Scene Classification Based On Saliency Detection And TMBP

Scene classification is called scene perception and scene recognition. It is achallenging research problem in the field of computer vision and cognitive science. Sceneclassification annotates automatically images based on a group of given semantic labels,which helps to provide effective contextual information on the higher level for imageunderstanding task such as object recognition. In recent years, the physiologicalexperiment showed that the biological visual systems rely on selective attention, visualmemory, quickly understand scene content, namely the image significance is closelyrelated with scene semantic understanding. Therefore, the combination of saliencydetection and scene classification is of great significance.In this paper, making the static image as the research object, we study the salientregion detection method, topic model by belief propagation and TMBP scene classificationmethod based on foreground and background. Our researches are as follows:1) For the problem of false detection when detecting image with multi-objects, wepropose a new multi-objects saliency detection method within the Bayesian framework.First, we get the low level features via Context-Aware saliency detection. Then, we obtainthe middle level cue by Ncut image segmentation which is category label information ofmulti-objects. The prior saliency map is computed with respected to both low and middlelevel cues. Last, we use a Bayesian formula to calculate the posterior saliency map. Theexperimental result shows that, our method can better solve the problem of false detectionof multi-object with higher detection precision.2) The the existing image scene classification methods based on the LDA model haveproblems such as slow calculating speed and high complexity. We use topic model methodwhich is based on the belief propagation inference. The topic model is in the framework ofthe LDA model, but it is different from traditional method of VB and GS inference. It uses the belief propagation to inference parameters and improves the speed of inference.Experimental results show that the scene classification method based on topic model bybelief propagation is faster than the traditional VB-LDA and GS-LDA, and computationalcomplexity is also lower.3) For the problem of low classification accuracy in scene classification, we propose ascene classification method which uses topic model on both foreground and background ofthe images. In this method, we first detect the salient areas as the foreground and the rest asbackground. Then we model the foreground and background areas respectively, so as toimplement scene classification. The experimental results show that the TMBP sceneclassification method based on foreground and background is higher than the traditionalmethods.