基于多尺度图像块分类的字典学习算法

Dictionary Learning Algorithm Based on the Classification of Multi-scale Image patches

针对传统字典学习算法未考虑训练数据集流形结构的问题,提出一种基于KD树分类的多尺度字典学习算法。首先在预处理阶段建立图像高斯金字塔,提取不同尺度下各层图像的角点并建立KD树进行分类;以各类角点为中心截取图像块并生成每层图像的训练数据集来完成各个子字典的学习。在字典训练阶段,提出一种基于局部保持投影的多原子更新算法,在保持字典中各类原子集的流形结构的情况下进行原子更新,高效训练出自适应稀疏字典。对测试图像进行压缩感知重构实验,仿真结果表明,该算法在保证图像重建精度的前提下,显著提高字典学习效率。

Aiming at the traditional dictionary learning algorithm which doesn’t consider the manifold structure of the training set,a multi-scale dictionary learning algorithm based on the classification of KD tree is purposed.Firstly,in the pre-processing stage,the image Gauss pyramid is established to extract the corners of each layer of image at different scales. And KD tree is established to classify the corners. Then the image blocks are intercepted and the training data sets of each layer are generated to complete sub-dictionary learning. In the stage of dictionary training,a algorithm of polyatomic updating based on local preserving projection is proposed. And the adaptive sparse dictionary is trained by updating atoms while maintaining the manifold structure of all kinds of atom sets.When the images of compressed sensing reconstruction are completed,the result shows that the proposed algorithm can significantly improve the efficiency of dictionary learning and the accuracy of image reconstruction is ensured.