Salient region Detection using Weighted Feature Maps based on the Human Visual Attention Model
The weight of this map is calculated mainly on the basis of the characteristics of different feature maps.
Itti and other people put forward a model based on human visual attention, to detect the significant area of an image, divided into color contrast, grayscale, direction of the characteristics to detect the significance of the distinction, the final synthesis of a significant figure. The methods of synthesis are supervised learning, non-linear local competition, non-linear global competition and so on. In this paper, we propose a feature saliency map to determine the weights of these feature map combinations of the final map, and introduce composite saliency indicator to measure feature Maps ' contribution to the final map.
1) Generate color, intensity, orientation feature map, using the method in [7] to calculate the threshold value, binary.
2) using gift wrapping algorithm to calculate convex polygons enclosing significant regions.
Gift Wrapping algorithm: Starting from the leftmost point, if there are multiple leftmost points, select the bottom of these points. Each time you select all points inside and beginning of the line with the x-axis of the minimum angle as the next point of the convex polygon. Repeat this until you return to the beginning. How to choose the least angle? Computes two vector axb, if the result is positive, then a is smaller than the angle of B to the x-axis.
3) Calculate the polygon area with trapezoid method.
4) define saliency density. If the point in the significant point set and its neighbors in the significant point set are small, the saliency density is large.
5) only feature map with convex polygon area less than 80% of image area is reserved. Choose one of the smallest convex polygon areas as reference map.
6) Only the distance from reference map (which can be European or otherwise) is less than feature maps of TT.
7) Calculate the weight of the feature maps left by the WI, as weighted and. The larger the convex polygon area of a feature map multiplied by the saliency density, the larger the map will weigh.
Results test: Due to the subjective nature of the problem, the author looked for 6 people to see 40 pictures, to determine whether Itti and his method of the significant part of the image reflects the most noticeable place. The result is that he is better than Itti.
ACCV2004 Salient Region Detection paper Reading