Feature quantization is an important component in Bag of word model. This paper proposes a novel method called nonnegative sparse locally linear coding (NSLLC) to improve the performance of locally linear coding. The core ides of NSLLC is to use nonnegative sparse representation to select the nearest neighbors in the same subspace and then encode the local feature with respect to the local coordinate consisting of these nearest neighbors. Experimental results have shown NSLLC has outperformed state-of-the-art local feature coding methods and is in favor of image classification problem.