基于稀疏编码的自然图像特征提取及去噪

主要讨论稀疏编码在自然图像统计特性中的应用，利用稀疏编码实现图像的特征提取以及消除图像中的高斯噪声。丈中利用双梯度算法对自然图像的基向量进行迭代学习。实验表明，提取的基向量在时域和频域上都有方向性和局部性。与小波收缩法相比，稀疏编码法提取的特征要优于小波法提取的特征。对特征提取的实际应用，就是利用稀疏编码收缩法对图像消噪，并通过仿真实验证明稀疏编码收缩法去噪效果要优于任何低通滤波方法。

Image Feature Extraction and Denoising Based on Sparse Coding

The application of sparse coding in the natural images statistics was discussed. Sparse coding is applied to image feature extraction and image denoising. The basis vectors of the natural images were obtained by using fast conjugate gradient algorithm. The experiments show that the basis vectors are localized and oriented in space and in frequency. Compared with wavelet shrinkage method, the features obtained based on sparse coding have important benefit over wavelet ones. As an application of such a feature extraction scheme, the sparse coding shrinkage was applied to reduce Gaussian noise. The simulative experiment shows that the result using sparse coding to denoise is better than that of other low-pass filters.