Research On Objects Classification Method In Using Saliency Detection And Bag Of Words Model

Object classification is an important research branch of computer vision, and also is the foundation of video surveillance system. It has a high research value and significance with widespread practical application in the image processing and video analysis. Due to the internal and external factors such as the background noise, illumination changes, objects occlusion and the object distortion, there are a lot of problems of modeling, training and classification. The most important problem of object classification at the present stage is how to classify objects quickly and accurately under the complex condition.During the past decades, many scholars have achieved great success in the object classification. The bag of words model plays an important role. Many object classification methods based on the bag of words model can effectively offset the scale changes, illumination changes and shape changes. Those methods have obvious advantages on object classification. However, the mixing of the foreground and background still affects the accuracy of object classification. Focus on this defect, this paper presents an object classification method based on the combination of saliency detection and bag of words model to reduce the interference of background noise.First of all, the full pixel visual saliency algorithm and the graph based visual saliency algorithm were employed to calculate a saliency map of the original image. The foreground object extracted from single saliency detection algorithm is incomplete. Due to the computer cannot accurately analysis the saliency map, two different saliency algorithms were employed to calculate saliency map for retaining the integrity foreground objects and eliminating background effectively.Secondly, optimized maximum difference method was used to calculate the saliency map threshold which can be used to calculate the region of interest. Then, the scale invariant feature transform descriptor was extracted from the region of interest to describe the object. Experimental results show these features extracted from the foreground object can effectively eliminate the interference of background noise.Finally, in order to overcome the defects of K-means clustering algorithm, the density peak clustering algorithm was employed to cluster the feature vectors. The transformation of K-nearest neighbor algorithm was employed to redivide the feature vectors which were abandoned by the density peak clustering algorithm in the process of gaining visual words. The experimental results show that accurate visual dictionary is conducive to improve the accuracy of object classification.