CN113689403B - Feature description system based on azimuth distance between features - Google Patents

Abstract

A feature description system based on inter-feature azimuth distance relates to the field of image feature extraction. All feature points of an image are detected by a feature detector, feature points are screened by a feature screening unit, then the azimuth and the distance between each main feature point and all other secondary feature points are calculated by a feature relation calculating unit, and the relation strength is further calculated. Finally, the feature description generating unit performs feature description, namely a feature description vector. The method adopts the information of the azimuth distance between the features to quantitatively describe the features, can describe the features without using scale space and gray gradient change information, and simultaneously maintains certain differentiation and scale invariance. The method has low computational complexity, is easy to realize in hardware, is beneficial to accelerating applications such as feature matching, image registration and the like, and brings potential to quick and real-time application.

Description

Feature description system based on azimuth distance between features

Technical field

The invention relates to the field of image feature extraction, and in particular to a feature description system for azimuth distances between features.

Background technique

Feature extraction is to detect features such as points and corners in the image, and use appropriate methods to quantitatively describe these features to prepare for subsequent processing such as image registration, image splicing, and image fusion. Therefore feature extraction is a key step for many image applications. Feature extraction is divided into two steps: feature detection and feature description. In terms of computational complexity, feature description has high computational complexity because it needs to make full use of the grayscale change information of the pixels around the detected feature points. To determine the direction, scale and other information. This makes many feature-based image processing limited in real-time applications.

Contents of the invention

In order to solve the problems of high computational complexity and poor real-time performance of feature description in the prior art, the present invention proposes a feature description system based on the azimuth distance between features.

A feature description system based on the azimuth distance between features, which includes a Gaussian smoothing unit, a feature detector, a feature screening unit, a relationship calculation unit between features and a feature descriptor generation unit;

The feature screening unit includes a secondary feature screening unit and a primary feature screening unit;

The feature descriptor generation unit includes a direction estimation unit and a direction intensity histogram statistics unit;

The Gaussian smoothing unit smoothes the input image I(x, y) and then performs feature detection through the feature detector;

The feature detector detects and obtains the feature point set F ^(I) of the image, and at the same time obtains the position of the feature point _fi ^(I) in the image Loc ( _fi ^(I) ) = (xi _, yi ₎ and the The response intensity Mag( _fi ^(I )) of the feature point _fi ^(I) , the feature point set is expressed as:

F ^(I) = {f _i ^(I) |i∈[1,2,...,N _f ]}

In the formula, i represents the ordinal number of the detected feature point _fi ^(I) , and the feature point set of the image includes N _f feature points;

The secondary feature screening unit performs secondary feature point screening on the feature point set F ^(I) detected and obtained by the feature detector; specifically:

The characteristic response intensity Mag _m after the introduction of modulation is expressed by the following formula:

In the formula, M, N represent the width and height of the image; sort all the feature points in descending order of the modulated feature response intensity, select the first half of the feature points as the secondary feature points, and then reassign the sequence numbers, that is, _fi ^(I,SF) ; the secondary feature points constitute the secondary feature point set F ^(I,SF) , which is expressed by the following formula:

F ^(I,SF) ={f _i ^(I,SF) |i∈[1,2,…,N _f /2]}

The secondary feature point set F ^{(I, SF)} is screened using the main feature screening unit to determine the main feature points. The specific process is:

The main feature point set F ^{(I, PF)} of the initialized image is an empty set, that is

For each secondary feature point _fi ^(I,SF) , define a circular domain D( _fi ^(I,SF) ) with a radius R, which is expressed by the following formula:

D(f _i ^(I,SF) )={(x,y)|(xx _i ) ² +(yy _i ) ² <R ² }

If there is no other secondary feature point f _j ^{(I, SF)} in the circular domain that satisfies Mag(f _j ^{(I, SF)} ) > Mag (f _i ^{(I, SF)} ), then the f _i ^{( I,SF)} as the main feature point, that is, f _i ^(I,SF) ∈F ^(I,PF) ;

Reassign the ordinal numbers to all main feature points in the main feature point set F ^{(I, PF)} , that is, obtain the main feature points f _i ^{(I, PF)} ;

Then the main feature point set satisfies the following formula:

The relative orientation and distance between each main feature point f _i ^{(I, PF)} and other secondary feature points f _j ^{(I , SF)} are calculated through the inter-feature relationship calculation unit,

The direction estimation unit will determine the main direction of each main feature point, use the main direction as a reference direction, and map the relative orientations of all secondary feature points and the main feature point clockwise to the range of 0 to 2π;

Among them, the main direction Ori(f _i ^(I,PF) ) is determined according to the relative orientation of the secondary feature point that has the strongest relationship with the main feature point; that is:

in satisfy:

Denote the relative orientation after remapping as Azim _rm (f _j ^(I,SF) |f _i ^(I,PF) );

The direction intensity histogram statistics unit will calculate the main direction Ori( _fi ^(I,PF) ) and the remapped relative orientation Azim _rm (f _j ^(I,SF) | _fi ^(I,PF) ) ) to perform statistics and generate feature description vectors; the specific method is as follows:

For each main feature point, taking this as the center and starting from its main direction, the image is divided into n sector-shaped areas, representing n direction intervals respectively;

Then the sum of the relationship strengths of all secondary feature points relative to the main feature point in each direction interval is calculated and becomes the strength of that direction interval;

After obtaining the intensity in all direction intervals, the feature description vector V(f _j ^{(I,PF )} |F ^(I,SF) ) that constitutes f _i (I,PF) is expressed by the following formula:

V(f _j ^(I,PF) |F ^(I,SF) )=[O ₁ (f _i ^(I,PF) |F ^(I,SF) ),…,O _n (f _i ^(I,PF) |F ^(I,SF) )]

Until the feature description vectors of all main feature points are obtained.

Beneficial effects of the present invention:

In the present invention, the secondary feature screening unit is used to determine the secondary features by combining the response intensity of the detected feature points and their position coordinates in the image to improve the repeatability of the system; the primary feature screening unit will determine the secondary features from the secondary features. The main feature points are determined to improve the system's robustness to changes in viewing angle.

When describing features, the system of the present invention uses information on the azimuth distance between features to quantitatively describe features. This method can describe features without using scale space and gray gradient change information, while maintaining a certain degree of distinction and scale. Immutability. This method has low computational complexity and is easy to implement in hardware, which is beneficial to accelerating applications such as feature matching and image registration, and brings potential for fast and real-time applications.

Description of drawings

Figure 1 is a functional block diagram of the feature description system based on the azimuth distance between features according to the present invention;

Figure 2 is a schematic diagram for determining the main feature points;

Figure 3 is a schematic diagram for calculating the orientation and distance between features;

Figure 4 is a histogram of feature relative orientation-relationship strength;

Figure 5 is a histogram of remapping orientation-relationship strength between features;

Figure 6 shows the azimuth interval histogram.

Detailed ways

This embodiment will be described with reference to Figures 1 to 6, which is a feature description system based on the azimuth distance between features. The system includes a Gaussian smoothing unit, a feature detector, a feature screening unit, a relationship calculation unit between features, and a feature descriptor generation unit; The feature screening unit includes a secondary feature screening unit and a primary feature screening unit, and the feature descriptor generation unit includes a direction estimation unit and a direction intensity histogram statistics unit;

First, the image I(x,y) is obtained from an external device (such as a camera, CMOS detector, etc.). The Gaussian smoothing unit smoothes the acquired image to reduce the uncertainty of noise on subsequent feature detection and improve subsequent feature screening and Repeatability in characterization and obtaining smoothed images.

The feature detector performs feature detection on the smoothed image, and simultaneously obtains the position of the feature point in the image and its response intensity. The method used for feature detection can be arbitrary, such as FAST, SIFT, SURF, etc. However, in order to improve computing performance, FAST is generally used. After obtaining all the features, a feature set in this image I(x,y) is obtained, including all feature points f _i ^(I) and their position coordinates Loc(f _i ^(I) ) = (x _i , y _i ) and its response intensity Mag(f _i ^(I) ); after detecting the features, a feature set is formed, that is

F ^(I) ={f _i ^(I) |i∈[1,2,...,N _f ]} (1)

In the above formula, i represents the ordinal number of the detected feature points, that is, the feature point set of the image contains N _f feature points.

In order to improve the repeatability of the system, the secondary feature point screening unit screens these detected feature points based on their response intensity and image coordinates. Since the feature points close to the center of the image can be described by the azimuth distance information of other feature points in each direction, while the feature points close to the edges or corners can only be described by other feature points in the half-azimuth or quarter-azimuth, Therefore, feature points closer to the center can be described more accurately. Therefore, the response intensity of features close to the center of the image is given greater weight, while the response intensity of features closer to the edges or corners of the image is given less weight. The characteristic response intensity Mag _m after the modulation is introduced, that is

In the above formula, M and N represent the width and height of the image. Sort all the feature points in descending order of the modulated feature response intensity, select the first half of the feature points as secondary feature points, and then redistribute the sequence numbers, that is, f _i ^{(I, SF)} . These secondary feature points form the secondary feature points. Feature point set F ^(I,SF) :

F ^(I,SF) ={f _i ^(I,SF) |i∈[1,2,…,N _f /2]} (3)

Since the feature detector detects features, the intensity of the returned feature response will change with the disparity of the image, which will affect the stability of feature description. In order to alleviate this problem, the main feature screening unit further determines the main feature points from the secondary feature points. The main feature point screening unit will further follow the following steps to determine the main feature points:

a) Initialize the main feature point set of the image to an empty set, that is

b) For each secondary feature point _fi ^(I,SF) , define a circular domain D( _fi ^(I,SF) ) with a radius of R:

D(f _i ^(I,SF) )={(x,y)|(xx _i ) ² +(yy _i ) ² <R ² } (1)

In the above formula, R is taken as 1/20·min(M,N);

c) If there are no other secondary feature points f _j ^(I,SF) in the circle domain such that Mag(f _j ^(I,SF) )>Mag(f _i ^(I,SF) ), then f _i ^{(I,SF )} is the main feature point, that is, _fi ^(I,SF) ∈F ^(I,PF) .

d) Reassign ordinal numbers to all main feature points in F ^{(I, PF)} , that is, get f _i ^{(I, PF)} (regardless of how to assign ordinal numbers).

As shown in Figure 2, assume that there are 14 secondary feature points A ~ N in an image. The numbers next to the labels represent the response loudness after modulation. Given the radius of the circle as shown in the figure, according to Equation (3) It can be determined that A~C, F, J, K, M, and N are the main feature points;

It can be seen that these feature point sets satisfy

After all main feature points are determined, the inter-feature relationship calculation unit will calculate the relative orientation and distance between each main feature point and all other secondary feature points: assuming there is a main feature point f _i ^{(I, PF)} , and another secondary feature point f _j ^(I,SF) , then the orientation between f _j ^(I,SF) relative to f _i ^{(I, PF} ) Azim(f _j ^(I,SF) |f _i ^{(I ,PF)} ) and distance Dist(f _j ^(I,SF) |f _i ^(I,PF) ):

Dist(f _j ^{(I, SF)} | f _i ^{(I, PF)} ) = (x _i -x _j ) ² + (y _i -y _j ) ² (6)

In the above formula (x _i , y _i )=Loc (f _i ^{(I, PF)} ), (x _j , y _j )=Loc (f _j ^{(I, SF)} ). And define the relationship strength between f _j ^{(I, SF)} and f _i ^{(I, PF)} as the reciprocal of its distance:

As shown in Figure 3, taking feature point A as an example, the remaining feature points will be regarded as secondary feature points. After selecting a direction as the absolute direction, use this absolute direction as a reference to find the distance between them and point A. The relative orientation θ _i and distance are obtained, and the relationship strength is obtained at the same time, as shown in Table 1.

Table 1

After determining the relationship between all main feature points and all other secondary feature points, the direction estimation unit will determine the main direction of each main feature point (Figure 3 also shows the main direction of a main feature point), and This direction is regarded as the reference direction (zero direction), and the relative orientations of all secondary feature points and the main feature point are mapped clockwise to the range of 0 to 2π. As shown in Figure 4, this histogram shows another example, including the relationship between a primary feature point and all other secondary feature points.

The main direction Ori( _fi ^(I,PF) ) is determined according to the relative orientation of the secondary feature point that has the strongest relationship with the main feature point. Right now

in satisfy:

Denote the relative orientation after remapping as Azim _rm (f _j ^(I,SF) |f _i ^(I,PF) )

If the main direction in Figure 4 is regarded as the reference direction, Figure 5 can be obtained by mapping the relative orientations of all secondary feature points and the main feature point clockwise to the range of 0~2π.

Based on this, the direction intensity histogram statistics unit will count this information and generate a feature description vector. The specific method is as follows: for each main feature point, with this as the center, starting from its main direction, the image is divided into n fan-shaped areas (n is generally more than 10, and the larger n is, the better the discriminability of the generated descriptor is. , but it will increase the calculation amount during feature matching), representing n direction intervals respectively. Then the sum of the relationship strengths of all secondary feature points in each direction interval relative to the main feature point is calculated, which is called the strength of that direction interval. For example, the sum of the relationship strengths of all secondary feature points in the p-th sector area relative to the primary feature point is O _p ( _fi ^(I,PF) |F ^(I,SF) ):

in is the indicator function:

After obtaining the intensity of all direction intervals, the feature description vector V(f _j ^{(I,PF )} |F ^{(I,SF) ) of f i (I,PF} ₎ is composed:

V(f _j ^(I,PF) |F ^(I,SF) )=[O ₁ (f _i ^(I,PF) |F ^(I,SF) ),…,O _n (f _i ^(I,PF) |F ^(I,SF) )] (12)

For example, in Figure 5, all directions are divided into n direction intervals starting from the main direction, where n=10, and then the intensity of the azimuth interval is calculated in each direction interval to obtain the azimuth interval histogram, which is the feature description vector of the feature, such as As shown in Figure 6.

Repeat the above steps to obtain the feature description vectors of all main feature points.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention, but not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that it can still be used Modifications are made to the technical solutions described in the foregoing embodiments, or equivalent substitutions are made to some of the technical features; however, these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims (3)

1. The feature description system based on the inter-feature azimuth distance is characterized in that: the system comprises a Gaussian smoothing unit, a feature detector, a feature screening unit, a feature relation calculating unit and a feature descriptor generating unit;

the feature screening unit comprises a secondary feature screening unit and a primary feature screening unit;

the feature descriptor generating unit comprises a direction estimating unit and a direction intensity histogram statistic unit;

the Gaussian smoothing unit performs smoothing processing on the input image I (x, y) and then performs feature detection through a feature detector;

the feature detector detects a feature point set F of the obtained image ^(I) At the same time obtain feature point f _i ^(I) Position Loc (f in image _i ^(I) )＝(x _i ,y _i ) The characteristic point f _i ^(I) Response intensity Mag (f) _i ^(I) ) The feature point set is expressed as:

F ^(I) ＝{f _i ^(I) |i∈[1,2,...,N _f ]}

wherein i represents the detectedCharacteristic point f _i ^(I) Is the ordinal number of the image, the feature point set of the image comprises N _f Feature points;

the secondary feature screening unit detects the obtained feature point set F by the feature detector ^(I) Secondary feature point screening is carried out; the method comprises the following steps:

introducing modulated characteristic response intensity Mag _m Expressed by the following formula:

wherein M, N represent the width and height of the image; all the characteristic points are ordered according to the descending order of the modulated characteristic response intensity, the characteristic points of the first half part are selected as secondary characteristic points, and then the serial numbers are reassigned, namely f _i ^(I,SF) The method comprises the steps of carrying out a first treatment on the surface of the The secondary feature points form a secondary feature point set F ^(I,SF) Expressed by the following formula:

F ^(I,SF) ＝{f _i ^(I,SF) |i∈[1,2,…,N _f /2]}

the secondary feature point set F ^(I,SF) The main characteristic screening unit is adopted for screening, and main characteristic points are determined, wherein the specific process is as follows:

initializing a set of principal feature points F of an image ^(I,PF) Is empty set, i.e

For each secondary feature point f _i ^(I,SF) Define a circle domain D (f) with radius R _i ^(I,SF) ) Expressed by the following formula:

D(f _i ^(I,SF) )＝{(x,y)|(x-x _i ) ² +(y-y _i ) ² ＜R ² }

if there are no other secondary feature points in the circleSatisfy->Then f is set to _i ^(I,SF) As the main characteristic point, i.e. f _i ^(I,SF) ∈F ^(I,PF) ；

For the main characteristic point set F ^(I,PF) Ordinals are reassigned to all main feature points in the table to obtain a main feature point f _i ^(I,PF) ；

The main feature point set satisfies the following equation:

calculating each principal feature point f by the inter-feature relationship calculating unit _i ^(I,PF) With other secondary characteristic pointsThe relative orientation and distance between them,

the direction estimation unit determines the main direction of each main characteristic point, takes the main direction as a reference direction, and maps the relative directions of all secondary characteristic points and the main characteristic point to a range of 0-2 pi clockwise;

wherein the main direction Ori (f _i ^(I,PF) ) Determining according to the relative orientation of the secondary characteristic point with the strongest relation with the primary characteristic point; namely:

wherein the method comprises the steps ofThe method meets the following conditions:

the relative position after remapping is recorded as

The direction intensity histogram statistics unit will calculate the direction intensity histogram for the main direction Ori (f _i ^(I,PF) ) And the remapped relative orientationCounting to generate a feature description vector; the specific method comprises the following steps:

for each main feature point, taking the main feature point as a center, dividing the image into n sector areas from the main direction of the main feature point, and respectively representing n direction intervals;

then, calculating the sum of the relation intensities of all secondary characteristic points in each direction interval relative to the main characteristic point to form the intensity of the direction interval;

after the intensity of all direction intervals is obtained, the composition f _i ^(I,PF) Feature description vector of (a)Expressed by the following formula:

until the feature description vectors of all the principal feature points are obtained.

2. The inter-feature bearing based characterization system of claim 1, wherein: the inter-feature relation calculation unit calculates each principal feature point f _i ^(I,PF) With other secondary characteristic pointsThe relative orientation and distance between the two parts are as follows:

setting a main characteristic point f _i ^(I,PF) And another secondary feature pointThen->Relative to f _i ^(I,PF) Orientation between->Distance->Expressed by the following formula:

in (x) _i ,y _i )＝Loc(f _i ^(I,PF) )，And define->And f _i ^(I,PF) The reciprocal between them is the relationship strength:

3. the inter-feature bearing based characterization system of claim 1, wherein: the sum of the relation intensity of all secondary characteristic points in the p-th fan-shaped area relative to the main characteristic point is O _p (f _i ^(I,PF) |F ^(I,SF) ) Expressed by the following formula:

wherein the method comprises the steps ofTo indicate a function, it is expressed by the following formula:

after the intensity of all direction intervals is obtained, the composition f _i ^(I,PF) Feature description vector of (a)Until the feature description vectors of all the principal feature points are obtained.