This paper focuses on the relationship between the number of interest points and the accuracy rate in scene classification. Here, we accept the common belief that more interest points can generate higher accuracy. But, few effort have been done in this field. In order to validate this viewpoint, in our paper, extensive experiments based on bag of words method are implemented. In particular, three different SIFT descriptors and five feature selection methods are adopted to change the number of interest points. As innovation point, we propose a novel dense SIFT descriptor named Octave Dense SIFT, which can generate more interest points and higher accuracy, and a new feature selection method called number mutual information (NMI), which has better robustness than other feature selection methods. Experimental results show that the number of interest points can aggressively affect classification accuracy.

This paper presents an efficient yet powerful codebook model, named classified codebook model, to categorize natural scene category. The current codebook model typically resorts to large codebook to obtain higher performance for scene categorization, which severely limits the practical applicability of the model. Our model formulates the codebook model with the theory of vector quantization, and thus uses the famous technique of classified vector quantization for scene-category modeling. The significant feature in our model is that it is beneficial for scene categorization, especially at small codebook size, while saving much computation complexity for quantization. We evaluate the proposed model on a well-known challenging scene dataset: 15 Natural Scenes. The experiments have demonstrated that our model can decrease the computation time for codebook generation. What is more, our model can get better performance for scene categorization, and the gain of performance becomes more pronounced at small codebook size.

We present a novel model, named Category Constraint-Latent Dirichlet Allocation (CC-LDA), to learn and recognize natural scene category. Previous work had to resort to additional classifier after obtaining image topic representation. Our model puts the category information in topic inference, so every category is represented in a different topics simplex and topic size, which is consistent with human cognitive habit. The significant feature in our model is that it can do discrimination without combined additional classifier, during the same time of getting topic representation. We investigate the classification performance with variable scene category tasks. The experiments have demonstrated that our learning model can get better performance with less training data.

Visually saliency detection provides an alternative methodology to image description in many applications such as adaptive content delivery and image retrieval. One of the main aims of visual attention in computer vision is to detect and segment the salient regions in an image. In this paper, we employ matrix decomposition to detect salient object in nature images. To efficiently eliminate high contrast noise regions in the background, we integrate global context information into saliency detection. Therefore, the most salient region can be easily selected as the one which is globally most isolated. The proposed approach intrinsically provides an alternative methodology to model attention with low implementation complexity. Experiments show that our approach achieves much better performance than that from the existing state-of-art methods.

Previous works show that the probabilistic Latent Semantic Analysis (pLSA) model is one of the best generative models for scene categorization and can obtain an acceptable classification accuracy. However, this method uses a certain number of topics to construct the final image representation. In such a way, it restricts the image description to one level of visual detail and cannot generate a higher accuracy rate. In order to solve this problem, we propose a novel generative model, which is referred to as multi-scale multi-level probabilistic Latent Semantic Analysis model (msml-pLSA). This method consists of two parts: multi-scale part, which extracts visual details from the image of diverse resolutions, and multi-level part, which concentrates multiple levels of topic representation to model scene. The msml-pLSA model allows for the description of fine and coarse local image detail in one framework. The proposed method is evaluated on the well-known scene classification dataset with 15 scene categories, and experimental results show that the proposed msml-pLSA model can improve the classification accuracy compared with the typical classification methods.