CN102855501A - Multi-direction object image recognition method - Google Patents

Abstract

The invention discloses a multi-direction object image recognition method and belongs to the field of intelligent computer recognition. The method comprises the following steps of: 1) reading an image to be detected and a trained unidirectional haar characteristic sequence, and rotating the characteristic sequence to form a characteristic sequence in a plurality of set directions; 2) setting the image to be detected as a grayscale image, and performing histogram equalization on the grayscale image; 3) performing pixel integration on the equalized image to be detected, and forming an integral image; 4) scaling the image to be detected, and calculating the characteristic sequence in each set direction and the characteristic value of the integral image; and 5) sliding a monitoring window in the image, and judging whether the calculated characteristic values are matched in each set direction, outputting a matched monitoring window range if the matching is successful, and detecting the recognition failure if the matching in all the set directions fails. According to the method, only once training is required, a multi-direction picture can be recognized, and the recognition efficiency is greatly improved.

Description

A kind of multi-direction subject image recognition methods

Technical field

The present invention relates to computer intelligence identification field, relate in particular to a kind of multi-direction subject image recognition methods.

Background technology

At present computer intelligence identification field object identification has a lot of methods.Its accuracy and speed has all reached the degree of practical application.

Wherein relatively more outstanding a kind of method is carried out object identification by the haar feature exactly.In this method, use very cleverly the reduced form of haar feature, and by using integral image to calculate the pixel integration value of rectangle to be detected, just can reach according to the calculating that rectangular characteristic is carried out the haar feature purpose of recognition object.

Haar simplifies feature

Haar simplifies feature and sees intuitively it is a kind of pixel comparison information based on rectangle.At first proposed by people such as Paul Viola, its effect be exactly use in the rectangle zoning pixel and, the area pixel that pixel color is dark and value can be lower; The area pixel that pixel color is shallow and value can be higher.Such as the definition of Haar_x3 feature among Fig. 1, there is the darker zone of color in the zone that meets this feature in the middle of must being, and the zone on both sides is more shallow.Such as Haar_y2 feature among Fig. 1, the zone that meets its feature must be that the upper area color is more shallow, and the bottom color is darker.The description of certainly complete haar feature must also have the position of area coordinate, the description of area size, also have field color and comparison threshold value.

The position of area coordinate is the relative position in the detection window that determines in training and when identification, and the monitoring window is to slide with convergent-divergent according to different scale in the zones of different of mapping to be checked.

In practical application, the description of this type of haar feature also determines that by describing three parts mainly comprise the coordinate of upper left angle point, peak width contrasts the degree threshold value in addition.

Its eigenwert can be formulated as

${\mathrm{feature}}_j=\underset{i\∈\left(\mathrm{1,2}\right)}{\mathrm{\Σ}}{\mathrm{\ω}}_i\mathrm{RectSum}\left(r_i\right)$

Wherein: feature _jRepresent the eigenwert of j kind Haar piece; ω _iThe weights of j Haar piece i sub-rectangle (black rectangle or white rectangle); RectSum (r _i) representing the pixel value sum of all picture elements that comprise in i the Haar piece, the weights in the eigenwert of different Haar block feature prototypes are than (ω ₁: ω ₂) can be in training determine progressively that a kind of method that initial value is set is exactly that inverse ratio according to area arranges (feature result calculate after near 0).If feature value result of calculation in image rectangular area is near 0 (less than a threshold value) then think that this rectangular area meets feature during follow-up judgement.

In order to calculate fast RectSum (r _i), adopt the integrogram method.Each pixel gray scale of the method definition piece image is i (x, y), and each the pixel value ii (x, y) in the integrogram of this width of cloth image is expressed as so:

$\mathrm{ii}(x,y)=\underset{x^{\&prime;}<x,y^{\&prime;}<y}{\mathrm{\&Sigma;}}i(x^{\&prime;},y^{\&prime;})$

For the integrogram value of piece image in the arbitrfary point, integral domain be upper left angle point to the rectangular pixels integrated value take this point as the lower right corner, can by to row and column accumulate once the circulation obtain

s(x，y)＝s(x，y-1)+i(x，y)

ii(x，y)＝ii(x-1，y)+s(x，y)

Wherein, s (x, y) is the row integrated value of point (x, y) position, but does not comprise the value of (x, y) point.

S when iteration is initial (x ,-1)=0, and ii (1, y)=0.The integrogram value of Fig. 2 (a) mid point (x, y) is the pixel gray-scale value summation in grey rectangle zone.Utilize integrogram to sue for peace to the gray-scale value in any one rectangle in the image easily, shown in Fig. 2 (b).

Identifying

Identifying at first requires the picture (preferably this object to be identified position in picture, shape, direction are relatively consistent) of a large amount of objects to be identified, is called positive sample.Also need in addition a large amount of not pictures of this object, be called negative sample.

Use above sample training, training process utilizes the haar feature to use the adaboost method through a large amount of circulations, uses the mass data training.Finally obtain to represent the sequence of the Haar feature of object to be identified.Sequence by these features can guarantee that object to be detected is identified according to discrimination and the false drop rate that the user needs.

There have been this data just can the picture to be detected of arbitrary content to be detected.

Testing process:

At present, realization has been arranged in the software project that some are increased income.It roughly uses flow process as follows:

1. read in image to be detected

2. read in the haar characteristic sequence (one direction) that has trained

3. picture to be detected is changed to gray level image

4. picture to be detected is carried out histogram equalization

5. picture to be detected is carried out pixel integration, and form integral image

6. picture to be detected is carried out proportional zoom

7. carry out eigenwert calculating in conjunction with haar characteristic sequence and integral image.

8. in image, carry out window sliding, and whether judging characteristic mates.

9. if coupling then detect and to return is not successfully mated then to detect and is returned failure

10. the window ranges of coupling is monitored in output

Present computer intelligence identification field has a lot of algorithms all to support object identification.Discrimination also reaches higher level.But a lot of problems are also arranged in the process of object identification, are exactly a larger problem for the identification of rotating rear object.Such as the most frequently used method for detecting human face.Often all can reach accuracy rate more than 90% for the detection of homo erectus's face.But in the uncertain situation of the shooting angle of some image.In a single day people's face rotates variation.The sorter that has trained just can not adapt to new variation, and the postrotational people's face of None-identified.In this model of cognition, often the feature of training is that what direction is so feature of identification of what direction also be.If needs detect different directions, need to be detecting again behind the picture integral-rotation.

Summary of the invention

The recognition methods that the present invention is directed to above-mentioned prior art only can be carried out the defective that one direction detects, and has developed a kind of new multi-direction object identification method, can carry out on the basis of original unidirectional training characteristics sequence the object identification of 4 directions.

Technical scheme of the present invention is: a kind of multi-direction object identification method comprises the steps:

1. read in image to be detected;

2. read in the one direction haar characteristic sequence that has trained;

3. image to be detected is made as gray level image;

4. gray level image to be detected is carried out histogram equalization;

5. the image to be detected after the equalization is carried out pixel integration, form integral image;

6. treat detected image and carry out proportional zoom;

7. one direction haar characteristic sequence is rotated to be the characteristic sequence (such as the characteristic sequence that becomes the four direction of 90 degree multiples with the training direction) of some direction initializations;

8. the eigenwert of calculating each direction initialization according to haar characteristic sequence and the integral image of each direction initialization;

9. whether the monitoring window slides in image, mate at each direction initialization judging characteristic;

10. if any direction at direction initialization has characteristic matching, the scope of the monitoring window of output matching then.

Described haar characteristic sequence is the characteristic rectangle sequence, and each eigenmatrix is with the upper left corner horizontal ordinate of this rectangle, upper left corner ordinate, and rectangle is wide, and the rectangle height represents.

The computing method of described eigenwert are:

${\mathrm{feature}}_j=\underset{i\&Element;\left(\mathrm{1,2}\right)}{\mathrm{\&Sigma;}}{\mathrm{\&omega;}}_i\mathrm{RectSum}\left(r_i\right)$

Wherein: feature _jRepresent the eigenwert of j kind Haar piece; ω _iThe weights of j Haar piece i sub-rectangle; RectSum (r _i) represent the pixel value sum of all picture elements that comprise in i the Haar piece.

Described sub-rectangle is white or black rectangle.

Described one direction haar characteristic sequence is rotated to be the characteristic sequence of four direction, its method is:

1) one direction haar characteristic sequence is copied as four parts;

2) four parts of haar characteristic sequences are spent by same direction rotation 0 degree, 90 degree, 180 degree and 270 respectively, obtain the haar characteristic sequence of four direction.

It is on the basis of one direction object identification system that the present invention realizes, carries out the identification of 4 directions and improves.In most of the cases.Use the fixing positive unidirectional object features of shape to train for object identification training process General Requirements, final identification also is to identify for the unidirectional object in front.Can maximize the accuracy rate of identification like this, and minimize the error rate of identification.Narrow but but the shortcoming of doing so but is the scope that has fixedly caused recognition result of sample, in case the identification of will having no idea the problem of picture rotation appears.The present invention has solved above problem to a certain extent.90 degree interval rotations by the haar rectangular characteristic.Only need to carry out its training result of use that positive unidirectional training just can four times of efficient, can solve the identification problem of four direction in the flow process of once identification, the differentiation angular interval of algorithm identified has also enlarged four times.Can only support the identification of direct picture such as original algorithm, but can be picture or the handstand picture one-off recognitions of positive and negative 90 degree of camera shooting by the method; Original identification angle only covers positive and negative 15 degree, and this method just can expand theoretic positive and negative 60 degree to so.And the method has minimized the outer treatment step of identifying, has improved the efficient of computing.

Reason has following several:

1. feature itself is carried out the rotation of 90 degree easily based on the shape of rectangle.Postrotational feature remains rectangle.Do not change for characteristic sequence.

2. the calculating of integral image as long as feature itself is rectangle, just can be recomputated integral image and directly carries out the calculating of haar feature.

3. do not change in the situation of rectangle size in rotation, the characteristic threshold value of corresponding moving window is constant.

4. all carry out in the situation of same angle (multiples of 90 degree) rotation at whole piece haar characteristic sequence.The detectability of characteristic sequence is constant.

Wherein as follows for the spinning solution of characteristic rectangle:

If the rectangle after transforming is r[k]; Rectangle before transforming is reck[k]; K rectangle in the k representative feature.Window_height and window_width be the Gao Yukuan of window that representative feature relies on respectively.R[k] .x, r[k] .y, r[k] .width, r[k] .heigh is respectively the upper left corner horizontal ordinate of rectangle, upper left corner ordinate, rectangle is wide, and rectangle is high.Reck[k] expression and r[k] the same.

But find that after scrutinizing there is easily spinning solution in Haar feature itself.

In conjunction with the Haar feature and to the thinking of its rotation, when judging, only need carry out feature the simple change of following four direction, both can obtain the judged result of four direction.

Feature rotation 0 degree

r[k]＝rect[k]

Take picture centre as initial point, dextrorotation turn 90 degrees,

r[k].x＝window_height-rect[k].y-rect[k].height；

r[k].y＝window_width-rect[k].x-rect[k].width；

r[k].width＝rect[k].height；

r[k].height＝rect[k].width；

Take picture centre as initial point, the dextrorotation turnback

r[k].x＝window.width-rect[k].x-rect[k].width；

r[k].y＝window.height-rect[k].y-rect[k].height；

r[k].width＝rect[k].width；

r[k].height＝rect[k].height；

Take picture centre as initial point, 270 degree turn clockwise

r[k].x＝rect[k].y；

r[k].y＝rect[k].x；

r[k].width＝rect[k].height；

r[k].height＝rect[k].width。

Compared with prior art, good effect of the present invention is:

The present invention only need train once, just can be high efficiency for handstand, and the picture of 90-degree rotation and 270 degree is identified.In daily life, and the maximum rotation of picture also be carry out 90 the degree 180 the degree and 270 the degree rotations.Therefore the invention solves above problem.

Description of drawings

Fig. 1 is that haar simplifies characterizing definition figure;

(a) Haar_x3 feature legend, (b) Haar_y2 feature legend,

(c) central point feature legend, (d) inclination central point feature legend;

Fig. 2 is the integral image synoptic diagram;

(a) any pixel integration surface area of minute dot image representative (from the image upper left corner);

(b) pixel integration obtains by the long-pending plus-minus of four dot product score values of rectangle in any rectangular area;

Fig. 3 is the inventive method process flow diagram.

Embodiment

Be illustrated in figure 3 as a kind of multi-direction object identification method of the present invention, comprise the steps:

1, reads in image to be detected;

2, read in the one direction haar characteristic sequence that has trained;

The haar characteristic sequence is the characteristic rectangle sequence, and each eigenmatrix is with the upper left corner horizontal ordinate of this rectangle, upper left corner ordinate, and rectangle is wide, and the rectangle height represents.

3, image to be detected is made as gray level image;

4, gray level image to be detected is carried out histogram equalization;

5, the image to be detected after the equalization is carried out pixel integration, form integral image;

6, one direction haar characteristic sequence is rotated to be the characteristic sequence of four direction;

One direction haar characteristic sequence is copied as four parts;

Four parts of haar characteristic sequences respectively by same direction rotation 0 degree, 90 degree, 180 degree and 270 degree, are obtained the haar characteristic sequence of four direction.

7, carry out eigenwert calculating in conjunction with haar characteristic sequence and integral image.(calculating the eigenwert of four direction)

The computing method of eigenwert are:

${\mathrm{feature}}_j=\underset{i\&Element;\left(\mathrm{1,2}\right)}{\mathrm{\&Sigma;}}{\mathrm{\&omega;}}_i\mathrm{RectSum}\left(r_i\right)$

Wherein: feature _jRepresent the eigenwert of j kind Haar piece; ω _iThe weights of j Haar piece i sub-rectangle; RectSum (r _i) represent the pixel value sum of all picture elements that comprise in i the Haar piece.Sub-rectangle is white or black rectangle.

8, the monitoring window slides in image to be detected, and (size variation, the minimum window that allows from feature begin the size that maximum is no more than image itself such as 2 pixels.Position movement begins to move to the lower right corner of image always from the upper left corner of image), whether mate at the four direction judging characteristic;

If 9 have direction coupling then detect and return successfully, and the scope of the monitoring window of output matching.Four direction does not mate then detection and returns failure

Claims (7)

1. a multi-direction subject image recognition methods the steps include:

1) reads in image to be detected; Read in the one direction haar characteristic sequence that has trained, and one direction haar characteristic sequence is rotated to be the characteristic sequence of some direction initializations;

2) image to be detected is made as gray level image, then this gray level image is carried out histogram equalization;

3) image to be detected after the equalization is carried out pixel integration, form integral image;

4) treat detected image and carry out proportional zoom, then calculate the haar characteristic sequence of each direction initialization and the eigenwert of integral image;

Whether 5) the monitoring window slides in image, mate at each direction initialization judging characteristic successively according to the eigenwert of calculating;

If characteristic matching on certain direction initialization, then the match is successful and the scope of the monitoring window of output matching, if all it fails to match for all direction initializations, then detects recognition failures.

2. the method for claim 1 is characterized in that the characteristic sequence of described some direction initializations is for becoming the characteristic sequence of the four direction of 90 degree multiples with the training direction.

3. method as claimed in claim 2 is characterized in that the acquisition methods of the characteristic sequence of described four direction is: at first one direction haar characteristic sequence is copied as four parts; Then four parts of haar characteristic sequences are spent by same direction rotation 0 degree, 90 degree, 180 degree and 270 respectively, obtain the haar characteristic sequence of four direction.

4. method as claimed in claim 3 is characterized in that take the testing image center as initial point, and the haar characteristic sequence turns clockwise.

5. as claimed in claim 1 or 2 or 3 or 4 method is characterized in that described haar characteristic sequence is the characteristic rectangle sequence,

Each eigenmatrix is with the upper left corner horizontal ordinate of this rectangle, upper left corner ordinate, and rectangle is wide, and the rectangle height represents.

6. method as claimed in claim 5 is characterized in that the computing formula of described eigenwert feature is:

Wherein, feature _jRepresent the eigenwert of j kind Haar piece, ω _iThe weights of j Haar piece i sub-rectangle, RectSum (r _i) represent the pixel value sum of all picture elements that comprise in i the Haar piece.

7. method as claimed in claim 6, its feature is white or black rectangle at described sub-rectangle.

Images