Abstract: Iris recognition based sparse recognition (SRIR) is very competitive with traditional recognition approaches on effectiveness and robustness. However, the recognition rate will drop dramatically when the available training samples per subject are very limited, and the computational cost is high. To solve this problem, iris recognition is operating collaborative representation on multi-scale patches and combining the recognition outputs of all patches. Instead of recognition the entire iris image directly, the iris image is divided into several non-overlapping patches with the same scale. Considering the fact that patches on different scales could have complementary information for classification, iris images are patched on multi-scale. The different multi-scale patches are recognized separately based collaborative representation which reduces the computational complexity, while the ensemble of multi-scale outputs is achieved by Bayesian fusion. Experimental results on iris databases show that, although both training and testing image per subject might be very limited, the proposed method outperforms the state-of-the-art recognition approaches on effectiveness and computational cost.