CN101894372A - New noise-containing remote sensing image segmentation method - Google Patents

Abstract

The invention provides a noise-containing remote sensing image segmentation algorithm which is based on wavelet packet analysis and is combined with improved 2D otsu, aiming at improving noise-containing image segmentation accuracy. The algorithm includes that firstly wavelet packet is used for decomposing a noise-containing remote sensing image, then adaptive threshold method is adopted to remove noise in the image, and then the two-dimensional histogram of the image is projected onto diagonal, so as to form a one-dimensional histogram with obvious 'double peak' distribution, and finally 2D otsu method is applied for searching optimal segmentation threshold on a line vertical to the diagonal, so as to realize image segmentation. The invention not only effectively removes interference of noise but also obtains better segmentation effect.

Description

The new method that noisy remote sensing images are cut apart

Technical field

The present invention relates to a kind of improved noisy remote sensing images 2D Otsu dividing method, belong to the remote sensing technology field.

Background technology

Image segmentation is one of very important content of image processing field, is the basis of realizing graphical analysis and understanding.Image segmentation can be understood as significant characteristic area in the image or needs the characteristic area of application to extract, and it can isolate target object, identifies the background with similar features parameter simultaneously.A large amount of Research of Image Segmentation achievements is arranged at present, in all these cutting techniques, threshold segmentation method is used widely because of algorithm is simple, applied widely, and most of threshold segmentation method is to choose optimal threshold by the grey level histogram of image, such as entropy threshold value and Otsu method or the like.

The Otsu method is based on the grey level histogram of image, be optimal threshold to the maximum with the inter-class variance of target and background and choose criterion, so claim maximum variance between clusters again, in the quality of image better or under the bigger situation of background and target difference, one dimension Otsu algorithm can be obtained gratifying segmentation effect, but when noise in the image is strong, because this method has only been considered pixel self gray scale, and do not have mutual relationship on the considered pixel space, therefore can not obtain desirable segmentation effect.At the relatively poor shortcoming of one dimension Otsu algorithm noise immunity, " the The automatic thresholding of gray-level picture via 2D Otsu method " that people such as J.Z.Liu delivered in 1933 represents the two dimensional gray histogram with the bivector of pixel grey scale and the combination of neighborhood average, two-dimentional Otsu algorithm has been proposed, this algorithm has improved the image segmentation effect greatly, particularly under the situation of low signal-to-noise ratio, still this method travelling speed is very slow.People such as L.J.Dong have proposed a kind of rapid image dividing method based on two-dimensional histogram (referring to document " The automatic thresholding of gray-level picture via 2D Otsu method " in 2005, Joumal of Northeastern University, 2005,26 (3): 220-223), but this method can not obtain the optimal segmentation effect for noisy image.People such as Zhang have proposed a kind of 2D Otsu histogram analysis based on wavelet transformation and have been used for image segmentation (referring to document " Image segmentation based on 2D Otsu method with histogram analysis " in 2008,2008 International Conference on Computer Science and Software Engineering), overcome cutting apart the problem of object magnitude.People such as Zhu have proposed a kind of based on improving histogrammic quick 2D Otsu thresholding algorithm (referring to document " A fast 2D otsu thresholding algorithm based on improved Histogram " in 2009, Pattern Recognition, CCPR 2009, Chinese Conference), though improved threshold search speed greatly, segmentation effect is unsatisfactory.People such as D.Y.Huang have proposed many threshold values of two-stage Otsu method (referring to document " Optimal multi-level thresholding using a two-stage Otsu optimization approach " in 2009, Pattern Recognit.Lett, vol.30, pp.275-284, Feb.2009), improve the efficient of Otsu method greatly, simultaneously different test patterns has been had less difference working time.

But 2D Otsu method can not obtain good segmentation effect sometimes, and especially to having comprised the remote sensing images of more noises, therefore, this paper has proposed the new algorithm that a kind of noisy remote sensing images are cut apart.

Summary of the invention

Technical matters: the objective of the invention is to propose the new algorithm that a kind of noisy remote sensing images are cut apart, this method is the Y-PSNR height not only, has good denoising effect, and has produced excellent segmentation performance.

Technical scheme: purpose of the present invention can realize by following scheme:

At first utilize the high-resolution characteristic of wavelet packet Space Time that image is decomposed, compare with wavelet analysis, wavelet packet analysis decomposes the low frequency and the HFS of image simultaneously, and, it can provide optimal base according to the feature of signal, for the data compression in the signal Processing and denoising provide one to decompose preferably and the reconstruct approach.In Texture classification, the orthogonality of basis function and tight property are necessary, therefore, this paper selects the db3 wavelet function in the Daubechies orthogonal wavelet family of functions that image is carried out WAVELET PACKET DECOMPOSITION, again because in picture breakdown, decomposed class is directly proportional with the boundary effect of image, and experiment shows, texture image through three grades of wavelet package transforms after, the details of its low frequency sub-band is smoothed basically, when decomposing the fourth stage, the energy of sub-band images is very little, therefore will select the db3 wavelet function in the Daubechies orthogonal wavelet family of functions that image is carried out three grades of WAVELET PACKET DECOMPOSITION.

After the WAVELET PACKET DECOMPOSITION, to adopt adaptive threshold to carry out denoising, compare with single threshold method, the adaptive threshold method can more effectively be removed noise and keep useful signal, it is according to the different characteristics of wavelet packet coefficient amplitude on every grade of yardstick, be considered as on every grade of separate yardstick wavelet packet coefficient being carried out threshold process respectively, seek a threshold value of mating most with it and carry out denoising, therefore, this paper uses SureShrink adaptive threshold denoise algorithm, it has adopted Stein unbiased error method of estimation, has kept the details of image when making the square error minimum and has eliminated noise.

Utilizing improved 2D Otsu method to make search space at last is one dimension by Simplified two-dimension, on the one dimension histogram, utilize the Otsu threshold method to obtain cut-point, point with the finger or gesticulate one perpendicular to cornerwise straight line by cutting apart, so just be easy to be partitioned into target and background, realize image segmentation on the both sides of straight line.

Beneficial effect: the present invention has the following advantages:

1) utilize the db3 wavelet function in the Daubechies orthogonal wavelet family of functions that image is carried out three grades of WAVELET PACKET DECOMPOSITION, and after using SureShrink adaptive threshold denoise algorithm that image is carried out denoising, compare with original noisy image, obviously known a lot, and kept most of details, eliminated the noise that comprises in the image simultaneously.

2) utilize improved 2D Otsu algorithm, make threshold search by original two-dimensional transform to one dimension, reduced calculated amount.

3) be used for image segmentation in conjunction with wavelet packet and improved 2D Otsu algorithm, obtained desirable segmentation effect.

Description of drawings

Accompanying drawing 1 expression two-dimensional histogram projects to the top view on the diagonal line, Fig. 2 is illustrated in and improves the top view of searching optimal segmenting threshold on the histogram, Fig. 3 represents original image, Fig. 4 represents that former gray level image adds the image after making an uproar, Fig. 5 represents the image after the wavelet packet denoising, Fig. 6 represents the result that former 2D Otsu algorithm is cut apart, Fig. 7 represents the result that document " A fast 2D otsu thresholding algorithm based on improved Histogram " is cut apart, and Fig. 8 represents to use the result that this paper algorithm is cut apart.

Embodiment

Below embodiments of the invention are elaborated: present embodiment is being to implement under the prerequisite with the technical solution of the present invention, provided detailed embodiment and concrete operating process, but protection scope of the present invention is not limited to following embodiment.

Present embodiment selects the db3 wavelet function in the Daubechies orthogonal wavelet family of functions that image is carried out three grades of WAVELET PACKET DECOMPOSITION at first according to the feature of noisy remote sensing images, uses SureShrink adaptive threshold denoise algorithm then:

$\mathrm{SURE}(\hat{T},Y)=N+\underset{i,j=1}{\overset{N}{\mathrm{\Σ}}}{\left[\min (\left|Y\right|,\hat{T})\right]}^2-2[1:N]---\left(1\right)$

$T=\arg \min \left(\mathrm{SURE}(\hat{T},Y)\right)---\left(2\right)$

Former 2D Otsu algorithm is represented two-dimensional histogram with the bivector of pixel grey scale and the combination of neighborhood average, adopt " square partitioning " to reach the selection of threshold strategy of " orthogonal straight lines to ": given two-dimentional threshold vector [s, t], with pair of orthogonal straight line f=s, g=t is divided into four rectangular region with two-dimensional histogram, near 45 ° of area 0 and zones 1 of diagonal line correspond respectively to background and target, away from cornerwise regional 2 and the zone 3 corresponding to a spot of strong edge and very noisy, but this method makes the threshold search scope rise to two dimension from one dimension, calculated amount increases fast with index, and can not obtain desirable segmentation effect sometimes.

If (x, gray level y) is L to image f, at point (x ₀, y ₀) gray-scale value located is f (x ₀, y ₀), the average gray value of n * n neighborhood is g (x around this point ₀, y ₀), establishing the number of pixels that satisfies gray scale f=i and average g=j is that (i, j), (x, gray level y) also is L to average as can be known neighborhood gray-scale map g to h.Fig. 1 represents that two-dimensional histogram projects to the top view on the diagonal line, and Fig. 2 is illustrated in and improves the top view (shown in description of drawings) of searching optimal segmenting threshold on the histogram.

Cut apart as shown in Figure 1 cornerwise length P and cornerwise discrete value x ∈ P (0,1,2 ..., P-1) be respectively:

$P=\left[\sqrt{2}(L-1)\right]---\left(3\right)$

$x=\left[\frac{\sqrt{2}}{2}(i+j)\right]---\left(4\right)$

Wherein, and i ∈ L (0,1,2 ..., L-1), j ∈ L (0,1,2 ..., L-1) represent the gray level of original image and average neighborhood gray level image respectively, [] expression bracket function.

When a some Q (i j) projects to some X on the diagonal line, h (i j) also projects on the diagonal line, make H (x) for h (i, the j) projection value on the diagonal line, then:

$H\left(x\right)=\underset{x=[\sqrt{2}(i+j)/2]}{\mathrm{\Σ}}h(i,j)---\left(5\right)$

Therefore, so just obtained having the obviously one dimension histogram of " bimodal " distribution, can on the one dimension histogram, utilize the Otsu threshold method to obtain cut-point K (as shown in Figure 2) then, draw one perpendicular to cornerwise straight line P by a K, so just be easy to be partitioned into target and background on the both sides of straight line P, threshold search to one dimension, has reduced calculated amount by original two-dimensional transform, the equation that straight line P can be expressed as:

g(x，y)＝-f(x，y)+2i i∈L(0，1，2，...，L-1) (6)

Below further specify and originally practice situation:

Experiment will be adopted the real scene shooting remote sensing image, because its size is bigger, can't completely show in experiment, so what intercept in the literary composition be 512 * 512 image that experimental situation is: matlabR2008a, PC processor are T6600, in save as 2GB.When WAVELET PACKET DECOMPOSITION, select the db3 wavelet function in the Daubechies standard orthogonal wavelet family of functions that image is carried out 3 grades of WAVELET PACKET DECOMPOSITION, the interpolation variance is 0.005 white Gaussian noise in the image, and the neighborhood size is got n=3, and the effect of denoising is weighed with Y-PSNR (PSNR).

Test the validity of verifying this paper algorithm by the algorithm of former 2D Otsu partitioning algorithm of contrast and document " A fast 2D otsu thresholding algorithm based on improved Histogram ", the Fig. 3 in experimental result such as the accompanying drawing is to shown in Figure 8.

According to the result that each algorithm is cut apart, can obtain the performance evaluation result of each Otsu algorithm, as shown in table 1 below:

Each Otsu algorithm performance evaluation result of table 1

	Threshold	Time/s	Psnr/dB	Effect
					2D?Otsu	(106，117)	18.354	36.148	bad
Method?in[7]	(103，106)	0.6346	31.661	bad
					proposed	(107，102)	5.4349	41.676	better

Can find out significantly that by above result the effect that this paper algorithm cuts apart is well more a lot of than former 2D Otsu partitioning algorithm and document " A fast 2D otsu thresholding algorithm based on improved Histogram " algorithm effects.Not only the isolated wrong branch of background is many for former 2D Otsu partitioning algorithm, and sliced time is longer, document " Afast 2D otsu thresholding algorithm based on improved Histogram " though the partitioning algorithm time shorter, but the assorted point of the white non-target in the black background is a lot, and effect is unsatisfactory.By contrast, though the algorithm that this paper proposes is longer than document " A fast 2D otsu thresholding algorithm based on improved Histogram " working time, can find out obviously that from above segmentation result segmentation effect is best, except the white object zone is more complete, black background is also very clean, the assorted point of the white that does not almost have mistake to divide, and, the Y-PSNR of this paper algorithm is also very high, has good denoising effect.This shows that this paper algorithm has produced excellent segmentation performance, is a kind of effective partitioning algorithm.

Claims (3)

1. the new method of cutting apart according to the described noisy remote sensing images of claim, it is characterized in that at first selecting the db3 wavelet function in the Daubechies orthogonal wavelet family of functions that image is carried out three grades of WAVELET PACKET DECOMPOSITION, adopt the method for SureShrink adaptive threshold to remove noise in the image then.

2. the new method of cutting apart according to the described noisy remote sensing images of claim is characterized in that after the wavelet packet denoising, the 2D Otsu dividing method of application enhancements.

3. the new method of cutting apart according to the described noisy remote sensing images of claim, it is characterized in that improved 2D Otsu dividing method at first projects to the two-dimensional histogram of image on the diagonal line, be formed with the obviously one dimension histogram of " bimodal " distribution, use 2D otsu method then and on perpendicular to cornerwise straight line, seek optimal segmenting threshold realization image segmentation.

Images