CN104268836A - Watershed segmentation mark point extraction method based on local area homogeneity indexes - Google Patents

Abstract

The invention discloses a watershed segmentation mark point extraction method based on local area homogeneity indexes. Firstly, an original image is quantized to obtain a quantized image, and a J-image is formed with the J-value of each pixel serving as the pixel value of the pixel; the J-image is defined, the original image is segmented, the texture complexity of each region is judged, and therefore distribution of mark points can be restrained through different threshold values. By means of the method, image detail loss caused by pre-filtering processing is avoided, more mark points are extracted in the regions with the complex texture so as to extract the detail features of the image, fewer mark points are adopted in the regions with the homogeneous texture so as to avoid over-segmentation, and therefore reasonable restraint on mark point distribution is achieved.

Description

A kind of watershed segmentation gauge point extracting method based on local homogeneous index

Technical field

The present invention relates to a kind of watershed segmentation gauge point extracting method based on local homogeneous index, on the watershed segmentation basis of tradition based on gauge point, achieving gauge point extracted in self-adaptive and distribution constraint by introducing local homogeneous index, belonging to Remote Sensing Image Processing Technology field

Background technology

In the last thirty years, remote sensing image, due to the feature such as have multiband, atural object huge number, textural characteristics is abundant, multiple dimensioned, coverage is broad, has been widely used in the every field of the social life such as land resource planning, Natural calamity monitoring.In many correlation techniques that remote sensing image process field relates to, Remote Sensing Image Segmentation achieves the extraction to the object in scene with physical significance, being carry out Images Classification and the basis realizing object-level change detection further, is also one of the study hotspot in remote sensing technology field.

Compared with normal image, the feature of Remote Sensing Image Segmentation mainly comprises: remote sensing image typically includes the data of multiple wave band, makes traditional single-range image division method be difficult to directly apply in multispectral or target in hyperspectral remotely sensed image segmentation; Moreover remote sensing image typically includes abundant texture information, can the spatial structure characteristic of the various atural object complexity of concentrated expression, make effective extract and the textural characteristics of statement object more difficult; Finally, the region that remote sensing image covers a wide range usually, size is large, and the disturbing factors such as cloud layer covering, atural object shade are numerous, therefore needs Image Segmentation more efficiently, sometimes also needs to introduce priori and improves segmentation precision.

In the last thirty years, scholars are to the extensive and deep research of Remote Sensing Image Segmentation and application start thereof, although it is pointed out that at present Remote Sensing Image Segmentation algorithm is a lot, for high-resolution remote sensing image partitioning algorithm research still specific aim and systemic in there is many deficiencies.The raising of spatial resolution brings abundant spectral information and the spatial information such as texture, shape.On the other hand, in the class of atural object of the same race, the inter-class separability of separability increase and variety classes atural object reduces, i.e. ubiquitous " same object different images " and " the different spectrum of jljl " phenomenon, and ground species more various in scene, baroque man-made target with etc. factor all successfully Image Segmentation cause difficulty.

Fractional spins is a kind of dividing method based on region, has the plurality of advantages such as global segmentation, anchored object edge is accurate, keeping object profile is complete, and has been successfully applied to high-resolution remote sensing image segmentation field.However, watershed algorithm usually needs to carry out filtering process to the disturbing factor such as isolated point, noise before Image Segmentation, easily causes the loss of detailed information in image; Gauge point fetch strategy based on single threshold value is difficult to the various atural object with difformity, size and texture complexity degree in accurate marker scene, therefore in segmentation result, easily produces over-segmentation and less divided phenomenon.

List of references

[1]Deng Y,Manjunath B S.Unsupervised segmentation of color-texture regions in images and video[J].Pattern Analysis and Machine Intelligence,IEEE Transactions on,2001,23(8):800-810.

[2] Zhang Bo, what is refined. the application of Watershed Transformation Algorithm in high-resolution remote sensing image multi-scale division [J] of improvement. and Earth Information Science journal, 2014,1 (16): 144-150.

Summary of the invention

Goal of the invention: for problems of the prior art with not enough, the invention provides a kind of watershed segmentation gauge point extracting method based on local homogeneous index, traditional gradient image is substituted by local homogeney index J-value, can the impact of the disturbing factor such as effectively overcoming noise.Meanwhile, propose the gauge point fetch strategy based on J-value, can according to the distribution of the complexity proper restraint gauge point of regional area textural characteristics.

Technical scheme: a kind of watershed segmentation gauge point extracting method based on local homogeneous index, comprises the steps:

First raw video quantized thus obtain quantification image, make the pixel value that position z (x, y) of each pixel in quantification image is pixel z, z (x, y) ∈ Z, Z are the set of all pixel compositions in specific dimensions window centered by pixel z.

In quantification image, definition N is the sum of all pixels centered by z in window, then average m:

$m=\frac{1}{N}\underset{z\∈Z}{\mathrm{\Σ}}z(x,y)---\left(1\right)$

Definition m _pfor belonging to all pixel averages of same grey level p in window, Z _pfor belonging to the set of all pixels of gray level p in window, P is the gray level sum in quantification image, then belong in window the variance of same gray-level pixels and S _wmay be defined as:

$S_W=\underset{p=1}{\overset{P}{\mathrm{\Σ}}}\underset{z\∈Z_p}{\mathrm{\Σ}}{\left|\right|z-m_p\left|\right|}^2---\left(2\right)$

Definition S _tpopulation variance for pixels all in window:

$S_T=\underset{z\∈Z}{\mathrm{\Σ}}{\left|\right|z-m\left|\right|}^2---\left(3\right)$

Then J-value is:

J＝(S _T-S _W)/S _W (4)

Utilize the pixel value of J-value as this pixel of each pixel, namely constitute J-image image.

Raw video is carried out piecemeal process, and respectively the texture complexity degree in every block region is differentiated, thus by the distribution of different threshold value constraint gauge points.

Accompanying drawing explanation

Fig. 1 is the method flow diagram of the embodiment of the present invention;

Fig. 2 is 9 × 9 pixel windows of the embodiment of the present invention;

Fig. 3 is Spot 5 multi-spectrum remote sensing image of the embodiment of the present invention;

Fig. 4 is the border that embodiment of the present invention method is extracted;

Fig. 5 is the border that method two extracts;

Fig. 6 is the border that inventive embodiments method is extracted

Fig. 7 is the border that method two extracts.

Embodiment

Below in conjunction with specific embodiment, illustrate the present invention further, these embodiments should be understood only be not used in for illustration of the present invention and limit the scope of the invention, after having read the present invention, the amendment of those skilled in the art to the various equivalent form of value of the present invention has all fallen within the application's claims limited range.

Based on the watershed segmentation gauge point extracting method of local homogeneous index, mainly comprise the multiple dimensioned J-image image sequence of structure, the gauge point of single yardstick extracts and region segmentation, the superposition of multi-scale division result and region merging technique.Algorithm flow is as shown in Figure 1:

1 local homogeney index calculate

JSEG algorithm is one of current multiple dimensioned color texture partitioning algorithm the most popular.In the cutting procedure of JSEG algorithm, a kind of J-image image can be produced, J-image essence is the gradient image that a width contains dimensional information, spectral distribution during wherein local homogeney index J-value reflects in raw video centered by each pixel specific dimensions regional area, and insensitive to the disturbing factor such as isolated point, noise.Therefore, the present embodiment is selected J-image image to substitute traditional gradient image to carry out gauge point extraction.The computation process of J-value is as follows:

First the quantization method adopting the people such as Deng to propose quantizes raw video thus obtains quantification image ^[1].Make the pixel value that position z (x, y) of each pixel in quantification image is pixel z, z (x, y) ∈ Z.Z is the set of all pixel compositions in the specific dimensions window centered by pixel z.Larger window size can reflect the texture complexity degree of bulk region, and less window size is conducive to the minutia extracting image.As the certain window set centered by pixel z is of a size of 9 × 9 pixels, as shown in Figure 2.Meanwhile, for ensureing the consistance of all directions as far as possible, the angle point in window is removed.

In quantification image, definition N is the sum of all pixels centered by z in window, then average m:

$m=\frac{1}{N}\underset{z\∈Z}{\mathrm{\Σ}}z(x,y)---\left(1\right)$

Definition m _pfor belonging to all pixel averages of same grey level p in window, Z _pfor belonging to the set of all pixels of gray level p in window, P is the gray level sum in quantification image, then belong in window the variance of same gray-level pixels and S _wmay be defined as:

$S_W=\underset{p=1}{\overset{P}{\mathrm{\Σ}}}\underset{z\∈Z_p}{\mathrm{\Σ}}{\left|\right|z-m_p\left|\right|}^2---\left(2\right)$

Definition S _tpopulation variance for pixels all in window:

$S_T=\underset{z\∈Z}{\mathrm{\Σ}}{\left|\right|z-m\left|\right|}^2---\left(3\right)$

Then J-value is:

J＝(S _T-S _W)/S _W (4)

Utilize the pixel value of J-value as this pixel of each pixel, namely constitute J-image image.

2 based on the gauge point extracted in self-adaptive of piecemeal

According to the definition of above J-value, J-value not only can reflect the texture complexity degree of regional area, and the J-value value that a certain pixel is corresponding is simultaneously larger, then more may be in the edge of object, otherwise, then may be positioned at the center of object.Based on this characteristic, raw video is carried out piecemeal process, and respectively the texture complexity degree in every block region is differentiated, thus by the distribution of different threshold value constraint gauge points.Gauge point fetch strategy is as follows:

Step1: first according to the size setup parameter M of image to be split, image is divided into the sub-block of size M × M pixel.Travel through all sub-blocks, calculate the J-value that each sub-block is corresponding, to judge the texture complexity degree of current sub-block inside.Enter Step 2.

Step2: in order to reflect the minutia of image as much as possible, the minimum window of the J-value of seletion calculation is of a size of 5 × 5 pixels, and calculates the value of the J-value that all pixels are corresponding in raw video, thus generates J-image image.

Step3: for a certain sub-block, if meet J ∈ (0.05,0.3], then think that this sub-block is normal areas.Calculate the J-value average of all pixels of J-image, and be defined as threshold value T _avg, in the region that this sub-block is corresponding in J-image, all threshold values are greater than T _avgpoint as gauge point.Otherwise, enter next step.

Step4: for a certain sub-block, if meet J ∈ (0.3,1), then thinks that this sub-block is texture complex region.For avoiding less divided phenomenon, definition threshold value T _min, and demand fulfillment T _min< T _avg, in the region that this sub-block is corresponding in J-image, all threshold values are greater than T _minpoint as gauge point.Otherwise, enter next step.

Step5: for a certain sub-block, if meet J ∈ (0,0.5), then thinks that this sub-block is average region.For avoiding over-segmentation phenomenon, definition threshold value T _max, and demand fulfillment T _max> T _avg, in the region that this sub-block is corresponding in J-image, all threshold values are greater than T _maxpoint as gauge point.Enter next step.

Step6: repeat Step3 ~ Step5, travel through all sub-blocks, the gauge point realized in J-image image extracts.

3 experimental results and analysis

For the performance of checking the present embodiment method, select the high-resolution remote sensing image partitioning algorithm based on improvement watershed transform that the people such as Zhang Bo propose herein ^[2](hereinafter referred method two) compares experiment.First the method adopts anisotropic filtering technology to carry out denoising and enhancing to image, and then adopt Morphological Gradient Extraction of Image gauge point, again by achieving image multi-scale division based on the watershed transform of gauge point, finally adopt the heterogeneous and shape heterogeneity index of spectrum to carry out region merging technique to obtain final segmentation result and achieve, good effect.However, still there is many deficiencies in the method: the Image semantic classification adopted is difficult to the loss avoiding image detail; Single threshold interval setting cannot retrain gauge point distribution; The region merging technique strategy adopted is difficult to " same object different images " phenomenon etc. accurately distinguishing adjacent area.

Experiment one adopts Chinese Shanghai area high resolving power SPOT 5 multispectral image, spatial resolution I 2.5 meters, be of a size of 256 × 256 pixels, as shown in Figure 3.

Through the T that the J-value average calculating all pixels of J-image obtains _avgvalue is 0.45, therefore chooses optimal value by many experiments, T in setting the present embodiment method _min=0.28, T _max=0.56.According to coefficient of diffusion in document [2] establishing method two the object bounds that the present embodiment method and method two extract and with raw video stack result as also shown in e.g. figs. 4-7:

Data set 1 is typical City scenarios as seen in Figure 3, atural object wide variety and textural characteristics is complicated, and these cause difficulty all to successful Image Segmentation.Can be found out by comparison diagram 6 and Fig. 7.The present embodiment method due to have employed based on local homogeneous index J-value gauge point extract and watershed segmentation, the edge of anchored object is accurate.Simultaneously owing to adopting gauge point distribution constraint strategy, therefore in the more image detail features of extracted region of texture complexity.Finally, because the present embodiment method have employed the region merging technique strategy based on color table standard difference, therefore substantially do not occur to merge phenomenon in segmentation result by mistake, but still there are some over-segmentation phenomenon in homogenous area such as roads.Method two is owing to having carried out filtering process, and therefore the border of anchored object exists obvious error.Meanwhile, be all based on identical threshold value because global mark point extracts, therefore there is obvious less divided phenomenon in the region of texture complexity.On the other hand, method two adopts the region merging technique strategy carried out based on gray feature, and the adjacent atural object making to have in segmentation result similar spectral feature is identified as same atural object by mistake.

Watershed segmentation gauge point extracting method based on local homogeneous index of the present invention, avoid the loss that pre-filtering process causes image detail, by extracting more gauge point at texture complex region to extract the minutia of image, and adopt a small amount of gauge point to avoid over-segmentation phenomenon in texture homogenous area, thus achieve the proper restraint to gauge point distribution.By testing the high-resolution remote sensing image of different sensors type, different resolution, demonstrate feasibility and the validity of proposed method.By comparing proof with existing algorithm further, target edges of the present invention is more accurate, more responsive to the minutia in scene.

Claims (2)

1., based on a watershed segmentation gauge point extracting method for local homogeneous index, it is characterized in that, comprise the steps:

First raw video quantized thus obtain quantification image, make the pixel value that position z (x, y) of each pixel in quantification image is pixel z, z (x, y) ∈ Z, Z are the set of all pixel compositions in specific dimensions window centered by pixel z;

In quantification image, definition N is the sum of all pixels centered by z in window, then average m:

$m=\frac{1}{N}\underset{z\&Element;Z}{\mathrm{\&Sigma;}}z(x,y)---\left(1\right)$

Definition m _pfor belonging to all pixel averages of same grey level p in window, Z _pfor belonging to the set of all pixels of gray level p in window, P is the gray level sum in quantification image, then belong in window the variance of same gray-level pixels and S _wmay be defined as:

$S_W=\underset{p=1}{\overset{P}{\mathrm{\&Sigma;}}}\underset{z\&Element;Z_p}{\mathrm{\&Sigma;}}{\left|\right|z-m_p\left|\right|}^2---\left(2\right)$

Definition S _tpopulation variance for pixels all in window:

$S_T=\underset{z\&Element;Z}{\mathrm{\&Sigma;}}{\left|\right|z-m\left|\right|}^2---\left(3\right)$

Then J-value is:

J＝(S _T-S _W)/S _W (4)

Utilize the pixel value of J-value as this pixel of each pixel, namely constitute J-image image;

Raw video is carried out piecemeal process, and respectively the texture complexity degree in every block region is differentiated, thus by the distribution of different threshold value constraint gauge points.

2., as claimed in claim 1 based on the watershed segmentation gauge point extracting method of local homogeneous index, it is characterized in that:

Gauge point fetch strategy is as follows:

Step1: first according to the size setup parameter M of image to be split, image is divided into the sub-block of size M × M pixel; Travel through all sub-blocks, calculate the J-value that each sub-block is corresponding, to judge the texture complexity degree of current sub-block inside; Enter Step 2;

Step2: in order to reflect the minutia of image as much as possible, the minimum window of the J-value of seletion calculation is of a size of 5 × 5 pixels, and calculates the value of the J-value that all pixels are corresponding in raw video, thus generates J-image image;

Step3: for a certain sub-block, if meet J ∈ (0.05,0.3], then think that this sub-block is normal areas; Calculate the J-value average of all pixels of J-image, and be defined as threshold value T _avg, in the region that sub-block is corresponding in J-image, all threshold values are greater than T _minpoint as gauge point; Otherwise, enter next step;

Step4: for a certain sub-block, if meet J ∈ (0.3,1), then thinks that this sub-block is texture complex region, needs more gauge point to extract the minutia of image, to avoid less divided phenomenon, and definition threshold value T _min, and meet T _min< T _avg, in the region that subimage is corresponding in J-image, all threshold values are greater than T _minpoint as gauge point; Otherwise, enter next step;

Step5: for a certain sub-block, if meet J ∈ (0.0.5), then thinks that this sub-block is average region, only needs less gauge point, to avoid over-segmentation phenomenon, and definition threshold value T _max, and meet T _max> T _avg, in the region that subimage is corresponding in J-image, all threshold values are greater than T _maxpoint as gauge point; Enter next step;

Step6: repeat Step3 ~ Step5, travel through all sub-blocks, the gauge point realized in J-image image extracts.