Enhancing Corner Detection: Leveraging 3 × 3 Structure Tensor Combined with Hourglass Filter

Baptiste Magnier, Guillaume Demoor

Abstract: In this study, an improvement of structure tensors for corner detection in digital images is presented, specifically focusing on comparing 2 × 2 and 3×3 tensors. Contextually, corner extraction serves as a crucial aspect of this analysis. While the 2 ×2 tensor primarily focuses on orthogonal gradients, the 3 ×3 tensor provides a more comprehensive analysis, encompassing diagonal second derivatives as well. Regarding the second-order 3 ×3 tensor, an eigenvalue-based strategy is developed to extract corner-like features in the image. Additionally, a novel approach is introduced by incorporating the hourglass filter, to explore its potential for corner detection. Remarkably, this technique excels at capturing intricate local features, while the hourglass filter provides additional insights into the global context of the image. Specifically, the hourglass filter allows for the detection of closely located corners without merging them together, thereby preserving their individual characteristics and enhancing detection accuracy. The robustness and reliability of the method are demonstrated through evaluation on a dataset where 60 images were subjected to contrast changes and Gaussian noise alterations. Finally, the visual results highlight the effectiveness of the proposed approach with accurately detected corners.

External IDs:dblp:conf/eusipco/MagnierD24