Invariant and equivariant bilinear operations for image feature detection

This paper presents four new feature extraction operators (based on gradient filters), which can be used to extract various edge and corner properties of images. Each of these operators consists of a bilinear combination of 1^st and 2^nd order derivatives of the luminance image, and is either invariant or equivariant to rotations of the image. A brief outline of these invariance and equivariance properties is given. The computational requirements of calculating the operators are stated, and some results of their application on clean and noisy images are presented. In these results, an image pyramid is used to show how a multiresolution approach can give robust estimates in the presence of noise