Scene Classification with the Discriminative Representation

Scene classification is an extremely challenging task owing to the complexity of the scene content. In this paper, a novel method is designed to harvest the discriminative representation for the scene classification. The proposed model simultaneously takes both discriminative patches and entire scene image into consideration. First, the discriminative patches are extracted from the raw scene image by an iterative algorithm. Then, the convolutional neural network is employed to capture the high-level semantic features from discriminative patches and entire image, respectively. Next, the modified VLAD is exploited to aggregate the features of discriminative patches into local feature. Finally, the global feature generated from the entire image is concatenated with the local feature to be the discriminative representation for classification. The proposed outperforms several state-of-the-art methods on the MIT 67 Indoor and Scene 15 database.