The low resolution problem in face recognition happens in video surveillance application and degrades the recognition rate dramatically. To overcome the low resolution problem, we introduce a novel face recognition method consisting of extracting multiresolution observation vectors, learning local similarity and making final decision based on top J local probabilities. There are two key contributions. One is to extract multiresolution local characteristics, and the other one is to select the top J local similarities automatically. The benefits of our method are to create multiresolution features and to exclude insignificant local features during recognition phase so that our method could achieve high recognition rate with a low resolution face image. The experimental results show that the proposed method reveals better performance for low resolution face recognition.