Robust Filtering-Based Thinning Algorithm for Pattern Recognition

We present a novel thinning algorithm to automatically extract skeletons from images without artefacts. It is well known that the major problem of existing thinning algorithms is the generation of artefacts such as redundant branches and lines due to noise in images. In this approach, we propose to use oriented Gaussian filters to determine principal directions and to classify ridges, valleys and edges. As oriented filters are low-pass filters in the principal directions, they are robust to noise and insignificant extremities. The thinning process of the proposed algorithm is guided by principal directions, thus it can remove edge points and valley points without the interference from noise and insignificant extremities. As a result, extracted skeletons of elongated shapes are smooth and without redundant branches and lines. The thinning algorithm is applied to handwriting recognition, fingerprint recognition and 3D plant analysis, where two 2D side-view images of cereal plants are available to convert 2D skeletons to 3D structures. Experimental results show that the proposed approach is able to handle noise and insignificant extremities and to generate smooth skeletons of objects, and also is used to automatically extract 3D structures of cereal plants.