Recently a novel learning algorithm called the broad learning system (BLS) has been established by the theory of pseudo-inverse and compressed sensing technology. As an alternative to deep learning, BLS has gained extensive attention in data analytics. Unlike the existing deep structures based learning tools (such as deep neural networks), it is able to rapidly achieve incremental learning, and does not require tedious retraining to remodel the system. However, given image data modeling tasks, one-dimensional BLS requires a conventional vectorization operation, the dimension of input will be quite large for high resolution image. This operation may result in huge computational complexity, thereby increasing the time of model construction. To resolve this issue, a two-dimensional version of BLS, namely 2D-BLS, is first proposed to fast construct randomized learner models by using matrix inputs. Specifically, the proposed method employs left and right projecting vectors to replace the usual high dimensional input weight. Moreover, the theoretical analyses on the superiority of 2D-BLS against BLS are presented in terms of the algorithm complexity and the difference of random parameter space. Empirical results based on MNIST dataset, NORB 3D object image recognition dataset, ORL dataset, YaleB dataset and Handwritten Digit dataset demonstrate that the proposed 2D-BLS outperforms the original BLS, some extensions of BLS and some 2-D machine learning methods in terms of the efficiency of model construction. This result also shows good potential of 2D-BLS for image data analytics.