关键词: 正态逆高斯（NIG）密度模型, 稳健主成分分析, 非负稀疏编码, 非负矩阵分解, 特征提取, 图像去噪

Abstract: This paper proposes a modified Non-Negative Sparse Coding（NNSC） neural network model based on the Normal Inverse Gaussian（NIG） density and robust Principal Component Analysis（PCA）.This model implements in fact a two-phase learning.Using this model，the V1 receptive fields of vision-perceptional system can be modeled successfully.Compared with Hoyer’s NNSC model，the NIG-based NNSC network behaves stronger capacity of adapting to the statistical properties of natural data.Otherwise，exploiting the similar wavelet shrinkage principle，this model can denoise efficiently the additive Gaussian noise in an image.The simulations of coding on natural images demonstrate this model’s plausibility in neuroscience view and feasibility in practical computation.

Key words: Normal Inverse Gaussian（NIG） density model, robust Principal Component Analysis（PCA）, non-negative sparse coding, non-negative matrix factorization, feature extraction, image denoising