Research On Saliency Region Detection Of Images

The ability to automatically find objects of interest in images is very useful in computer vision. In recent years, it is always a hot topic in computer vision. Saliency region detection is widely used in the areas of image compression, image retrieval, re-targeting, and so on. There are two classes of such algorithms:those that find any object of interest with no prior knowledge, independent of task, and those that find specific objects of interest known a priori. The former class of algorithms tries to detect objects in images that standout, by virtue of being different from the rest of the image. The detection is generic in this case as there is no specific object we are trying to locate. The latter class of algorithm often requires training using features extracted from known examples and detects specific known objects of interest. This paper proposes a purely computational method, independent of specific object, to detect the region of interest.Firstly, the paper proposes a pure computational method, combining graph laplacian technology and region-based global contrast visual saliency detection model. We extend the region-based global contrast visual saliency detection model, and use the saliency map produced by this model as the initial input. The region whose value is greater than some threshold is viewed as the most saliency region, the region whose value is lower than some threshold is viewed as the most common region. The saliency detection problem can be casted as linear equation solver problem, combing the constraints and graph laplacian. The proposed method is fast, simply to implement, and produces better results.Then, the paper proposes another improved model based feature selection. In own method, every pixel is represented by a feature vector, but not all features have positive effect on visual saliency detection. Consider a symmetric Markov process running on the affinity graph, we find that maximization of the mixing rate of markov process will result in the maximization of the second smallest eigenvalue of the correspondence graph laplacian. And this states that the graph is easy to separate into two groups. Then build graph laplacian for every feature, the weight for all features can be generated by formulating the maximization of the second smallest eigenvalue as a semi-definite programming problem. The effect of saliency region detection can be improved after apply the feature weights into the model proposed before.The proposed model has three advantages:First, this is unsupervised visual saliency detection technology; second, our model produces the satisfied results based the local and global feature; third, our model produces full resolution saliency map. Experimental results on a public available dataset show the effectiveness of our method compared with several previous method. And the last we discuss the future work.