Abstract: Scene recognition is an important and challenging topic in the research filed of high level image understanding. Traditional researches of scene recognition focused on handcrafted features, which result in limited discriminative and generalization ability. In addition, finding regions in a scene with rich information is always very challenging. This paper presents an effective method for scene recognition based on learned features from multi-scale salient regions. The method first finds multi-scale salient regions in a scene and then extracts the features from the regions via transfer learning using convolutional neural networks (ConvNets). Experiments on two popular scene recognition datasets show that our proposed method is effective and has good generalization ability for scene recognition, compared with the benchmarks on both of the datasets.