问题的提出：

　　　Some algorithms use Hough transform to detect the irises　of both eyes [9,11]. However, since the irises are considerably　small as compared to the face size, it is diJcult to　correctly detect the irises by a direct application of Hough　transform to the whole face region. Chow and Li [9] proposed
an algorithm to solve this problem. The algorithm selects　eye windows from valleys in the intensity image. But,　the performance of this algorithm is sensitive to the threshold　value used to extract valleys.

[9] G. Chow, X. Li, Toward a system for automatic facial feature　detection, Pattern Recognition 26 (1993) 1739–1755.

　　请解释为什么对于阈值特别敏感．

　　Lin and Wu[12] proposed a novel eye detection　algorithm. The algorithm computes a cost for each pixel in　the face region using a generic feature template and selects　pixels with the largest costs as eyes. We veri,ed by our　experiments that it is diJcult to detect eyes by only using the
algorithm of Ref. [12] because pixels with the largest costs　are not always eyes. But, the algorithm is very attractive　because the algorithm can detect the candidates for eyes even　when size and orientation of the image face are unknown　and even when some regions of the face exhibit low contrast.　

　[12] C.H. Lin, J.L. Wu, Automatic facial feature extraction by　genetic algorithms, IEEE Trans. Image Process. 8 (6) (1999)　834–845.

　　请描述Ref【１２】中算法的基本原理，并解释其为什么对于脸部尺寸和方位以及对比度具有鲁棒性？