot of time training and are limited by specific training data and the like. This article introduces the "zero-shot" super-resolution concept, which can reuse existing photos, noise images, biological data, and related internal information of other images with unknown or non-ideal processes to improve their resolution. In the test phase, a small image-specific convolutional neural network is trained, and only the example extracted from the input image itself is trained; then, the internal information of the image is reproduced by a single image, and the internal information of the image is further utilized to enhance the image resolution. The experimental results show that the proposed algorithm can speed up the training and does not require a standard training set. The image resolution is better than the existing super-resolution method based on convolutional neural network and the existing unsupervised super-resolution method.