Shape extraction: A comparative study between neural network-based and conventional techniques

Abstract

Extraction of the skeletal shape of an elongated object is often required in object recognition and classification problems. Various techniques have so far been developed for this purpose. A comprehensive comparative study is carried out here between neural network-based and conventional techniques. The main problems with the conventional methods are noise sensitivity and rotation dependency. Most of the existing algorithms are sensitive to boundary noise and interior noise. Also, they are mostly rotation dependent, particularly if the angle of rotation is not a multiple of 90°. On the other hand, the neural network based technique discussed here is found to be highly robust in terms of boundary noise as well as interior noise. The neural method produces satisfactory results even for a very low (close to 1) Signal to Noise Ratio (SNR). The algorithm is also found to be efficient in terms of invariance under arbitrary rotations and data reduction. Moreover, unlike the conventional algorithms, it is grid independent. Finally, the neural technique is easily extendible to dot patterns and grey-level patterns also.