Scene Image Categorization Based on Content Correlation

Scene image categorization is a basic problem in the field of computer vision. A content correlation based scene image categorization method is proposed in this paper. First of all, dense local features are extracted from images. The local features are quantized to form visual words, and images are represented by the “bag-of-visual words” vector. Then a logistic-normal distribution-based generative model is used to learn themes in the training set, and themes distribution on each image in the training set. Finally, an SVM based discriminative model is used to train the multi-classifier. The proposed approach has the following advantages. Firstly, the approach uses logistic normal distribution as the prior distribution of themes. The correlation of themes is induced by the covariance matrix of logistic normal distribution, which makes the theme distribution of subjects more accurate. Secondly, manually tagging image content is not required in learning process, so as to avoid the heavy human labor and subjective uncertainty introduced in the process of labeling. A new local descriptor is proposed in this paper, which combines the gradient and color information of local area. Experimental results on natural scene dataset and manmade scene dataset show that the proposed scene image categorization method achieves better results than traditional methods.