AbstractThis paper proposes a new bags-of-words (BoW)-based algorithm for scene/place recognition. Current scene recognition works that adopt BoW as the framework usually use a single codeword to represent the clusters obtained by k-means. Further, most of them often assign a hard value to a certain codeword to construct the BoW histogram. Using a single codeword to represent each cluster in fact is very preliminary since different clusters usually have different mean and covariance values. This causes using only mean value-based codeword will lose the covariance information and also makes the hard assignment to the codeword become biased. Considering this, this paper proposes an effective BoW-based technique to perform scene recognition. It first uses k-means algorithm to cluster the feature vectors into a certain number of clusters, in addition with an occurrence matrix. Gaussian mixed model (GMM) is then used to model the distribution of each cluster. Each GMM will be used as the new “codeword” of the codebook. Finally we propose to establish a new soft BoW histogram to represent each image through the soft assignment of the image features to each GMM. Support vector machine (SVM) is used to train these BoW histograms. Experimental results on the 15 categories dataset show that the proposed new BoW-based approach is very effective for scene/place recognition.