基于随机森林的单幅图像超分辨重建算法研究

发布时间：2018-02-08 20:05

本文关键词： 超分辨率重建随机森林决策树增强预测层级学习　出处：《中国地质大学(北京)》2017年硕士论文　论文类型：学位论文

【摘要】：随着信息化时代的到来,图像作为信息的载体因其生动形象直观便捷的优点越来越凸显其重要性,而更高分辨率的图像往往能提供丰富的细节,具备更高的应用价值。现实中,数字成像系统获取的图像分辨率较低。在这种背景下,图像超分辨率重建的概念应运而生,它的目标是从模糊退化的低分辨率图像重建出清晰的高分辨率图像。目前超分辨率重建算法已被广泛地应用于安防、医学、视频和遥感等领域。本文首先探讨图像超分辨率重建的相关理论和技术基础,以及国内外在该领域的研究成果,重点叙述了基于学习的超分辨率重建算法。本文的主要工作分为以下4点:(1)基于随机森林的图像超分辨率重建算法:在训练阶段,采用外部训练数据训练含有多棵决策树的随机森林,在决策树的叶子节点学习高低分辨率图像块之间的映射关系,即回归模型;图像重建时,寻找输入图像块对应叶子节点的回归模型预测高分辨率图像块,综合多棵树的预测结果重建高分辨率图像。(2)增加训练数据的方法提高重建图像质量:我们通过适当增加开源的图像数据集来扩展训练数据,比较增加训练数据前后重建图像的质量。(3)增强预测的方法提高重建图像质量:通过不同的旋转角度以及对应的翻转变换得到8张低分辨率图像,对每一张图像进行超分辨率重建,然后逆转图像的变换,平均8张图像的重建结果作为最后的高分辨率图像。(4)层级学习的方法提高重建图像质量:随机森林中决策树之间不再相互独立,而是一个层级的关系。每一层决策树预测的高分辨图像块会进入下一层决策树继续预测,层级推动预测的高分辨率图像块接近真实的高分辨率图像块。大量实验结果表明本文方法在图像重建质量和重建效率上都有较好的表现。
[Abstract]:With the arrival of the information age, the image as the carrier of information has become more and more important because of its vivid image, intuitive and convenient, and the higher resolution image can often provide rich details, and has higher application value. The image resolution obtained by digital imaging system is relatively low. In this context, the concept of super-resolution image reconstruction emerges as the times require. Its goal is to reconstruct clear high-resolution images from degraded low-resolution images. At present, super-resolution reconstruction algorithms have been widely used in security, medicine, In this paper, we first discuss the theoretical and technical basis of super-resolution image reconstruction, as well as the domestic and foreign research results in this field. The main work of this paper is divided into the following 4 points: 1) Image super-resolution reconstruction algorithm based on random forest: in the training stage, Using the external training data to train the random forest with multiple decision trees, learning the mapping relationship between high and low resolution image blocks at the leaf node of the decision tree, that is, regression model, when the image is reconstructed, The regression model of finding the input image block corresponding to the leaf node is used to predict the high resolution image block. Combining the prediction results of multiple trees to reconstruct high-resolution images. 2) improving the quality of reconstructed images by adding training data: we extend the training data by appropriately adding open source image data sets. Comparing the quality of reconstructed images before and after the increase of training data to improve the quality of reconstructed images: eight low-resolution images were obtained by different rotation angles and corresponding flipping transformations. The super-resolution reconstruction of each image, and then the reversal of the transformation of the image, The reconstruction results of an average of 8 images are used as the final high-resolution image. 4) level learning method to improve the quality of reconstructed images: decision trees in random forests are no longer independent of each other. It's a hierarchical relationship. Each layer of decision tree predicts a high-resolution image block that goes into the next layer of the decision tree and continues to predict. A large number of experimental results show that the proposed method performs well in image reconstruction quality and efficiency.
【学位授予单位】：中国地质大学(北京)
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】