Image representation for content-based scene recognition is considered as one of the most challenging problems in computer vision. Motivated by the successful text information retrieval application through a histogram of word counts for each text document, the promising bag of features (BoF) model (or the bag of visual words model) has been developed in recent years and clearly demonstrated its potential as a powerful image representation technique for visual information retrieval application. The BoF model represents each image as a collection of clustered local features, and the centroid of each cluster is known as a visual word, which is analogous to a text word. Since the contextual information (i.e., the inter-word relationship) among the visual words is omitted in the conventional BoF model, a novel feature weighting method, coined as the FeatureRank (FR) in this thesis, has been proposed to address this issue. Furthermore, a non-parametric clustering scheme has been developed, which does not require any prior knowledge regarding the number of clusters involved. Finally, the FeatureRank-based weighting scheme has been developed and incorporated into the BoF model for improving the performance of scene recognition.