Gaussian random matrix (GRM) has been widely used to generate linear measurements in compressed sensing (CS) of natural images. However, there actually exist two disadvantages with GRM in practice. One is that GRM has large memory requirement and high computational complexity, which restrict the applications of CS. Another is that the CS measurements randomly obtained by GRM cannot provide sufficient reconstruction performances. In this paper, a Deep neural network based Sparse Measurement Matrix (DSMM) is learned by the proposed convolutional network to reduce the sampling computational complexity and improve the CS reconstruction performance. Two sub networks are included in the proposed network, which are the sampling sub-network and the reconstruction sub-network. In the sampling sub-network, the sparsity and the normalization are both considered by the limitation of the storage and the computational complexity. In order to improve the CS reconstruction performance, a reconstruction sub-network are introduced to help enhance the sampling sub-network. So by the offline iterative training of the proposed end-to-end network, the DSMM is generated for accurate measurement and excellent reconstruction. Experimental results demonstrate that the proposed DSMM outperforms GRM greatly on representative CS reconstruction methods