Estimating adaptive kernels from local image grey value changes

The extraction of image edges is a fundamental task in early computer vision. The successful edge detection depends on the selection of optimal convolution kernels that are appropriate to the local grey value changes. Unlike previous attempts that use a bank of filters, we introduce in this paper a computational method of estimating adaptive kernels from the covariance matrix of local grey value changes. Such an adaptive kernel can be deformed at any scale in an arbitrary direction. Some results on edge detection using adaptive kernels are also presented in this paper