Interest Point Detection by Limiting Form of Median Log Filter

Abstract

Interest point detection has been widely used in image analysis applications. However, some interest points, including small structures and large angle corners, could not be effectively extracted. This paper proposes a limiting form of median Laplacian of Gaussian (LMLG) filter, which combines the superiority of the traditional Laplacian of Gaussian (LoG) filter and a limiting form of the weighted median LoG filter. A detector is also proposed based on the LMLG filter. The LMLG filter aims to improve the detection of LoG-based methods for interest points, especially small structures and large angle corners. Also, it could detect blobs, edges, and local structures. We conduct the repeatability and discrimination experiments on the Oxford dataset. Moreover, we conduct the recall rate experiment on the DTU dataset. The experiments show that the proposed method achieves comparable performance with state-of-the-art methods. In order to verify the utility of the LMLG detector, we carry out a series of interest point detector-based applications: face recognition, infrared-visible image registration, and image classification. The results demonstrate that the LMLG detector performs better than the nine detectors in face recognition. The LMLG detector outperforms the nine detectors and Hrkać's, Han's and Liu's methods in infrared-visible image registration. Our method also gives a comparable result on image classification. The source code of the proposed LMLG detector is made publicly available at https://github.com/chenjzBUAA/LMLG-detector.

Keywords

• Feature extraction
• image registration
• interest point detection
• Laplacian of Gaussian filter