Subpixel spot localization using multiscale anisotropic Gaussian tensor

Abstract

Spot localization is a fundamental but crucial issue in many optical applications. However, the existing approaches either have limited localization accuracy for overlapping elliptic spots or are noise-sensitive. To overcome these limitations, a novel subpixel spot localization method based on the multiscale anisotropic Gaussian tensor is proposed. Firstly, an anisotropic Gaussian spot model is developed to better describe spots under perspective projection. Then, the anisotropic scale-space tensor representation is generated to separate overlapping spots and determine the initial centers. Afterwards, the scale-adapted anisotropic Gaussian kernel and its second-order partial derivates are constructed to obtain subpixel centers. The simulation results show that the highest score of 1.00 is achieved on overlapping elliptic spot images. When estimating homography matrix, the proposed method yields the reprojection error of 0.06 pixels and is more robust to noise. Thus, the proposed method achieves robust subpixel localization for overlapping elliptic spots.