CS-FREAK: An Improved Binary Descriptor - Advances in Image and Graphics Technologies - page 130

accelerating the process by introducing integrating graphics into its process of
describing [8]. But dimensions of these descriptors remain high, which are unsatisfac-
tory in terms of real-time practice, therefore the binary descriptors become hotspot
recently. A clear advantage of binary descriptors is that the Hamming distance
(bitwise XOR followed by a bit count) can replace the usual Euclidean distance,
eliminating the common matching strategy such as building a K-d tree. Calonder et al.
put forward the BRIEF[9]which is obtained by comparing the intensity of 512 pairs of
pixels after applying a Gaussian smoothing to reduce the noise sensitivity; Ruble et al.
improved the traditional FAST by adding orientation information and proposed the
Oriented Fast and Rotated BRIEF (ORB)[11],which gets strong robustness to noise
and rotation, which is obtained by comparing the intensity of random pixels pairs;
Leutenegger et al. put forward BRISK descriptor[10]which is invariant to scale and
rotation, their BRISK is obtained by comparing a limited number of points in a specif-
ic sampling pattern; Alahi et al, heuristically proposed FREAK[1]descriptor accord-
ing to human retina system, that a cascade of binary strings is computed by efficiently
comparing image intensities over a retinal sampling pattern.
In this paper, we propose a new method for feature description and a novel match-
ing strategy based on FREAK descriptor. We simplified the 8-layer circle model to a 5-
layer model which shortcut the description time, and added the neighborhood informa-
tion of the fixed sampling points as the vote data to ensure the matching accuracy.

1.1 Two-Step Matching Strategy

FREAK descriptor is a binary bit string descriptor by thresholding the difference be-
tween pairs of receptive fields with their corresponding Gaussian kernel, which is
called a binary test. FREAK takes a 8-layer model consist of 43 sampling points, [12]
pointed out that, compared with BRIEF, conducting the binary test with the utilization
of fixed sampling point improved training efficiency, but fixed sampling pattern may
reject the optimal point collection, Alexandre Alahi also pointed out that FREAK is
inspired by the human visual system and more precisely the retina [1], the focus of the
study in terms of points selected is still important. [12] proposed a binary descriptor
matching algorithm based on hierarchical learning method which combines the advan-
tages of the fixed-point sampling mode and random sampling mode. proposed a fixed
point of first use (3 layer 17 points) for training, and then point to where the circle from
the candidate within the stratified random sampling for training learning model that
combines the advantages of different sampling modes, thereby improving the learning
efficiency. Drawing on the basis of this idea, we proposed a new method for feature
description based on FREAK and a novel matching strategy based on voting.

1.2 Simplified Retina Modal

To utilize the neighborhood information of the key points more fully, we presents a
two-step matching strategy, the first step is to use the simplified binary descriptors for
matching, then the neighborhood information of the fixed sampling points is utilized
to form a vote data, determining the final match result according to the number of