An integration of top-down and bottom-up visual attention for categorization of natural scene images

An integrated technology of top-down and bottom-up visual attention used to the solution of the feature selection and saliency detection problems in object extraction and categorization of natural scene images is proposed. A decision criterion based on the top-down, goal-driven component is introduced to select the features of desired detection object which best distinguish the object from the other parts of scenes in image. The bottom-up, image-driven component of visual attention can be tuned by the learnt knowledge according to the optimal features to optimize the saliency detection. The saliency detection of the interest objects is performed through each of the images by the optimized bottom-up component. The parameter value of category confidence is computed to conduct the categorization of images. The experiment on a set of natural images is carried out to test the adaptability of saliency discrimination and accuracy of image categorization provided in this paper. The experimental evidence shows that the integrated technology introduced in this paper has the capability of capture the intrinsic information with respect to image categorization.