Abstract: Feature extraction is always the key point in image processing as an important step in pattern recognition. Recently, the theory of deep learning has drawn more and more attention from scholars as a new deep feature extraction model. In this paper, a new algorithm of sparse feature extraction will be proposed based on multi-layered deep metric subspace learning. This algorithm can hierarchically map images through metric matrix. It also can maximize interclass variations. After the mapping, this algorithm guarantees the sparsity of extracted result by the sparse iteration. First of all, it is necessary to build the function for image distance measurement. After that, the optimal metric matrix should be calculated by maximizing inter-class variations. At the same time, feature mapping results should be iterated through L₁ Norm Sparse to in prove the noise robustness. Then, the unit of basic characteristics will be reformed according to the principle of deep subspace. In the second layer, the same operations will happen as well. In the end, sparse feature extraction model will be accomplished based on multi-layered deep metric subspace learning. Comparing with the existing subspace model, this paper introduces the mechanism of metric self-learning in the process of feature mapping. Meanwhile, every feature layer will be added visually pleasing sparse constraint to generate the result of feature extracted. The experimental results on face database of FERET、AR、Yale and target database of MNIST、CIFAR-10 show that this feature extraction model can achieve high recognition rate, robustness for illumination, expression and pose. Additionally, the introduced algorithm has more clear structure and faster convergent rate than deep learning theory of convolutional neural networks or the others.