CN105225238A - A kind of gray space division methods of the Image semantic classification based on mean filter - Google Patents

Abstract

Based on a gray space division methods for the Image semantic classification of mean filter, comprise the following steps the importance according to RGB tri-components and other index, three components are weighted average calculating operation with different weights; Take the first-selected advanced column hisgram correction of histogram equalization, utilize greyscale transformation function to be modified to the histogram of original image and be uniformly distributed, and then carry out histogram equalization; Carry out mean filter to image, taking is Global thresholding, and in binarization, only use the method for a global threshold T, the gray-scale value of each pixel of image and T compare by it, if be greater than T, are then taken as foreground; Otherwise, be taken as background colour; Using the maximum center point P of respective pixel number in L grey level range as initial classes average.When i-th iteration, investigate each pixel, calculate the spacing between it and the average of each gray level, each pixel is composed the average class nearest apart from it, for j=1,2 ... l, calculates new cluster centre, upgrades class average, all pixels is investigated one by one.

Description

A kind of gray space division methods of the Image semantic classification based on mean filter

Technical field

The present invention relates to technical field of image processing, provide a kind of gray space division methods of the Image semantic classification based on mean filter.

Background technology

Iamge Segmentation is exactly image is divided into several are specific, have peculiar property region and proposes technology and the process of interesting target.It is by the committed step of image procossing to graphical analysis.Existing image partition method mainly divides following a few class: the dividing method based on threshold value, the dividing method based on region, the dividing method based on edge and the dividing method etc. based on particular theory.Since 1998, researchist updates original image partition method and some new theories of other subject and new method for Iamge Segmentation, proposes much new dividing method.The target extracted after Iamge Segmentation may be used for image, semantic identification, picture search etc. field.

Digital image processing techniques are fields interdisciplinary.Along with the development of computer science and technology, image procossing defines the scientific system of oneself gradually with analysis, and new disposal route emerges in an endless stream, although its developing history is not long, causes the extensive concern of each side personage.First, vision is the most important perception means of the mankind, and image is again the basis of vision, and therefore, the digital picture scholars become in the numerous areas such as psychology, physiology, computer science study the effective tool of visually-perceptible.Secondly, image procossing has ever-increasing demand in the large-scale application such as military affairs, remote sensing, meteorology.

Since 1998, artificial neural network recognition technology has caused and has paid close attention to widely, and is applied to Iamge Segmentation.Basic thought based on the dividing method of neural network obtains linear decision function by training multi-layer perception（MLP）, and then classifying to pixel with decision function reaches the object of segmentation.This method needs a large amount of training datas.There is the connection of flood tide in neural network, easily introduces spatial information, can solve the noise in image and problem of non-uniform preferably.Which kind of network structure is selected to be the subject matter that this method will solve.

Iamge Segmentation is the vital pre-service of image recognition and computer vision.Do not have correct segmentation just can not have correct identification.But, carry out splitting brightness and color that only foundation is pixel in image, when automatically processing segmentation by computing machine, will all difficulties be run into.Such as, uneven illumination is even, the impact of noise, there is unsharp part in image, and shade etc., usually there is segmentation errors.Therefore Iamge Segmentation is the technology needing research further.People wish the method introducing some artificial knowledge guiding and artificial intelligence, for correcting the mistake in some segmentation, being up-and-coming methods, but which again increases the complicacy of dealing with problems.

In the field of communications, image Segmentation Technology is very important to the transmission of the live images such as videophone, movable part in image and static background is needed to separate, also will regions different for displacement in movable part separately, to the different coding transmission in region of Activity, to reduce the code check needed for transmission.

Summary of the invention

The object of the invention is to improve nicety of grading and accuracy.Can real-time stabilization to Target Segmentation extract, the gray space division methods that segmentation effect is good.

The present invention is to achieve these goals by the following technical solutions:

Based on a gray space division methods for the Image semantic classification of mean filter, it comprises the following steps:

Step 1, according to the importance of RGB tri-components and other index, three components are weighted average calculating operation with different weights.Because the susceptibility of human eye to green is high, low to the susceptibility of blueness, therefore average calculating operation can be weighted according to different weights to RGB tri-components and can obtain more rational gray level image.

Step 2, take the first-selected advanced column hisgram correction of histogram equalization, utilize greyscale transformation function to be modified to the histogram of original image and be uniformly distributed, and then carry out histogram equalization;

Step 3, medium filtering is carried out to image,

Step 4, take to be Global thresholding, in binarization, only use the method for a global threshold T, the gray-scale value of each pixel of image and T compare by it, if be greater than T, are then taken as foreground; Otherwise, be taken as background colour;

Step 5, using the maximum center point P of respective pixel number in L grey level range as initial classes average μ ₁ ⁽¹⁾, μ ₂ ⁽²⁾..., μ _l ^(l).

Step 6, when i-th iteration, investigate each pixel, calculate the spacing between it and the average of each gray level, namely the distance D of it and cluster centre, composes the average class nearest apart from it, namely by each pixel

D|x _p-μ _l ⁽ⁱ⁾|＝min{D|x _p-μ _j ⁽ⁱ⁾|,(j＝1,2,…l)}

D is that two pixel grey scale value differences are less than determining deviation;

X _p(p=0,1 ..., 255) and be the gray-scale value of pixel;

Then it is the pixel set being assigned to class j after i-th iteration;

Step 7, for j=1,2 ... l, calculates new cluster centre, upgrades class average:

${{\mathrm{\μ}}_j}^{(i+1)}=1/N_j\underset{x\∈Q_j^{\left(i\right)}}{\Σ}{x}_p$

In formula, N _jbe in number of pixels;

Step 8, all pixels to be investigated one by one, if j=1,2 ... K, has μ _j ⁽ⁱ⁺¹⁾=μ _j ⁽ⁱ⁾, then algorithm convergence, terminates; Otherwise return step 6 and continue next iteration;

After step 9, above cluster process terminate, each pixel of segmentation result is using cluster centre gray-scale value as such final gray scale.

In technique scheme, medium filtering comprises the following steps:

First a neighborhood of a point centered by certain pixel is determined, then the gray-scale value of pixel each in neighborhood is sorted, get the new value of intermediate value as center pixel gray scale;

After neighborhood moves in the picture up and down, utilize median filtering algorithm to the smoothing process of image.

In sum, owing to have employed technique scheme, the invention has the beneficial effects as follows:

The output image of Mean Filtering Algorithm gets average with pixels all in window according to certain mathematical operations, arithmetic wave filter effectively can remove Gaussian noise and the little salt-pepper noise of intensity, and geometric mean wave filter can retain more image detail relative to arithmetic wave filter.

Adopt the dividing method edge of the application more clear, segmentation result both highlighted target, remained detailed information again, reached good segmentation effect.Therefore, this algorithm can be split gray level image effectively, can obtain more target information from the image after segmentation.

Embodiment

All features disclosed in this instructions, or the step in disclosed all methods or process, except mutually exclusive feature and/or step, all can combine by any way.

Information in piece image comprises target object, background and noise three parts, a kind of image processing method that the binaryzation of image be target object in order to obtain in image and produce, and after binaryzation, in image, all pixels will become white or black.When only comprising prospect and background two parts information in image, just the pixel value of prospect can be set to 1, the pixel value of background is set to 0, and this sampled images is just binarized.The method of binaryzation has a variety of, is generally divided into Global thresholding and local thresholding method

Global thresholding refers to the method only using a global threshold T in binarization.The gray-scale value of each pixel of image and T compare by it, if be greater than T, are then taken as foreground (white); Otherwise, be taken as background colour.Determine a threshold value according to the histogram of text image or gray space distribution, realize the conversion of gray document image to bianry image with this.Wherein Global thresholding can be divided into again based on the threshold method of point and the threshold method based on region.The result of thresholding method depends on the selection to threshold value to a great extent, and therefore the key of the method how to select suitable threshold value.

Typical global threshold method comprises Otsu method, maximum entropy method etc.Global thresholding algorithm is simple, target and background is obviously separated, histogram distribution is that bimodal image effect is good, but for making histogram distribution due to reasons such as uneven illumination are even, noise is larger not in bimodal image, binaryzation successful is deteriorated.

Local smoothing method method is also referred to as neighborhood averaging, and the wave filter be made up of its principle is called mean filter, and this is a kind of typical linear filtering [15].It is a kind of technology of directly carrying out picture smooth treatment in spatial domain.Suppose that image is made up of the fritter that many gray-scale values are constant, very large spatial coherence is there is between neighbor, and noise is self-existent, the gray-scale value that therefore this pixel can be replaced original with the gray level mean value of each pixel in neighborhood, gray-scale value like this between each pixel just all has correlativity, just can remove noise, realize the level and smooth of image.Mean filter is also divided into arithmetic equal value filtering and geometric mean filtering.Arithmetic equal value filtering is as the term suggests be the gray-scale value substituting window center with the arithmetic mean of pixel gray-scale values all in window, this is the simplest mean filter, and the principle of geometric mean filtering is averaged with geometric operation by the gray-scale value of image pixel in window.

Known neighborhood averaging is exactly using the one simple denoising way of the average gray in each for present image neighborhood of pixels as its output valve.

Mean Filtering Algorithm key step is as follows:

Select 3 × 3 windows, pixel value when center pixel is the input of this point, remaining is the value of pixel in its neighborhood.

Then the average of these pixel values is asked, as the pixel value exported, neighborhood averaging ratio juris that Here it is.

In addition for the first row and last column, first row and last row, eight pixels adjacent with oneself can not be found, therefore keep their data constant, finally the gray matrix of the data of these several ranks composition diagram picture together with the data assemblies after those changes.

According to the value output image of this gray matrix.

The image read in MATLAB is all eight, its maximal value only has 255 therefore after nine numbers are added, just to exceed 255, so therefore it can get 255. automatically in im2double is double type by uint8 data type conversion, then after whole data operation is over, calls im2uint8 double data is converted to uint8 and shows.

The output image of Mean Filtering Algorithm gets average with pixels all in window according to certain mathematical operations, arithmetic wave filter effectively can remove Gaussian noise and the little salt-pepper noise of intensity, geometric mean wave filter can retain more image detail relative to arithmetic wave filter, but owing to lacking the consideration to keeping containing object edge in image in algorithmic procedure, comprising sign mutation place to all pixels in image all to carry out level and smooth, use mean filter can cause the annihilation of the fuzzy of edge and details, therefore mean filter also creates bad impact while image denoising, this method also makes the detail section of image become fuzzyyer while level and smooth picture signal, can verify when neighborhood obtains larger image can be fuzzyyer.

Greyscale transformation is basic picture point computing, is the one very basic space area image disposal route in image enhancement processing.Greyscale transformation refers to goes pointwise to change each grey scale pixel value in original image according to certain goal condition according to certain transformation relation, object is to improve image quality, to make the display effect of image better clear, therefore greyscale transformation is also called as the Contrast enhanced of image.Dynamic range of images after greyscale transformation becomes large, and contrast can strengthen, and image can become more clear, and feature is also more obvious.Greyscale transformation mainly utilizes point processing to change the gray-scale value of image slices vegetarian refreshments, does not change the spatial relationship in image, and except converting according to certain specific transforming function transformation function, greyscale transformation can be thought simply to copy pixel.The expression formula of greyscale transformation is:

g(x,y)＝T[f(x,y)](1)

Wherein function T is greyscale transformation function, it defines the conversion condition between input picture gray scale and output image gray scale.If so gamma function determines, so greyscale transformation is just fully defined.The method of greyscale transformation has a variety of, such as gradation of image is negated, gray scale stretching, gray scale cutting, the adjustment of gray scale dynamic range and gray level correction etc.The treatment effect of above several method to image is different, but all must use point processing in their processing procedures.Point processing can be divided into this three major types of linear transformation, piecewise linear transform and nonlinear transformation usually.

(1) linear transformation

Assuming that the intensity value ranges of input picture f (x, y) is [a, b], the intensity value ranges of the output image g (x, y) after conversion extends to [c, d], then for the gray-scale value (x, y) of any point of image, its expression formula is as follows:

$g(x,y)=\frac{d-c}{b-a}\×\[f(x,y)-a\]+c---\left(2\right)$

If the gray level of original image major part pixel is in interval [a, b], maxf is original image gray scale maximal value, only has the gray level of particular not in interval, then in order to improve image enhancement effects, and Ke Yiling:

$g(x,y)=\left\{\begin{array}{cc}c& o\≤f(x,y)\≤a\\ \frac{d-c}{b-a}\×\[f(x,y)-a\]+c& a\≤f(x,y)\≤b\\ d,& b\≤f(x,y)\≤\max f\end{array}\right.---\left(3\right)$

Therefore linear transformation is applicable to those under-exposed or excessive images, and their gray scale may be distributed in a very little scope, and the image at this moment obtained is fuzzyyer, not have a gray-level image.Adopt above-mentioned linear transformation each pixel to image to carry out gray scale and make linear stretch, effectively will strengthen the quality of image.

(2) piecewise linear transform

Piecewise linear transform and linear transformation similar, difference is in order between gray area interested in outstanding image, relatively suppresses, between unwanted gray area, can carry out piecewise linear transform, and it carries out the segmentation of two to multistage to gradation of image interval.When converting, 0-255 gray value interval is divided into several line segment, all corresponding linear transformation function of each line segment.

(3) nonlinear transformation

Nonlinear transformation, as the term suggests be exactly utilize non-linear transform function to convert image, is divided into exponential transform and log-transformation.Exponential transform, just refer between the gray-scale value of output image pixel and input picture gray-scale value it is exponential relationship, its general formulae is:

g(x,y)＝b ^f(x,y)(4)

Log-transformation namely refers to present logarithmic relationship between the gray-scale value of output image pixel and the gray-scale value of input picture, and its general formulae is:

g(x,y)＝lg[f(x,y)](5)

Visible exponential transform will far away higher than the interval of low gray scale for the divergence between high gray area, so index greyscale transformation was generally applicable to bright image.Contrary with exponential transform, log-transformation is comparatively large for divergence between low gray area, so be generally used for processing excessively bright image.

Claims (1)

1., based on a gray space division methods for the Image semantic classification of mean filter, it is characterized in that: comprise the following steps

Step 1, according to the importance of RGB tri-components and other index, three components are weighted average calculating operation with different weights;

Step 2, take the first-selected advanced column hisgram correction of histogram equalization, utilize greyscale transformation function to be modified to the histogram of original image and be uniformly distributed, and then carry out histogram equalization;

Step 3, carry out mean filter to image, select n × n window, pixel value when center pixel is the input of this point, remaining is the value of pixel in its neighborhood, then asks the average of these pixel values, as the pixel value exported;

Step 4, take to be Global thresholding, in binarization, only use the method for a global threshold T, the gray-scale value of each pixel of image and T compare by it, if be greater than T, are then taken as foreground; Otherwise, be taken as background colour;

Step 5, using the maximum center point P of respective pixel number in L grey level range as initial classes average μ ₁ ⁽¹⁾, μ ₂ ⁽²⁾..., μ _l ^(l);

Step 6, when i-th iteration, investigate each pixel, calculate the spacing between it and the average of each gray level, namely the distance D of it and cluster centre, composes the average class nearest apart from it, namely by each pixel

D|x _p-μ _l ⁽ⁱ⁾|＝min{D|x _p-μ _j ⁽ⁱ⁾|,(j＝1,2,…l)}

D is that two pixel grey scale value differences are less than determining deviation;

X _p(p=0,1 ..., 255) and be the gray-scale value of pixel;

Then it is the pixel set being assigned to class j after i-th iteration;

Step 7, for j=1,2 ... l, calculates new cluster centre, upgrades class average:

${{\mathrm{\μ}}_j}^{(i+1)}=1/N_j\underset{x\∈Q_j^{\left(i\right)}}{\Σ}{x}_p,$

In formula, N _jbe in number of pixels;

Step 8, all pixels to be investigated one by one, if j=1,2 ... K, has μ _j ⁽ⁱ⁺¹⁾=μ _j ⁽ⁱ⁾, then algorithm convergence, terminates; Otherwise return step 6 and continue next iteration;

After step 9, above cluster process terminate, each pixel of segmentation result is using cluster centre gray-scale value as such final gray scale.