CN105403890A - Related target detection method based on row and column characteristic vectors - Google Patents

Abstract

The invention belongs to the field of digital image processing, and specifically relates to a related target detection method based on row and column characteristic vectors and used in sonar image underwater target detection. The method comprises the following steps: performing binarization processing on an image through a threshold customized binarization method, performing morphological processing on the image after binarization, performing an edge extraction on the image after the morphological processing through a Sobel edge detection algorithm, performing arc detection by utilizing edge information through an arc detection algorithm based on Hough conversion, performing normalization processing on a suspected region, taking an ideal circle image as a template, and getting a row characteristic vector and a column characteristic vector of the ideal circle image. The invention provides the characteristic vectors capable of simply and efficiently representing circular or approximately circular target region characteristics, and non-target regions can be effectively removed through row and column characteristic screening, so the false alarm probability of target detection can effectively reduced.

Description

A kind of object detection method relevant based on ranks proper vector

Technical field

The invention belongs to digital image processing field, be specifically related to a kind of object detection method relevant based on ranks proper vector being mainly used in detecting underwater object in sonar image.

Background technology

Target detection refers to be determined whether to have target in single image or sequence image and the one of being separated in interested target area and uninterested background area operation.It often utilizes known priori, calculates relevant characteristic quantity, and utilizes these characteristic quantities to distinguish object and background.The fields such as current Underwater Acoustic Object Detection Techniques are widely used in petroleum prospecting, underwater topography is observed, deep-sea pipeline is detected, underwater object salvaging, sea fishery.Military aspect, the great difficult problem of neritic area naval mine detection naval of Cheng Liao various countries.And imaging sonar can detect region under water on a large scale, imaging sonar is therefore utilized to detect torpedo target Cheng Liao various countries study hotspot.

Carrying out naval mine detection in neritic area is very difficult.Torpedo target is little and be often hidden in some man-made obstructions.Zhan Yan center, U.S. sea has carried out entering research in 10 years in the automatic or semi-automatic detection of torpedo target, utilizes non-commutative group's frequency analysis theoretical, have developed many algorithm fusions torpedo target detection method that effectively can reduce false-alarm probability.The method can fast processing sonar image, extracts suspicious region.

Canada Burlinton researchist MartinG.Bello has also carried out large quantity research to the torpedo target detection based on sonar image, takes the target detection recognition methods based on neural network.This algorithm can be divided into four parts, is respectively: the pre-service of sonar image, suspicious region extraction, suspicious region feature extraction, neural network recognization.Can find out that this system has lower false-alarm probability under the condition that detection probability is certain from its characteristic working curve.

Current foreign scholar has carried out large quantity research to torpedo target detection method, obtains some efficiently based on the object detection method of sonar image.And domestic corresponding research is also in the starting stage, also there is no shaping corresponding product.The result that target detection is also extracted directly will affect the result of target identification, and the research therefore for object detection method is significant.

What the present invention mainly proposed is the object detection method that a kind of based target ranks proper vector is relevant.By contrasting with document [1] described method, confirm that the method effectively can reduce false-alarm probability while effectively extracting target.

List of references related to the present invention comprises: J.ChenandZ.Gong, ADetectionMethodBasedonSonarImageforUnderwaterPipelineTr acker.MACE2011, pp.3766-3769.

Summary of the invention

The object of this invention is to provide a kind of object detection method relevant based on ranks proper vector

The object of the present invention is achieved like this:

(1) by the binarization method of self-defined threshold value, binary conversion treatment is carried out to image;

(2) Morphological scale-space is carried out to the image after binaryzation, to reduce spotted noise in image and to fill up the leak in target area;

(3) by Sobel edge detection algorithm, edge extracting is carried out to the image after Morphological scale-space;

(4) utilize marginal information to carry out circular-arc detection by the circular-arc detection algorithm based on Hough transform, and the circular arc region exceeding detection threshold is considered as suspicious region;

(5) normalized is done to suspicious region, go summation and its row proper vector of row read group total and row proper vector;

(6) using circular ideal image as template, ask its row proper vector, row proper vector; The related coefficient of calculation template proper vector and suspicious region proper vector, and threshold value is set, region related coefficient being exceeded threshold value is considered as target area:

Compared with prior art, the invention has the beneficial effects as follows:

The present invention proposes a kind of proper vector that can represent circle or sub-circular target area feature simply efficiently, proper vector and row proper vector at once.This feature vectors by calculate suspicious region row and, row and can obtain, calculated amount is less, and can screen suspicious region by means of only the related coefficient asking for template and suspicious region.Because this feature vectors can the feature in effectively expressing circular target region, effectively can remove nontarget area by the screening of ranks proper vector, therefore can effectively reduce the false-alarm probability of target detection.

Accompanying drawing explanation

Fig. 1 is image to be detected, and wherein RED sector is target area;

Fig. 2 is the pseudo color image of image to be detected;

Fig. 3 is the result that Hough transform extracts suspicious region;

Fig. 4 is the suspicious region intercepted out;

Fig. 5 (a) is template image;

The row proper vector that Fig. 5 (b) is template image;

The row proper vector that Fig. 5 (c) is template image;

Fig. 5 (d) is suspicious region one;

The row proper vector that Fig. 5 (e) is suspicious region one;

The row proper vector that Fig. 5 (f) is suspicious region one;

Fig. 5 (g) is suspicious region two;

The row proper vector that Fig. 5 (h) is suspicious region two;

The row proper vector that Fig. 5 (i) is suspicious region two;

Fig. 5 (j) is suspicious region three;

The row proper vector that Fig. 5 (k) is suspicious region three;

The row proper vector that Fig. 5 (l) is suspicious region three;

Fig. 6 is the former figure of another width side-scanning sonar image, and wherein RED sector is target area;

Fig. 7 is based on the relevant target detection result one of ranks proper vector;

Fig. 8 is the target detection result one based on Gabor wavelet;

Fig. 9 is based on the relevant target detection result two of ranks proper vector;

Figure 10 is the target detection result two based on Gabor wavelet.

Embodiment

Below in conjunction with accompanying drawing, the present invention is described further.

The invention provides one in high noisy image, detect circular (or sub-circular) order calibration method.This is a kind of object detection method utilizing target row and column proper vector to screen noisy image suspicious region.In most cases, artificial immersed body all can have available circular arc to exist.Therefore first we carry out Morphological scale-space to noisy image, then uses Hough transform to mark circular arc, find suspicious region.For each suspicious region, we calculate its row and with row and, respectively as row proper vector (RowSumVector, RSV) and row proper vector (ColumnSumVector, CSV).Ask the related coefficient of the proper vector of gained proper vector and priori, large related coefficient means that suspicious region and target prior imformation have high shape similarity.By this principle, we can filter out target area from target area suspicious.The present invention includes:

(1) binary conversion treatment is carried out to gray level image;

(2) Morphological scale-space is carried out to the image after binaryzation;

(3) Sobel operator is utilized to carry out edge extracting;

(4) utilize Hough transform to detect circular arc, and discernible circular arc region is considered as suspicious region;

(5) suspicious region is normalized, and it is cumulative to carry out row cumulative sum row respectively, obtains row proper vector and the row proper vector of suspicious region;

(6) ask for the related coefficient of the ranks proper vector of circular shuttering image and the ranks proper vector of suspicious region, and arrange threshold value, region related coefficient being exceeded threshold value is considered as target area;

(7) utilize two width side-scanning sonar images to carry out testing and contrast with a kind of Target Recognition Algorithms based on Gabor transformation, demonstrating this method, under the condition that Detection results is similar, there is lower false-alarm.

Wherein:

(1) binaryzation is carried out to pending image

Pending gradation of image figure as shown in Figure 1, is the sonar image obtained by side-scan sonar, and containing much noise, target area is red by painting.In order to Pseudo Col ored Image is carried out to it in clearer display-object region, result as shown in Figure 2.

First carry out binary conversion treatment to gray level image, what adopt here is a kind of method of self-defined threshold value.

(2) Morphological scale-space is carried out to the image after binaryzation

Morphological operation mainly comprises expansion, burn into opens operation, closed operation etc.Morphological scale-space effectively can remove the spotted noise in image, and fills up the hole in target area.

(3) Sobel operator is utilized to carry out edge extracting to it after pre-service

Sobel edge detection operator is a kind of detection method of gradient, and its principle is by being weighted summation in neighborhood of a point to be detected, determines whether measuring point to be checked is extreme point, and then determines whether center pixel is marginal point.

The mathematic(al) representation of Sobel edge detection operator is:

S(i,j)＝|f _x|+|f _y|

Wherein, f _xand f _yavailable following convolution mask represents:

(4) Hough transform is utilized to carry out loop truss

Hough transform is circle detection method conventional in Digital Image Processing, and its basic parameter detected is central coordinate of circle and radius of circle.Set up an office collection { (x _i, y _i) | i=1,2 ..., n} is by forming a little on circle a certain in detection space, and the center of circle is designated as (a, b), and radius is r, and this circle equation under rectangular coordinate system is:

(x _i-a) ²+(y _i-b) ²＝r ²

Then on image, a bit (x, the y) equation in parameter coordinate system can be written as:

(a-x) ²+(b-y) ²＝r ²

This equation is three conical surfaces, and therefore the upper every bit of circle is all corresponding with three conical surfaces in parameter space, and the point on same circle will meet at a bit in parameter space, and the coordinate of this point will show position and the radius information of circle in rectangular coordinate system.Concrete steps are as follows:

1) utilize the sensing range of priori determination radius of circle, and set up parameter space, by every bit zero setting according to this scope and image actual size.

2) every bit on image mapped to parameter space and add up.

3) arrange threshold value, in parameter space, carried out threshold test, then these positional informations in parameter space will reflect the parameter information of detected circle.

4) circle detected is marked out on gray-scale map come.

Hough transform carries out the result of loop truss as shown in Figure 3.

(5) ranks proper vector correlation method is utilized to screen suspicious region

First suspicious region is intercepted out, intercept result as shown in Figure 4.The template image of target, as shown in Fig. 5 (a), carries out to it that row cumulative sum row are cumulative can obtain row proper vector and row proper vector.Then row proper vector can be expressed as

Row proper vector can be expressed as:

The row proper vector of template image and row proper vector are respectively as shown in Fig. 5 (b) He Fig. 6.In gray-scale map, suspicious region is intercepted by square, ask its row proper vector and row proper vector respectively.

(6) suspicious region is normalized, and calculates the related coefficient of suspicious region proper vector and template characteristic vector, to carry out target area screening.

Ask for the related coefficient of row proper vector with template image and row proper vector respectively, related coefficient can be expressed as:

Wherein the proper vector of x and the y proper vector and suspicious region that are respectively template image carries out threshold test, and related coefficient exceedes the be regarded as target area of certain value.

Fig. 5 (d) (g) (j) is respectively three kinds of typical suspicious regions, wherein Fig. 5 (d) is target area, Fig. 5 (e) (f) is respectively its row proper vector and row proper vector, can find that its row, column proper vector is all comparatively similar to the row, column proper vector of template image, related coefficient reaches 0.7625 and 0.9360 respectively.And the row proper vector of Fig. 5 (g) is comparatively similar, row proper vector difference is comparatively large, and related coefficient is respectively 0.6167 and 0.0113, is therefore got rid of.And the row proper vector of Fig. 5 (j) is comparatively similar, row proper vector difference is comparatively large, and related coefficient is respectively 0.0864 and 0.2724, therefore also by it eliminating.

Testing result marks the most at last, as shown in Figure 7.Wherein RED sector is target area.

(7) contrast with a kind of Target Recognition Algorithms based on Gabor transformation

Algorithm described in document [1] is a kind of algorithm of target detection based on Gabor wavelet.Compare after it realization with result of the present invention, find that algorithm of the present invention has less false-alarm under the condition that all target area can be marked.After again the sonar image that another width contains circular target is processed, its former figure is as Fig. 6, and result is as shown in Figure 9.Find that the present invention has less false-alarm equally.

Claims (1)

1., based on the object detection method that ranks proper vector is relevant, it is characterized in that, comprise the steps:

(1) by the binarization method of self-defined threshold value, binary conversion treatment is carried out to image;

(2) Morphological scale-space is carried out to the image after binaryzation, to reduce spotted noise in image and to fill up the leak in target area;

(3) by Sobel edge detection algorithm, edge extracting is carried out to the image after Morphological scale-space;

(4) utilize marginal information to carry out circular-arc detection by the circular-arc detection algorithm based on Hough transform, and the circular arc region exceeding detection threshold is considered as suspicious region;

(5) normalized is done to suspicious region, go summation and its row proper vector of row read group total and row proper vector;

$RSV=\underset{j}{\Σ}I(i,j);$

$CSV=\underset{i}{\Σ}I(i,j)$

(6) using circular ideal image as template, ask its row proper vector, row proper vector; The related coefficient of calculation template proper vector and suspicious region proper vector, and threshold value is set, region related coefficient being exceeded threshold value is considered as target area:

$r=\frac{\underset{i=1}{\overset{n}{\Σ}}\left(x_i-\stackrel{\‾}{x}\right)\left(y_i-\stackrel{\‾}{y}\right)}{\sqrt{\underset{i=1}{\overset{n}{\Σ}}{\left(x_i-\stackrel{\‾}{x}\right)}^2\·\underset{i=1}{\overset{n}{\Σ}}{\left(y_i-\stackrel{\‾}{y}\right)}^2}}=\frac{n\underset{i=1}{\overset{n}{\Σ}}{x}_i{y}_i-\underset{i=1}{\overset{n}{\Σ}}{x}_i\·\underset{i=1}{\overset{n}{\Σ}}{y}_i}{\sqrt{n\underset{i=1}{\overset{n}{\Σ}}{x}_i^2-{\left(\underset{i=1}{\overset{n}{\Σ}}{x}_i\right)}^2}\·\sqrt{n\underset{i=1}{\overset{n}{\Σ}}{y}_i^2-{\left(\underset{i=1}{\overset{n}{\Σ}}{y}_i\right)}^2}}.$