A multi-scale refinement corner detection algorithm based on Shi-Harris

Abstract

Corner detectors provide more structural and localization information with less redundancy than interest point detectors, which are very important in image processing tasks. This paper proposes a novel multi-scale refinement corner detection algorithm based on the Harris response function improved by Shi-Tomasi, integrating scale information into every step of corner extraction in the extremum point space. First, a new image information criterion, the spectral product, is proposed by analyzing the relationship between amplitude and frequency after Fourier transform of an image. This criterion combines the number of intensity variations represented by the frequency with the response strength represented by the amplitude. It is used to compute the mathematical relationship between scale and feature quantity, thereby constructing a corner density space. The number of stable corners at each scale is adaptively determined. Secondly, a new multi-scale separable gradient kernel is designed to adapt to scale variations and precisely compute image gradients. The proposed method further refines the sizes of the corner detection window and the non-maximal suppression window using hierarchical scale information. Finally, the criteria on region repeatability evaluation based on the Vggaffine dataset (under affine image transformation, JPEG compression, and viewpoint transformations) and Hpatches dataset (under viewpoint transformations) are used to evaluate the proposed detector against ten state-of-the-art methods. Experimental results demonstrate that the proposed detector outperforms all the tested methods.