An Approach Making the Shape Matching Techniques Robust to Noise

Computer vision has been an active research area for the past few decades. It is being used in many fields of industries for example digit recognition and handwriting recognition in banks, industrial parts recognition etc. Shape matching is an important sub-domain of Computer Vision, and relies on effective shape representation techniques. These techniques must be developed in order to effectively estimate similarity between the shapes. Major similarity estimation results correspond to effective shape matching techniques. These shape matching techniques or shape descriptors must be invariant to variances or noise in the shape for accurate shape matching results. But, unfortunately, these descriptors are not invariant to noise like small cracks or slits present in the contour of the image. These cracks changes the shape boundary which changes the shape altogether. In this paper, we propose a noise removal approach that significantly improves the performance of sophisticated shape matching techniques. Existing shape matching techniques focus on complex algorithms without giving much consideration to the noise removing preprocessing techniques. These preprocessing techniques can reasonably improve the accuracy where the shapes are affected by cracks. We present a preprocessing approach to identify and remove small cracks and slits in the shape contour. It has been observed that these cracks or slits introduce changes in a manner which makes classification of shapes difficult. This technique is simple and effective and may be used to preprocess the images before applying sophisticated shape representation techniques to improve their accuracies. The effectiveness of proposed pre-processing approach is verified using publicly available shape datasets.