Abstract: In order to solve the problem that single-frame image can be super-resolution reconstructed under different magnification conditions, this paper proposes a good method of super-resolution reconstruction based on Adam optimization for convolutional neural network (CNN). Firstly, ISODATA (iterative selforganizing data analysis) clustering algorithm is used to classify the trained image sets. Then, feature mapping is obtained by feature extraction and nonlinear mapping of input images in the convolution neural network optimized by Adam. Finally, the feature mapping is reconstructed by deconvolution to obtain a multi-scale enlarged reconstruction image in the convolution neural network optimized by Adam. It is verified by experiments that the reconstruction effect of this method under different magnification conditions is better than the traditional algorithm, and it has better performance in visual effect.

Key words: super-resolution reconstruction, convolution neural network (CNN), iterative selforganizing data analysis (ISODATA) clustering algorithm, Adam optimization algorithm

摘要： 为了使单帧图像在不同放大倍数的条件下进行超分辨率重建能得到较好的效果，提出了一种Adam优化的卷积神经网络（convolutional neural network，CNN）超分辨率重建方法。该方法首先使用ISODATA（iterative selforganizing data analysis）聚类算法对训练的图像集进行分类处理，然后在Adam优化的卷积神经网络中对输入图像进行特征提取和非线性映射得到特征映射图，最后在Adam优化的卷积神经网络中对特征映射图进行反卷积重建得到多尺度放大的重建图像。通过实验验证使用该方法在不同放大倍数条件下的重构效果优于传统算法，在视觉效果上有较好的表现。

关键词: 超分辨率重建, 卷积神经网络（CNN）, ISODATA聚类算法, Adam优化算法