深度稀疏编码模型及应用

发布时间：2018-02-26 06:16

本文关键词： 稀疏编码降噪可预测稀疏分解物体识别　出处：《上海交通大学》2014年硕士论文　论文类型：学位论文

【摘要】：最近这几年稀疏编码(Sparse Coding)在很多领域取得了巨大的成功，包括数据挖掘、机器学习和计算机视觉等。稀疏编码提供了一种学习出一组超完备基函数的无监督算法，，从而能够通过线性组合基函数的一小部分子集来重构原有的信号。当前大多数稀疏编码算法的一个缺点是需要进行一些迭代最小化计算来为测试点推断出稀疏表示(Sparse Representation)，这说明这些算法很难进行新样本点的扩展。通过额外训练一个把输入非线性地映射为稀疏表示的回归变量，可预测稀疏分解(Predictive Sparse Decomposition)可以很自然地应用于新样本点的扩展。因此可预测稀疏分解成为了非常流行的一种稀疏编码算法。然而，当训练数据太少而不足以捕获样本中的变化时，可预测稀疏分解可能在实际的应用中无法取得令人满意的效果。这篇论文提出了一个新的模型——降噪可预测稀疏分解(Denoising Predictive Sparse Decomposition)算法，来获得更鲁棒的稀疏编码。针对真实的图像物体识别任务的实验结果表明，降噪可预测稀疏分解算法在实际应用中的效果可以大大超越可预测稀疏分解算法。此外，本文还提出了深度降噪可预测稀疏分解算法以及深度多路径降噪可预测稀疏分解算法，验证了降噪可预测稀疏分解算法在深度模型中能大大提升识别效果。
[Abstract]:In recent years, sparse coding has achieved great success in many fields, including data mining, machine learning and computer vision. Sparse coding provides an unsupervised algorithm for learning a set of supercomplete basis functions. Thus it is possible to reconstruct the original signal by a small subset of linear combinatorial basis functions. One disadvantage of most current sparse coding algorithms is that some iterative minimization calculations are required to infer sparse tables for test points. Sparse representation, which shows that these algorithms are difficult to extend new sample points by training an additional regression variable that maps input nonlinearly to sparse representation. Predictive Sparse decomposition) can be applied naturally to the expansion of new sample points. Thus predictable sparse decomposition has become a very popular sparse coding algorithm. However, when the training data is too little to capture changes in the sample, In this paper, a new model-Denoising Predictive Sparse decomposition algorithm is proposed. Experimental results for real image object recognition show that the denoising predictive sparse decomposition algorithm can outperform the predictive sparse decomposition algorithm in practical applications. In this paper, the predictive sparse decomposition algorithm for depth noise reduction and the predictive sparse decomposition algorithm for deep multipath denoising are also proposed, which verify that the denoising predictive sparse decomposition algorithm can greatly improve the recognition effect in depth model.
【学位授予单位】：上海交通大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN911.2

【共引文献】