Task Specific Top-Down Visual Attention Based on Local Pattern Analysis and Histogram Backprojection for Fast Object Localization

In this paper, we present a novel task specific topdown visual attention method via local pattern analysis as well as histogram backprojection for fast localizing a target. First, two features (i.e., the color feature and the local pattern) are extracted for each pixel of training samples. Then, two histograms of each feature are computed for the target and the whole areas of these samples, respectively. The ratio histograms of these two features are generated through dividing the histogram of the target by that of the whole samples. After that, the task relevant color feature saliency map and local pattern saliency map are obtained by backprojecting the ratio histogram into the image based on its color feature and local pattern. The total saliency map is computed by linearly combining these two normalized saliency maps. At last, the object is located where the maximum saliency exists in the total saliency map. Furthermore, we investigate the effect of the context information in feature space. Experimental results show that modification to the ratio histogram of color feature based on the context information in feature space improves the performance and our method performs faster and better than the state-of-the-art top down saliency methods.