A new method to recognize objects by means of multiscale features and
Hopfield neural networks is proposed in this paper. The feature vector
consists of the multiscale wavelet transformed extremal evolution. Th
e evolution contains the information of the contour primitives in a mu
ltiscale manner, which can be used to discriminate dominant points, he
nce a good initial state of the Hopfield network can be obtained. Such
good initiation enables the network to converge more efficiently. A n
ew normalization scheme, wavelet normalization, was developed to make
our method scale invariant and to reduce the distortion resulting from
normalizing the object contours. The Hopfield neural network was empl
oyed as a global processing mechanism for feature matching. The Hopfie
ld network was modified to guarantee unique and more stable matching r
esults. A new matching evaluation scheme, which is computationally eff
icient, was proposed to evaluate the goodness of matching, images of i
ndustrial tools were used to test the performance of the proposed meth
od under noisy, occluded and affine conditions. Experimental results h
ave shown that our method is robust and more efficient than the Mokhta
rian-Mackworth's method. (C) 1998 Elsevier Science B.V. All rights res
erved.