Abstract: To improve the robustness of visual features,we propose a feature mining algorithm of maximally stable extremal regions in scale space.The algorithm detects features in scale space,mines the robust features by analyzing local stability,and then filters features with low discriminability according to entropy;on that basis,by exploiting scale consistency of maximally stable extremal regions in scale space,we propose an efficient retrieval algorithm based on the maximally matched scale,which can improve the retrieval performance of locality sensitive hashing high-dimensional index.Experimental results show that the precision of our algorithm is 10% higher on average than existing methods while the efficiency is also improved.