CN102800066B - Local extremum and pixel value gradient based improved image enhancement method - Google Patents

Abstract

The invention discloses an image gradient and local extremum based image edge-preserving filter algorithm. According to the gradient value of pixel values of an image and the distribution information of local extremum points, the position of an edge in the image is judged. Because according to the algorithm disclosed by the invention, the position of an edge of an image is judged through the gradient of pixel values and the local extremum points of pixel values, the edge of the image can be determined more accurately. In the process of filtering noises and detail textures in the image, a filter reduces the influences on the edge of the image in the process of filtering according to the previous judgment, therefore, the filter can separate the contour of an object in the image from an original image.

Description

A kind of image enchancing method based on local extremum and pixel value gradient of improvement

Technical field

The present invention relates to digital image processing field, be specifically related to a kind of improvement based on Local Extremum distribution and the image enchancing method of pixel value gradient.

Background technology

Holding edge filter device, as the basic tool of image procossing, because range of application is very extensive, is current basal image processing tool.Its objective is according to actual needs, by secondary information (such as noise, the grain details etc.) filtering in image, but retain important information (edge of such as object).Certainly, according to faced by practical problems different, even same piece image, its important information and secondary information not immobilize.Such as, when people wish to remove noise in image and obtain high-quality picture, the now noise just elimination as a kind of secondary information; On the contrary, when interference type suffered by the pattern of noise in people's Water demand image thus hypothetical system, the noise now in image must be retained as important information again.

As far back as the 1950's at initial stage of image processing techniques, because its predecessor---image filter---basic status and image processing techniques in widespread attention in multiple application, so the figure of image filter can be seen in Aero-Space, military guidance, robot vision, police and judicial, medical science detection and biomedical engineering etc.In image filtering technology in early days, research emphasis how image (such as noise, edge fog etc.) lower for quality is in a certain respect passed through process, improves this aspect quality and form output image.Based on this background, academia uses the Fourier transform in frequency-region signal treatment technology, definition image filtering.The signal transacting thought of one dimension is extended to two dimensional image signal field by it.

In recent years, in practical application to image processing techniques no longer as only requiring that image processing techniques is improved a certain width input picture single aspect quality or carries out multiple tracks process to image simply thus improve the lower image quality index of several relevance in early days; It requires that image processing techniques provides a kind of comprehensive method more, to improve the multiple index of image, or even some conflicting indexs.In fact, edge keeps and these two targets of filtering are conflicting to a certain extent: the object of wave filter itself is noise in removal of images and grain details, but this understands the edge in blurred picture to a certain extent.This is because from traditional image processing techniques angle, noise, grain details and image border three are the HFSs in image, use traditional method (such as Gaussian wave filter and Butterworth wave filter) can not obtain desirable effect so simple.Boundary filter is just for tackling the requirement of this contradiction and the wave filter proposed.

The inventive method may produce the defect of halation etc. for existing holding edge filter device, proposes a kind of filtering method of improvement, is separated better, has good treatment effect to the object edge profile in image with article surface vein.

Summary of the invention

A kind of image border based on local extremum and pixel value gradient of improvement is the object of the present invention is to provide to keep filtering method, can be used for image texture and strengthen the stylization process of image (such as fuzzy and that contrast is low image is strengthened or), the Digital Image Processing such as the tone reproduction of high-dynamics image, improve the overall visual effect of image.

The present invention, a kind of image border based on partial gradient and pixel gradient value keeps filtering method, on the basis of bilateral filtering method and local extremum filtering method, by the mode that above-mentioned two kinds of different filtering methods are merged, propose a kind of local extremum distributed intelligence according to neighborhood of pixel points and pixel gradient and carry out the method that image border keeps filtering.Being described below of new method:

Note input picture I (x, y) size is (b-a) × (d-c), and at two dimensional surface region Ω: on [a, b] × [c, d], output image is O (x, y),

(1) by output image O (x, y] be initialized as zero;

(2) the subimage I of a k × k is defined in the input picture upper left corner _band establish the step-length of its level and vertical direction movement to be respectively h and v;

(3) to subimage I _bcarry out image procossing according to the filtering method based on local extremum and pixel gradient, require to process all pixels of whole subgraph block, its result is outputted in the position of the output image O (x, y) corresponding to subimage central point;

(4) sub-block is moved horizontally block to move horizontally step-length h, if sub-block does not exceed image boundary, repeat step (4), otherwise enter next step;

(5) with the vertical mover segment of vertical moving step length v, if sub-block does not exceed image boundary, repeat step (4), otherwise enter next step;

(6) after above step completes, output image O (x, y) will be obtained.

The described filtering method based on local extremum and pixel gradient, its method is described below:

(1) the Local Extremum number of original image pixels point is added up

For reducing noise in image to judging in image that whether certain any be the impact of extreme point, the extreme point of this method in the following way in check image.As point (x ₀, y ₀) pixel value at most than the pixel value hour of k-1 point in the k × k neighborhood centered by it, algorithm is thought (x ₀, y ₀) be the Local modulus maxima of image I.Similarly, algorithm judges the local minizing point of image in the same way.Above-mentioned account form is equivalent to think that the pixel value vibration interval that the texture of image causes only has at most k pixel.By adjusting the size of parameter k, the texture of various frequency is removed by final algorithm.

By said method, the number I of statistical pixel point extreme point in k × k neighborhood _extrma.

(2) filter template based on local extremum and pixel gradient is calculated

The present invention on original two-sided filter basis, by introducing the local extremum information template to two-sided filter.Pixel domain weight in the template of two-sided filter is added to the Local Extremum number I of corresponding pixel points _extrmaas correction term.Now, based on the difference of the pixel value between image each point, filter template not only judges whether this point is positioned at the edge of object, judges simultaneously, thus avoid edge texture large for Oscillation Amplitude being mistaken for object according to the number of extreme point in neighborhood of pixel points.

A kind of image border based on local extremum and pixel gradient that the present invention proposes keeps filtering method, has the feature of following two aspects:

(1) this method is on the basis of two-sided filter method and local extremum filtered method, by utilizing the template of Local Extremum distributed intelligence correction two-sided filter, proposes one more accurate image border maintenance filtered method.

(2) method after improving is compared with local extremum wave filter with two-sided filter, can produce good effect, and the complexity of algorithm is identical with two-sided filter.

Accompanying drawing explanation

Fig. 1 is passenger counting system integral module block diagram of the present invention;

Embodiment

Filtered method is kept to be described in detail below in conjunction with accompanying drawing to a kind of image border based on local extremum and pixel gradient of the present invention.

The present invention is improving one's methods of proposing on two-sided filter method basis.Therefore, the image border first specifically describing traditional gauss low frequency filter method and two-sided filter keeps filtering method.

For convenience of discussing, here only for single channel image.Note input picture I size is (b-a) × (d-c), at two dimensional surface region Ω: on [a, b] × [c, d].To any point s ∈ Ω in image, defining the point set that its neighborhood point forms is Q _s.The pixel value of some s is I _s, and after remembering the t time iteration, this pixel value is I _s ^t.According to above-mentioned notation convention, the result of the secondary rear gained of Gauss spatial domain filter iterative computation (t+1) can be expressed as:

$I_s^{t+1}=\frac{1}{k\left(s\right)}\underset{p\∈Q_s}{\mathrm{\Σ}}g(I_p^t-I_s^t)I_p^t,$

Wherein function g (x) is Gauss function, and k (s) is the normalized factor of coefficient, namely has

$g\left(x\right)=\frac{1}{\mathrm{\σ}\sqrt{2\mathrm{\π}}}\exp (-\frac{x^2}{2{\mathrm{\σ}}^2}),$ $k\left(s\right)=\underset{p\∈Q_s}{\mathrm{\Σ}}g(I_p^t-I_s^t).$

Similarly, the result of the secondary rear gained of two-sided filter iterative computation (t+1) can be expressed as:

$I_s^{t+1}=\frac{1}{K\left(s\right)}\underset{p\∈Q_s}{\mathrm{\Σ}}{g}_S\left(\right|p-s\left|\right)g_R(I_p^t-I_s^t)I_s^t,$

Wherein function g _s(x) and g _rx () is all Gauss function, act in spatial domain and pixel domain respectively; K (s) is coefficient normalized factor, namely has

$K\left(s\right)=\underset{p\∈Q_s}{\mathrm{\Σ}}{g}_S\left(\right|p-s\left|\right)g_R(I_p^t-I_s^t).$

The present invention is by Local Extremum number I in computed image neighborhood _extrma, revise the weight of Filtering Template in two-sided filter, particularly,

$I_s^{t+1}=\frac{1}{T\left(s\right)}\underset{p\∈Q_s}{\mathrm{\Σ}}[g_R(I_p^t-I_s^t)+I_{\mathrm{Extrma}}\left(s\right)]g_S\left(\right|p-s\left|\right)I_p^t---\left(1\right)$

Wherein function T (s) is the normalized factor of coefficient, namely has

$T\left(s\right)=\underset{p\∈Q_s}{\mathrm{\Σ}}[g_R(I_p^t-I_s^t)+I_{\mathrm{Extrma}}\left(s\right)]g_S\left(\right|p-s\left|\right)$

Rudimentary algorithm process flow diagram please refer to Figure of description 1.

Above content is in conjunction with concrete preferred implementation further description made for the present invention, can not assert that specific embodiment of the invention is confined to these explanations.For general technical staff of the technical field of the invention, without departing from the inventive concept of the premise, some simple deduction or replace can also be made, all should be considered as belonging to protection scope of the present invention.

Claims (1)

1. the image enchancing method based on local extremum and pixel value gradient improved, it is characterized in that: on the basis of bilateral filtering method and local extremum filtering method, by the mode that above-mentioned two kinds of different filtering methods are merged, propose the method, comprise the following steps:

A. remember that input picture I (x, y) size is (b-a) × (d-c), at two dimensional surface region Ω: on [a, b] × [c, d], output image is that O (x, y) adopts following steps afterwards;

B. output image O (x, y) is initialized as zero;

C. the subimage I of a k × k is defined in the input picture upper left corner _band establish the step-length of its level and vertical direction movement to be respectively h and v;

D. to subimage I _bcarry out image procossing according to the filtering method based on local extremum and pixel gradient, require to process all pixels of whole subimage, its result is outputted in the position of the output image O (x, y) corresponding to subimage central point;

E. by subimage with horizontal direction moving step length h, if subimage does not exceed image boundary, repeat step e, otherwise enter next step;

F. by subimage with vertical direction moving step length v, if subimage does not exceed image boundary, repeat step e, otherwise enter next step;

G., after above step completes, output image O (x, y) will be obtained;

Wherein, the computing method step of described steps d is as follows:

(1). the Local Extremum number of statistics original image pixels point: as point (x0, y0) pixel value is at most than the pixel value hour of k-1 point in the k × k neighborhood centered by it, algorithm is thought (x0, y0) be the Local modulus maxima of image I, similarly, algorithm judges the local minizing point of image in the same way, and above-mentioned account form is equivalent to think that the pixel value vibration interval that the texture of image causes only has at most k pixel; By adjusting the size of parameter k, the texture of various frequency is removed by final algorithm, by described method, and the number IExtrma of statistical pixel point extreme point in k × k neighborhood;

(2). calculate the filter template based on local extremum and pixel gradient: on original two-sided filter basis, by introducing the template of local extremum information to two-sided filter, the Local Extremum number IExtrma of corresponding pixel points is added as correction term to pixel domain weight in the template of two-sided filter, now, based on the difference of the pixel value between image each point, filter template not only judges whether this point is positioned at the edge of object, judge according to the number of extreme point in neighborhood of pixel points simultaneously, thus avoid edge texture large for Oscillation Amplitude being mistaken for object.