The Research And Application Of Topic Model And Eye Movement In Scene Classification

Scene classification is a basic problem in image understanding, and its task is to annotate the given scene image with the right semantic label. Two steps are included in the process of scene classification. The first one is to represent a scene with appropriate information, and the second mainly concerns the selection of algorithms and classifiers that are used to achieve classifying. Our work mainly focus on the first step.When extracting the local features with the method of single scale, if the scale is too large, there will be redundant information in the extracted features; and if the scale is small, the spatial information will be lost. So we propose a new method which is based on multi scale, and the difference lies in that the scale will vary with the position of the image region. Furthermore, we also propose to improve the quality of codebook with the method of clustering in every specific scene. The experiment with LDA as the classifier show that the classification accuracy was high, so it verify the effectiveness of our method.In most of the scene classification method, the whole scene is needed to achieve categorization. The features overload could lead to low efficiency. We all know that human’s visual attention is easily attracted by the salient region of the scene. So there exists an relationship between the gaze zone and the saliency. And this kind of high-level cognitive behavior can be reflected by the eye movement. We propose that bringing the eye movement into the stage of extracting features, and get the salient images. In order to validate the effective of this method, they are used as the image set in the implement of classification. The classifiers were the combination of Spatial Pyramid Matching(SPM) and SVM. The results show that classification performance is desirable in some of the scenes, and the classification efficiency promote a lot in the training and testing stages.Considering that some improvement has been made in the model that include the eye movement, we try to make use of the analysis of eye movement to understand the visual perception mechanism from the view of top-down. For this purpose, two experiments about the eye movement were organized. Human visual attention mechanism, which include the location and entropy of the gaze zone, the fixation duration, and eye movement path, are analyzed on the data of eye movement. The experiment results verifies that the gaze zones are informative, and there is a strong correlation between the scene category and the gaze zone. So it provides the argument basis for developing the scene classification model which is more close to the mechanism of human visual perception.