DFOB: Detecting and describing features by octagon filter bank for fast image matching

https://doi.org/10.1016/j.image.2015.12.001Get rights and content

Highlights

  • A novel robust and efficient binary descriptor.

  • An efficient and accurate method for feature point orientation computation.

  • A variant of the CenSurE detector for feature point detection.

  • The detector and descriptor are created from the same filter bank.

Abstract

Feature correspondence is vital in image processing and computer vision. To find corresponding pairs efficiently, in this paper it is proposed that feature detector and descriptor are constructed from the same octagon filter bank (DFOB). The DFOB method is a novel method for the detection, orientation computation, and description of feature points, and is very efficient as computationally implemented by integral images. The matching capability of DFOB is close to the prevalent methods such as SIFT and SURF, because they all detect blob-like image structures as interest features and describe these features using histogram of oriented gradients. Experimental results on benchmark datasets demonstrate that the matching performance of DFOB is comparable with the SIFT and SURF algorithms, while the computational cost is much lower, especially the proposed descriptor is about 50 times faster than SURF descriptor.

Introduction

Image matching is fundamental for many computer vision applications such as object recognition, 3D structure reconstruction, image stitching, visual mapping, and target tracking. Extracting stable and repeatable features and encoding these features into robust descriptors with high discriminability are two key steps in image matching. During the last decade, researchers were racing to find faster and better approaches to address this issue. Among them, SIFT [1] and SURF [2] are two of the most famous approaches due to their high quality with respect to the detector repeatability and the descriptor discriminability under a variety of different image geometric and photometric deformations. However, the two methods are memory consuming and computationally expensive, which are not suitable for real-time tasks or applications on mobile devices with limited computational capability. Recently, high computational speed and low memory consumption are achieved by using the combination of the variant of the FAST [3] detector and variant of the BRIEF [4] descriptor, such as ORB [5], BRISK [6] and FREAK [7]. Compared to these methods, the proposed method is more efficient, this is because the descriptor and the detector in proposed method are created from the same filter bank and this makes a dynamic programming strategy applicable to accelerating the computation. The main contribution of our work are highlighted in the following three trickles:

  • A novel binary descriptor based on histogram of oriented gradients, which is about 50 times faster than SURF descriptor with about the same robustness and discriminability for matching.

  • A stable method for feature point orientation computation, which is efficient and accurate.

  • A variant of the CenSurE detector for feature point detection, which allows the detection to run faster and cover more scales than the original.

The DFOB method has been tested on standard benchmarks and the experimental results demonstrate that the matching performance of DFOB is close to SIFT and SURF which are famous for the high quality in matching, and the time efficiency is higher than BRISK and ORB which are known for their high computation speed. Another advantage of our method is that both the detector and the descriptor are constructed by octagon filters without pre-processing steps such as Gaussian-smoothing or data training. So, the implementation of our method is very simple, and has quite good efficiency.

Section snippets

Related works

In this section we briefly introduce several popular feature point detectors and descriptors related to DFOB method.

DFOB: the method

The DFOB method mainly consists of 3 parts: feature point detection, feature point orientation computation and feature point description. Each method in the 3 parts can be used in combination with the methods in other popular image matching algorithms, such as SIFT and SURF. It׳s worth noting that different to other methods building the detector and descriptor separately, DFOB method builds them together. That is, the descriptor is built using the intermediate result produced in the feature

Experimental results and discussion

We compared our method to SIFT, SURF, FAST, ORB, BRISK and FREAK in respect to detector repeatability, the accuracy of the feature point orientation or the descriptor recognition rate. The SIFT code implemented by Rob Hess [29] is used in our experiment. For the SURF algorithm, we use the original implementation released by its authors. The other methods in our experiment are implemented in OpenCV 2.4.10. The methods and datasets proposed by Mikolajczyk and Schmid [28], [20] are used for the

Conclusions

In this paper we have introduced a novel algorithm, named DFOB, for fast image matching. The best advantage of DFOB lies on its efficiency, and meanwhile, its matching performance is very close to the best accurate methods. Without using any prior knowledge of the scene and camera poses and without preprocessing procedures such as Gaussian smoothing and data training, the DFOB method is very convenient for use. Due to the good properties, DFOB could be a favourable choice for many computer

Acknowledgements

Thank the Editor and Reviewers for time and effort spent handling this paper. This work was supported in part by the National Natural Science Foundation of China under Grant 61571313, Grant 61173182 and Grant 61411130133 and in part by the Sichuan Province under Grant 2014HH0048 and Grant 2014HH0025.

References (30)

  • H. Bay et al.

    Speeded-up robust features (SURF)

    Comput. Vis. Image Underst.

    (2008)
  • S.-C. Pei et al.

    Design of fir bilevel Laplacian-of-Gaussian filter

    Signal Process.

    (2002)
  • J. Matas et al.

    Robust wide-baseline stereo from maximally stable extremal regions

    Image Vis. Comput.

    (2004)
  • P.L. Rosin

    Measuring corner properties

    Comput. Vis. Image Underst.

    (1999)
  • D.G. Lowe

    Distinctive image features from scale-invariant keypoints

    Int. J. Comput. Vis.

    (2004)
  • E. Rosten et al.

    Faster and bettera machine learning approach to corner detection

    IEEE Trans. Pattern Anal. Mach. Intell.

    (2010)
  • M. Calonder, V. Lepetit, C. Strecha, P. Fua, BRIEF: binary robust independent elementary features, In: Computer...
  • E. Rublee, V. Rabaud, K. Konolige, G. Bradski, ORB: an efficient alternative to SIFT or SURF, In: Computer Vision...
  • S. Leutenegger, M. Chli, R.Y. Siegwart, BRISK: binary robust invariant scalable keypoints, In: 2011 IEEE International...
  • A. Alahi, R. Ortiz, P. Vandergheynst, FREAK: fast retina keypoint, In: 2012 IEEE Conference on Computer Vision and...
  • H.P. Moravec, Obstacle avoidance and navigation in the real world by a seeing robot rover, Technical Report, DTIC...
  • C. Harris, M. Stephens, A combined corner and edge detector, In: Alvey Vision Conference, vol. 15, Manchester, UK,...
  • K. Mikolajczyk, C. Schmid, An affine invariant interest point detector, In: Computer Vision ECCV 2002, Springer,...
  • K. Mikolajczyk et al.

    Scale & affine invariant interest point detectors

    Int. J. Comput. Vis.

    (2004)
  • E. Mair, G.D. Hager, D. Burschka, M. Suppa, G. Hirzinger, Adaptive and generic corner detection based on the...
  • Cited by (6)

    View full text