CN109410223A - A kind of SAR image segmentation method based on watershed algorithm and dictionary learning - Google Patents

Abstract

The invention discloses a SAR image segmentation method based on a watershed algorithm and dictionary learning. The method includes: using a denoising algorithm to preprocess an original image to make the image smoother, and extracting three-layer wavelet features at the same time; using K-means The algorithm performs the initial segmentation of the original image; the morphological gradient image of the training image is obtained, the floating-point moving image is calculated, and then the watershed algorithm is used for segmentation; the data in each small area is regarded as a data set of a type of data, and the set is All the data in the nearest neighbors are calculated with the atoms in each type of dictionary, corresponding to each type of dictionary, we can get a statistical histogram, calculate the histogram corresponding to each type and the histogram of the type obtained from the original training data The error between the graphs; to determine which category has the smallest error, divide the data in the set into which category, and divide all the small areas in turn to obtain the final segmentation result. Compared with the traditional SAR image segmentation method, the present invention has higher effectiveness and accuracy in the identification process and lower algorithm complexity.

Description

A kind of SAR image segmentation method based on watershed algorithm and dictionary learning

Technical field

The invention belongs to synthetic aperture radar (SAR) image application fields, are related to a kind of SAR image segmentation method, especially It is a kind of SAR image segmentation method based on watershed algorithm and dictionary learning.

Background technique

Since SAR imaging technique has the advantage round-the-clock, the factors such as climate do not influence so that it in national product and Very big effect has been played in construction.And also just become the hot spot and emphasis of people's research to the processing of SAR image, wherein SAR image segmentation is as to the fundamental operation during image procossing, the also concern by more and more people.Spectral clustering Algorithm is common a kind of algorithm in processing image segmentation problem, but since the algorithm is difficult to handle the data of magnanimity, so its Using being restricted.

Dividing method based on mathematical morphology is method important always in the field of image segmentation.Based on Mathematical Morphology A kind of mathematical tool of the method based on image aspects, wherein more classical algorithm, such as watershed algorithm (watershed algorithm).Watershed algorithm has many good qualities, his scale is intuitive, and the width of cut-off rule is single pixel And closure is continuous, so being paid attention in recent years by many people, the method for dictionary learning is successfully applied to many fields, such as: Classification, segmentation, identification, super-resolution, denoising etc..

It is that everybody is common using the pixel in image as the method that a sample is handled in image segmentation problem Method, and traditional clustering method has spectral clustering etc., but the algorithm is difficult to that a large amount of data are effectively treated.To understand The problem of certainly above-mentioned SAR image is divided, the invention proposes the SAR image segmentation sides based on watershed algorithm and dictionary learning Method.Initial segmentation is carried out to image first, obtains the training sample of some tape labels, and the data of every a kind of tape label, benefit It obtains a statistic histogram under such obtained dictionary with arest neighbors, counts in dictionary atomic distance training data most Close number.Then in order to divide an image into some pockets, make over-segmentation of watershed, then by inside with picture Vegetarian refreshments is that unit is handled, and by using the method for statistic histogram, is sorted out, obtains final segmentation result.

Summary of the invention

In view of above-mentioned deficiencies of the prior art, the technical problem to be solved by the present invention is to how be directed to high resolution SAR Image carries out the segmentation identification of efficiently and accurately using the principle of watershed algorithm and dictionary learning.

To achieve the above object, the present invention provides the SAR image segmentation methods of a kind of watershed algorithm and dictionary learning. Its feature includes:

(1) SAR image pre-processes: since there are a large amount of speckle noises for SAR image itself, therefore first with Denoising Algorithm pair Original image is pre-processed, and so that image is become more smooth, while extracting three layers of wavelet character；

(2) initial partitioning is carried out to original image using K-means algorithm: selects some lean on from the segmentation result of every one kind The data of such nearly cluster centre；Every one kind data obtain such dictionary, the instruction of every one kind using K-means algorithm The arest neighbors for practicing data calculating and the category dictionary atom, obtains a statistic histogram, and every one kind data all obtain a histogram Figure；

(3) initial partitioning is carried out using watershed algorithm: to treated in (2) training image, finding out its morphocline Image calculates floating point activities image, then is split with watershed algorithm；

(4) region division: to the image after segmentation in (3), the data in each pocket regard a kind of data as Data acquisition system corresponds to every a kind of word by all data in the set respectively with the atom computing arest neighbors in each category dictionary We can obtain a statistic histogram to allusion quotation, calculate and be somebody's turn to do corresponding to what the histogram and original training data of every one kind obtained Error between the histogram of class；Judge that the error on which kind of is minimum, which kind of the data in set are just divided into, according to It is secondary to divide all pockets, obtain final segmentation result.

Further, in described (1) using open form state reconstruction filter remove in image some inessential details and Small noise.When calculating wavelet character, three layers of Stationary Wavelet Transform are carried out to original image, obtain coefficient matrix coef_m1(i₁, j₁), m1=1,10, as m1=1, coef_m1(i₁,j₁) represent low frequency coefficient；As m1 > 1, coef_m1(i₁,j₁) represent height Frequency coefficient.10 dimension sub-belt energy feature e (i, j)=[e are extracted to each pixel₁(i,j),...,e₁₀(i,j)]^T, as this The wavelet character of pixel:

Wherein, w × w is the size of sliding window, coef_m1(i₁,j₁) it is i-th in stationary wavelet subband₁Row jth₁Column correspond to Coefficient value.

The classification number for needing to cluster in (2) is K, is selected in the segmentation result of every one kind some in such cluster The data of the heart；Every one kind data obtain such dictionary using K-means algorithm, the training data of every one kind calculate and The arest neighbors of the category dictionary atom, obtains a statistic histogram, and every one kind data all obtain a histogram；

Morphometric characters subtract corrosion transformation by dilation transformation and obtain in (3):

Wherein, b is structural element, and Θ indicates erosion operation,Dilation operation is indicated, if having grey scale change in image very Big pocket, then the value of structural element will suitably reduce, and otherwise may be filtered.Floating point activities image definition are as follows:

Fimg (f)=grad (f) * grad (f)/255.0

The pocket for generating watershed algorithm over-segmentation in (4) acquires statistics with the training sample of tape label Histogram seeks residual error by the statistic histogram as dictionary, to determine which kind of the region belongs to.Scheme with traditional SAR As dividing method is compared, validity and accuracy of the present invention in identification process are higher, and algorithm complexity is lower.

The technical effect of design of the invention, concrete scheme and generation will be described further below, with fully Solve the purpose of the present invention, feature and effect.

Specific embodiment

A specific embodiment of the invention addressed below

(1) remove some inessential details and small noise in image using open form state reconstruction filter.It is counting When calculating wavelet character, three layers of Stationary Wavelet Transform are carried out to original image, obtain coefficient matrix coef_m1(i₁,j₁), m1=1,10, when When m1=1, coef_m1(i₁,j₁) represent low frequency coefficient；As m1 > 1, coef_m1(i₁,j₁) represent high frequency coefficient.To each pixel Point extracts 10 dimension sub-belt energy feature e (i, j)=[e₁(i,j),...,e₁₀(i,j)]^T, wavelet character as the pixel:

Wherein, w × w is the size of sliding window, coef_m1(i₁,j₁) it is i-th in stationary wavelet subband₁Row jth₁Column correspond to Coefficient value.

(2) initial partitioning is carried out to original image using K-means algorithm: selects some lean on from the segmentation result of every one kind The data of such nearly cluster centre；Every one kind data obtain such dictionary, the instruction of every one kind using K-means algorithm The arest neighbors for practicing data calculating and the category dictionary atom, obtains a statistic histogram, and every one kind data all obtain a histogram Figure；

(3) corrosion transformation is subtracted by dilation transformation and obtains morphometric characters:

Wherein, b is structural element, and Θ indicates erosion operation,Dilation operation is indicated, if having grey scale change in image very Big pocket, then the value of structural element will suitably reduce, and otherwise may be filtered.Floating point activities image definition are as follows:

Fimg (f)=grad (f) * grad (f)/255.0

(4) data in each pocket regard the data acquisition system of a kind of data as, by all data in the set Respectively with the atom computing arest neighbors in each category dictionary, corresponding to each category dictionary, we can obtain a statistics histogram Figure calculates the error between such histogram that the histogram for corresponding to every one kind and original training data obtain；Judge Error on which kind of is minimum, which kind of the data in set are just divided into, successively divides all pockets, Obtain final segmentation result.

The preferred embodiment of the present invention has been described in detail above.It should be appreciated that the ordinary skill of this field is without wound The property made labour, which according to the present invention can conceive, makes many modifications and variations.Therefore, all technician in the art Pass through the available technology of logical analysis, reasoning, or a limited experiment on the basis of existing technology under this invention's idea Scheme, all should be within the scope of protection determined by the claims.

Claims (4)

1. a SAR image segmentation method of watershed algorithm and dictionary learning, is characterized in that, comprises: Step (1) SAR image preprocessing: Since the SAR image itself has a large amount of speckle noise, the original image is preprocessed by a denoising algorithm to make the image smoother, and three-layer wavelet features are extracted at the same time; Step (2) Use the K-means algorithm to segment the original image for the first time: select some data close to the cluster center of this type from the segmentation results of each type; use the K-means algorithm for each type of data to obtain a Dictionary, each type of training data calculates the nearest neighbor to the dictionary atom of this type, and gets a statistical histogram, and each type of data gets a histogram; Step (3) uses watershed algorithm to perform initial segmentation: for the training image processed in (2), obtain its morphological gradient image, calculate floating-point moving image, and then use watershed algorithm to segment; Step (4) Region division: For the image segmented in (3), the data in each small area is regarded as a data set of a type of data, and all the data in the set are respectively associated with the atoms in each type of dictionary. Calculate the nearest neighbor, corresponding to each type of dictionary, we can get a statistical histogram, calculate the error between the histogram corresponding to each type and the histogram of the type obtained from the original training data; determine which type of If the error is the smallest, the data in the set is divided into which category, and all the small areas are divided in turn to obtain the final segmentation result. 2. utilize open morphological reconstruction filter to remove some insignificant details and tiny noise in the image as claimed in claim 1, when calculating wavelet feature, carry out three-layer stationary wavelet transform to original image, obtain coefficient matrix coef _m 1 (i ₁ ,j ₁ ),m1=1,,10, when m1=1, coef _m1 (i ₁ , j ₁ ) represents low-frequency coefficients; when m1>1, coef _m1 (i ₁ , j ₁ ) represents high-frequency coefficients , extract the 10-dimensional sub-band energy feature e(i,j)=[e ₁ (i,j),...,e ₁₀ (i,j)] ^T for each pixel, as the wavelet feature of the pixel : Among them, w×w is the size of the sliding window, and coef _m1 (i ₁ , j ₁ ) is the coefficient value corresponding to the i ₁ row and the j ₁ column in the stationary wavelet subband. 3. The number of categories that need to be clustered in claim 1 is K, and some data close to the cluster center of this class are selected in the segmentation result of each class; each class of data uses the K-means algorithm to obtain a dictionary of this class , each type of training data calculates the nearest neighbor to the dictionary atom of this type, and obtains a statistical histogram, and each type of data obtains a histogram; In the step (3), the morphological gradient image is obtained by subtracting the erosion transformation from the dilation transformation: where b is a structuring element, Θ is an erosion operation, Represents the expansion operation. If there are small areas with large grayscale changes in the image, the value of the structural element should be appropriately reduced, otherwise it may be filtered out. The floating-point active image is defined as: fimg(f)=grad(f)*grad(f)/255.0 4. As claimed in claim 1, the small area generated by the watershed algorithm is over-segmented, and a statistical histogram is obtained with the labeled training sample, and the residual is obtained by using the statistical histogram as a dictionary to determine which area this area belongs to. kind.