Abstract: Considered the prior sparse structure and the feature subspace locality of a nature image，on the basis of Local Non-negative Matrix Factorization（LNMF），the neural network of LNMF with the sparse constraint，denoted by SC-LNMF，is proposed in this paper.Using three types of nature images to train the SC-LNMF neural network with the different level of image’s dimension，local features of all training images can be extracted successfully.Compared with other feature methods，such as NMF and LNMF，the experimental results prove that the SC-LNMF can simulate the receptivity field of V1 in the primary visual system in brain，and prove that this SC-LNMF method is indeed efficient and application in image local feature extraction.

Key words: sparse constraint, Local Non-negative Matrix Factorization（LNMF）, natural images, feature extraction

摘要： 考虑自然图像的先验稀疏结构及其特征子空间的局部性，在局部非负矩阵分解（LNMF）算法的基础上，提出一种具有稀疏约束的局部非负矩阵分解（SC-LNMF）神经网络算法。使用两类自然属性不同的图像在不同的维数下对SC-LNMF网络进行训练，该方法都能成功地提取出训练图像的局部特征。与NMF、LNMF特征提取方法相比，实验对比结果证明了SC-LNMF算法能够模拟大脑初级视觉系统V1区感受野的特性，进一步证实了该算法在图像局部特征提取中的有效性和实用性。

关键词: 稀疏约束, 局部非负矩阵分解（LNMF）, 自然图像, 特征提取