Contour Fine Tune and Extraction

After finding all the control points in the M sections along the closed boundary, the outline of the desired object can be found similar to the human input initial points in the classical active contour models. However, compared with the human inputs, the identified control points are much more efficient and objective, especially in the situation where large amount of image sequences need to be processed.

Based on these control points, the complete contour can be found by applying the MPA algorithm to every two adjacent control points. To improve the performance in the optimization process, we dynamically frame the path searching range instead of applying it to the whole image. This can avoid the irrelevant sites in the image and hence reduce the computation. Assume a pair of control points are P0 and Pi, in our implementation, the searching range is defined as a square containingp0 andp1, as illustrated in Fig. 8.8, in which P0 and P1 are the middle points of the edges. The shape of searching range is often decided by two factors:

(i) It must guarantee that the minimal path goes through this reduced search region.

(ii) The implementation of this region boundary control must be easy in case excessive computation is involved.

In practical applications, searching range is totally decided by the characteristics of the desired objects. A simple design is a band with certain width along the region boundary found by region segmentation. However, this may

Figure 8.8: Illustration of the dynamic searching range frame.

involve a lot calculation to control its boundary. For simplicity and generality, we chose the square region to limit searching range. Another benefit of this restriction is that it can work as a control of the overall object shape and prevent the occurrence of "wild divergence" distorted by noise.

The procedure of region boundary splitting, control point searching, and curve fine-tuning is illustrated in Fig. 8.9.


- control point

- section separator

Figure 8.9: Illustration of the procedure used to apply the MRF-based active contour model on object boundary tracking. (a) The QHCF segmented region with the boundary divided into six sections. (b) In each section, a control point is searched based on maximum reliability criterion. (c) The final fine-tuned contour is found by linking the curves between each two adjacent control points, which are searched with minimal path approach.

0 0

Post a comment