Object detection on inertia surface by improved watershed transform

This paper presents a new scheme for object detection in a complex background. Firstly a Difference Offset of Gaussian filter is introduced to calculate a feature inertia surface of an image, this feature inertia image preserves certain region ridges in an image, while reducing insignificant details. After skeletonization on the inertia surface, an improved marker extraction for watershed transform is carried out to detect objects, followed by a merging operation based on a criterion suggested according to the measurement of texture similarity. Each located area is finally verified by Nearest Neighboring classifiers trained for different kinds of objects. Detection experiments on face areas and character regions have shown its feasibility.