|
Contour detection algorithms which operate directly on an image are greatly affected by small magnitude transformations to the image. A contour detection method is developed which takes advantage of the Curvature Scale Space representation. This method improves detection performance by increasing the detector invariance to affine transformations including: scaling, rotation, and translation. Invariance to noise is also demonstrated. The addition of a histogram matching criterion further improves performance by sanity-checking the results of the Curvature Scale Space detection. The combination of the two techniques produces a robust detector well-suited for object detection in noisy images and video. Below, two images are displayed. The first is a silhouette of Sherlock Holmes. Next, a pure rotation of 45° is applied to the image. In the pixel domain, these two images are dramatically different.
|
|
Below, the Curvature Scale Space representation of each image is presented. The invariance to an affine rotation is immediately evident. In fact, these representations are identical up to a circular rotation in the CSS domain.
|
