CN106340005A - High-resolution remote sensing image unsupervised segmentation method based on scale parameter automatic optimization - Google Patents

Abstract

The invention discloses a non-supervised segmentation method of high-resolution remote sensing images based on automatic optimization of scale parameters, which mainly includes three steps: 1) adaptive SP selection based on J-value of local homogeneity index; 2) based on scale 3) Image segmentation based on object boundary constraint strategy; 3) Region merging based on multi-features. Through experiments on multiple sets of high-resolution remote sensing images of different sensor types, and comparing with well-known commercial software eCognition and traditional supervised segmentation methods, it is proved that the proposed method is more accurate in locating the edge of the object, extracting the outline of the object is more complete, and segmentation The process requires no human intervention and is a general and effective unsupervised solution.

Description

An Unsupervised Segmentation Method for High-Resolution Remote Sensing Images Based on Automatic Optimization of Scale Parameters

technical field

The invention relates to a non-supervised segmentation method of high-resolution remote sensing images based on automatic optimization of scale parameters, and belongs to the technical field of remote sensing image segmentation.

Background technique

In recent years, OBIA (object-based image analysis) is getting more and more attention in the field of GIScience (geographic information science) and remote sensing (especially the application of high-resolution remote sensing images). Image segmentation is one of the core steps in OBIA, which realizes the extraction of contour information of geographical objects in the scene, and is the basis and premise of subsequent feature extraction and target recognition. Compared with ordinary images, remote sensing images have many characteristics such as wide coverage, multi-band, and multi-spatial resolution, and contain more types of ground objects. Therefore, traditional segmentation methods are difficult to directly apply to remote sensing images. At the same time, with the continuous improvement of the spatial resolution of remote sensing satellites, such as SPOT 5, IKONOS, QuickBird, etc., representing meter-level and sub-meter-level high-resolution data, have been widely used in crop yield surveys, urban land planning, and disaster monitoring. Therefore, image segmentation technology for high-resolution remote sensing images has become one of the research hotspots in the field of remote sensing.

Compared with medium- and low-resolution remote sensing images, high-resolution remote sensing images bring richer spectral and texture features, and spatial details such as the structure and shape of objects can be clearly expressed. It is effectively solved, thus improving the separability between classes of adjacent ground objects. On the other hand, the improvement of spatial resolution has also brought new difficulties and challenges to image segmentation: the phenomenon of "same object but different spectrum" is more prominent, that is, the spectral characteristics of the same type of ground objects may have significant differences; At the same time, the influence of interference factors such as object shadows, noise, and cloud cover is more significant; the changing ecological environment, rich types of land objects, and man-made objects with complex structures in urban scenes have caused difficulties in accurately extracting geographical objects.

To this end, scholars have proposed some solutions, one of the most important means is to introduce a multi-scale segmentation strategy, so as to better reveal the spatial structure characteristics of objects at different scales. For example, C.Burnet et al. proposed a fractal-based multi-scale segmentation algorithm, which achieved good results by estimating the homogeneity and heterogeneity of spectral features in local areas and performing iterative optimization ^[1] ; the well-known remote sensing commercial software eCognition The fractal net evolution algorithm (FNEA) is used for multi-scale segmentation, which makes full use of the object's spectrum, texture, shape, hierarchy and inter-class information. It should be pointed out that the multi-scale segmentation methods in the prior art all need to determine the scale parameter (scale parameter, SP) through manual interpretation or trial and error, and these methods cannot be called automatic image segmentation. At present, there are not many multi-scale segmentation solutions with adaptive SP parameters, which has become one of the main bottlenecks restricting the wide application of OBIA technology.

references

[1] Burnett C, Blaschke T.A multi-scale segmentation/object relationship modeling methodology for landscape analysis[J].Ecologicalmodelling,2003,168(3):233-249.

[2] Shao P, Yang G, Niu X, et al. Information Extraction of High-Resolution Remotely Sensed Image Based on Multiresolution Segmentation [J]. Sustainability, 2014, 6(8): 5300-5310.

[3] 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.

[4]Baraldi A, Boschetti L. Operational automatic remote sensing image understanding systems: Beyond Geographic Object-Based and Object-Oriented Image Analysis (GEOBIA/GEOOIA). Part 2: Novel system architecture, information/knowledge representation, algorithm design and implementation [J ]. RemoteSensing, 2012, 4(9): 2768-2817.

Contents of the invention

Purpose of the invention: Aiming at the problems and deficiencies in the prior art, the present invention introduces the multi-scale J-image sequence in the traditional color texture segmentation method JSEG into high-resolution remote sensing image segmentation, and proposes a method based on the local homogeneity index. The adaptive SP selection strategy of J-value, the boundary constraint strategy of inter-scale objects, and the region merging strategy based on multi-features realize automatic multi-scale segmentation. Through experiments on remote sensing images with different sensor types and different spatial resolutions, and comparing with the experimental results of two supervised segmentation methods (eCognition and Document 2), it is proved that the proposed algorithm not only locates the edge of the object more accurately, but also extracts the outline of the object more accurately. Complete and without human intervention, it improves the automation and robustness of the segmentation process.

Technical solution: an unsupervised segmentation method for high-resolution remote sensing images based on automatic optimization of scale parameters, which mainly includes three steps: 1) Adaptive SP selection based on J-value of local homogeneity index, thus determining the best Multi-scale J-image sequence; 2) Image segmentation based on inter-scale object boundary constraint strategy, realizing multi-scale segmentation from coarse to fine; 3) Region merging based on multi-features to deal with possible over-segmentation in the segmentation results Phenomenon.

SP adaptive selection

The J-image sequence was chosen as the multi-scale analysis platform, and an adaptive selection strategy for SP was proposed.

The calculation process of the multi-scale J-image is as follows: Firstly, the color quantization is performed on the original image in LUV space. In the quantized image, set a window Z with a size of M×M (M is SP) pixels centered on the pixel z, and use the coordinate z(x,y) of each pixel in the window as its pixel value, and z(x,y)∈Z. At the same time, the corner points in the window are removed.

Let the total number of gray levels in the quantized image be P, let Z _p be the set of all pixels belonging to gray level p in window Z, and m _p be the mean value of all pixels corresponding to gray level p, then the window Z belongs to The sum of variances of pixels with the same gray level can be expressed as:

${\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; W</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; P</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; Z</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; |</span>{<span\; class="notranslate">\; |</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>$

The overall variance of all pixels within window Z can be expressed as:

${\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; z</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; Z</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; z</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; |</span>{<span\; class="notranslate">\; |</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$

Among them, m is the mean value of all pixels in the window Z. Then the local homogeneity index J-value can be defined as:

J-value＝(S _T -S _W )/S _W (3)

At this time, the J-value corresponding to the pixel z is used as the pixel value of the pixel, and the entire quantized image is traversed to obtain the J-image when the SP is M. By changing the SP, a multi-scale J-image sequence can be obtained.

Adaptive SP selection strategy based on J-value:

Step1: Calculate the J-image sequence when SP is M (M=5,6....N), where M=5 is the minimum window size allowed by J-image, and N represents the J-image with the roughest scale.

Step2: Calculate the mean value of pixel J-value under all scales J-image and build curve.

Step3: in many Among the inflection points of the curve, only some of the most prominent inflection points are selected, and these inflection points should satisfy

Constrained Multiscale Segmentation

In the segmentation stage, a segmentation strategy based on inter-scale object boundary constraints is proposed. Assuming that the optimal J-image sequence contains L scales, which can be expressed as S _k (k=1,2...L), the specific implementation process is as follows:

Step1: First segment the roughest scale S ₁ . The threshold T ₁ is determined according to formula (4) for seed region extraction, where μ _k and σ _k represent the mean and variance of J-value of all pixels in scale S _k , respectively.

T _k =μ _k -0.2σ _k , (k=1,2...L) (4)

In S ₁ , all pixels whose J-value is smaller than T ₁ use the four-connection method to form a connected area, which is used as a seed area. Starting from the seed area, the area is grown in the order of J-value from small to large in the four directions of up, down, left, and right. The boundary when adjacent areas meet constitutes the segmentation result under S ₁ .

Step2: Map the object boundary of the previous scale to the current scale and correct it. Convert the current scale J-image into a binary image, retain only the object boundaries extracted by inter-scale mapping, and perform morphological expansion operations. The size of the dilated structure unit is set to M×M pixels, where M is the SP of the current scale. Use the expanded boundary to divide the current scale J-image into independent seed regions, and grow these seed regions according to the J-value value from small to large. The boundary when adjacent regions meet is the boundary correction. result.

At this time, the segmentation at the current scale is only performed inside the object extracted by the corrected boundary. In order to avoid over-segmentation, objects with a high degree of internal homogeneity are considered to have matched the actual object types, and no segmentation is performed at the current scale. The judgment rule is that the average J-value inside the object should be smaller than the threshold T _k corresponding to the current scale (see formula 4). On this basis, the remaining objects are segmented according to the threshold _Tk pairs, the segmentation process is the same as scale 1, and the segmentation results are mapped to the next scale.

Step3: Repeat the segmentation process of Step2 until the scale L is segmented to obtain preliminary segmentation results.

Multi-featured region merging

Calculate the multi-scale J-image sequence corresponding to each band of the original image according to the best SP. Assuming that the original image contains F bands, for any object q, define the feature vector J _qf =(J _q1 ,J _q2 ,...,J _qF ), where each component represents the object q in each band L scale J- The J-value mean under the image, then for any band f (f=1,2...F), there are According to the definition of J-value, it can be seen that J-value comprehensively reflects the spectrum, texture and scale information of the local area (object), so the similarity can be judged by judging the Euclidean distance between the feature vectors of adjacent objects q _A and q _B , as shown in formula (5).

$\mathrm{<span\; class="notranslate">\; D.</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; q</span>}}_{\mathrm{<span\; class="notranslate">\; A</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; q</span>}}_{\mathrm{<span\; class="notranslate">\; B</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; J</span>}}_{{\mathrm{<span\; class="notranslate">\; q</span>}}_{\mathrm{<span\; class="notranslate">\; A</span>}}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; J</span>}}_{{\mathrm{<span\; class="notranslate">\; q</span>}}_{\mathrm{<span\; class="notranslate">\; B</span>}}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; )</span>$

Use RAG (Region Adjacency Graphics) to merge regions. The specific process is as follows:

Step1: According to the segmentation results in scale L, generate RAGs of all adjacent objects.

Step2: Select all objects adjacent to any object q _A , and calculate the Euclidean distance according to formula (5).

Step3: If there is an object q _B satisfying D(q _A , q _B )≤0.1, it is considered that q _A and q _B belong to the same object, and q _A and q _B are merged to generate a new RAG. Otherwise, return to Step2.

Step4: Repeat Step2 to Step3 to traverse all objects to obtain the final segmentation result.

Description of drawings

Fig. 1 method flowchart of the present invention;

Fig. 2 is the specific size window Z when M=9;

Figure 3 is the image of QucikBird in 2005;

Figure 4 is Curve and optimal SP selection;

Fig. 5 is the segmentation result of the method of the present invention;

Figure 6 is the method two segmentation results;

Fig. 7 is the method three segmentation results;

Figure 8 is the image of QucikBird in 2005;

Figure 9 is Curve and optimal SP selection;

Fig. 10 is the segmentation result of the method of the present invention;

Figure 11 is the method two segmentation results;

Fig. 12 is method three method segmentation results;

Figure 13 is Experiment 1 Accuracy evaluation;

Figure 14 is Experiment 2 Accuracy evaluation.

detailed description

Below in conjunction with specific embodiment, further illustrate the present invention, should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention, after having read the present invention, those skilled in the art will understand various equivalent forms of the present invention All modifications fall within the scope defined by the appended claims of the present application.

As shown in Figure 1, the unsupervised segmentation method of high-resolution remote sensing images based on automatic optimization of scale parameters mainly includes three steps: 1) Adaptive SP selection based on the J-value of the local homogeneity index, thus determining the optimal Jia multi-scale J-image sequence; 2) Image segmentation based on the inter-scale object boundary constraint strategy, which realizes multi-scale segmentation from coarse to fine; 3) Region merging based on multi-features to deal with possible overshoots in the segmentation results Segmentation phenomenon.

SP adaptive selection

Multi-scale segmentation usually mainly relies on a core control parameter, namely SP, to divide an image into independent geographic objects. SP controls the spectral homogeneity within the object in the segmentation result and the average size of the object, and actually affects the number of objects in the segmentation result. Therefore, selecting an appropriate multi-scale analysis tool and how to determine the SP is one of the key issues to be solved in the present invention.

Multi-scale J-image sequence

The multi-scale J-image sequence adopted by JSEG can fully reflect the homogeneity of the spectral distribution in the local area, and it can also avoid the traditional multi-scale analysis tools (such as wavelet transform, contourlet transform) when calculating multi-scale images. However, it also faces the problem of reasonable selection of SP. Therefore, this method chooses the J-image sequence as the multi-scale analysis platform, and proposes an adaptive selection strategy for SP.

The calculation process of the multi-scale J-image is as follows: Firstly, the color quantization is performed on the original image in LUV space. In the quantized image, set a window Z with a size of M×M (M is SP) pixels centered on the pixel z, and use the coordinate z(x,y) of each pixel in the window as its pixel value, and z(x,y)∈Z. At the same time, in order to ensure consistency in all directions, corner points in the window are removed. Taking M=9 as an example, the window Z centered on pixel z is shown in Figure 2:

Let the total number of gray levels in the quantized image be P, let Z _p be the set of all pixels belonging to gray level p in window Z, and m _p be the mean value of all pixels corresponding to gray level p, then the window Z belongs to The sum of variances of pixels with the same gray level can be expressed as:

${\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; W</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; P</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; Z</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; |</span>{<span\; class="notranslate">\; |</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>$

The overall variance of all pixels within window Z can be expressed as:

${\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; z</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; Z</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; z</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; |</span>{<span\; class="notranslate">\; |</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$

Among them, m is the mean value of all pixels in the window Z. Then the local homogeneity index J-value can be defined as:

J-value＝(S _T -S _W )/S _W (3)

At this time, the J-value corresponding to the pixel z is used as the pixel value of the pixel, and the entire quantized image is traversed to obtain the J-image when the SP is M. By changing the SP, a multi-scale J-image sequence can be obtained.

Optimal Scale Selection Based on J-value

Using J-image sequence as a multi-scale analysis tool, an adaptive SP selection strategy based on J-value is proposed.

Step1: Calculate the J-image sequence when SP is M (M=5,6....N), where M=5 is the minimum window size allowed by J-image, N (can be adjusted appropriately according to the actual image size, In this paper, N=30) represents the J-image with the roughest scale.

Step2: Calculate the mean value of pixel J-value under all scales J-image and build curve. According to the definition of J-value, The inflection point of the curve reflects the sharp increase in the homogeneity of the spectral distribution at the current scale compared with the front and back scales. Therefore, we assume that these inflection points indicate that some representative object types in the scene are suitable for segmentation at the current scale. That is, at these inflection points, the objects in the segmentation results just match the actual object types, and have the same or similar spectral homogeneity, and these representative objects can Curves have a noticeable effect.

Step3: Among the many inflection points, only select some of the most prominent inflection points, and these inflection points should satisfy In order to effectively extract the most representative types of ground objects in the scene as far as possible. In addition, in order to preserve the detailed information of the image, the finest scale (ie M=5) is always chosen. At this point, the SP corresponding to the selected inflection points and the finest scale can be integrated to jointly determine the optimal multi-scale J-image sequence.

Constrained Multiscale Segmentation

In the segmentation stage, a segmentation strategy based on inter-scale object boundary constraints is proposed. Assuming that the optimal J-image sequence contains L scales, which can be expressed as S _k (k=1,2...L), the specific implementation process is as follows:

Step1: First segment the roughest scale S ₁ . The threshold T ₁ is determined according to formula (4) for seed region extraction, where μ _k and σ _k represent the mean and variance of J-value of all pixels in scale S _k , respectively.

T _k =μ _k -0.2σ _k , (k=1,2...L) (4)

In S ₁ , all pixels whose J-value is smaller than T ₁ use the four-connection method to form a connected area, which is used as a seed area. Starting from the seed area, the area is grown in the order of J-value from small to large in the four directions of up, down, left, and right. The boundary when adjacent areas meet constitutes the segmentation result under S ₁ .

Step2: Map the object boundary of the previous scale to the current scale and correct it. According to the definition of J-image, the J-image multi-scale sequence has the same size as the original image. Therefore, the object boundaries extracted from the previous coarse-scale segmentation results can be mapped to the same position at the current fine-scale according to the coordinates, and the segmentation at the current scale is constrained. Although the boundary extracted at the coarse scale can determine the position of the object and its approximate outline, it is difficult to accurately locate the edge of the object, so it needs to be corrected at the current scale. The process is as follows.

Convert the current scale J-image into a binary image, retain only the object boundaries extracted by inter-scale mapping, and perform morphological expansion operations. The size of the dilated structure unit is set to M×M pixels, where M is the SP of the current scale. Use the expanded boundary to divide the current scale J-image into independent seed regions, and grow these seed regions according to the J-value value from small to large. The boundary when adjacent regions meet is the boundary correction. result.

At this time, the segmentation at the current scale is only performed inside the object extracted by the corrected boundary. In order to avoid over-segmentation, objects with a high degree of internal homogeneity are considered to have matched the actual object types, and no segmentation is performed at the current scale. The judgment rule is that the average J-value inside the object should be smaller than the threshold T _k corresponding to the current scale (see formula 4). On this basis, the remaining objects are segmented according to the threshold _Tk pairs, the segmentation process is the same as scale 1, and the segmentation results are mapped to the next scale.

Step3: Repeat the segmentation process of Step2 until the scale L is segmented to obtain preliminary segmentation results.

Multi-featured region merging

Although objects with high internal homogeneity have been discriminated before each scale is segmented, over-segmentation is still unavoidable, and further region merging is required. Due to the prominent phenomenon of "same object with different spectrum" and "same spectrum with different object" in high-resolution remote sensing images, only using the spectral features inside the object may cause mis-merging problems. Therefore, a multi-feature region merging strategy is proposed.

Calculate the multi-scale J-image sequence corresponding to each band of the original image according to the best SP. Assuming that the original image contains F bands, for any object q, define the feature vector J _qf =(J _q1 ,J _q2 ,...,J _qF ), where each component represents the object q in each band L scale J- The J-value mean under the image, then for any band f (f=1,2...F), there are According to the definition of J-value, it can be seen that J-value comprehensively reflects the spectrum, texture and scale information of the local area (object), so the similarity can be judged by judging the Euclidean distance between the feature vectors of adjacent objects q _A and q _B , as shown in formula (5).

$\mathrm{<span\; class="notranslate">\; D.</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; q</span>}}_{\mathrm{<span\; class="notranslate">\; A</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; q</span>}}_{\mathrm{<span\; class="notranslate">\; B</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; J</span>}}_{{\mathrm{<span\; class="notranslate">\; q</span>}}_{\mathrm{<span\; class="notranslate">\; A</span>}}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; J</span>}}_{{\mathrm{<span\; class="notranslate">\; q</span>}}_{\mathrm{<span\; class="notranslate">\; B</span>}}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; )</span>$

We use RAG (Region Adjacency Graphics) to merge regions. The specific process is as follows:

Step1: According to the segmentation results in scale L, generate RAGs of all adjacent objects.

Step2: Select all objects adjacent to any object q _A , and calculate the Euclidean distance according to formula (5).

Step3: If there is an object q _B satisfying D(q _A , q _B )≤0.1, it is considered that q _A and q _B belong to the same object, and q _A and q _B are merged to generate a new RAG. Otherwise, return to Step2.

Step4: Repeat Step2 to Step3 to traverse all objects to obtain the final segmentation result.

Experiment and Analysis

In order to verify the validity and reliability of the proposed method, experiments were carried out on two sets of multispectral high-resolution remote sensing images with different resolutions and different sensor types. The supervised high-resolution remote sensing image segmentation method (referred to as "method 3") proposed by et al. (document 2) was compared.

Among them, eCongnition is an internationally renowned object-oriented remote sensing image classification software developed by Germany’s Definiens Imaging. has excellent performance. eCongnition is mainly controlled by three parameters in segmentation, namely: scale parameter (Scale Parameter), which mainly controls the average size of the object in the segmentation result; shape parameter (Shape Parameter), which helps to maintain the integrity of the object outline during the segmentation process ;CompactnessParameter (CompactnessParameter), which helps to improve the inter-class separability of objects. The method proposed by Shao P et al. introduces traditional edge detection into high-resolution remote sensing image segmentation, and realizes multi-scale feature extraction and segmentation of objects by constructing object levels. good effect. The setting of relevant parameters in the above two methods needs to be realized through manual interpretation, and the present invention determines the optimal parameter combination through trial and error in experiments.

Experiment 1 Results and Analysis

Experiment 1 used QucikBird four-band color fusion data in 2005, located in Wuhan, China, with a spatial resolution of 2.4m and an image size of 512×512 pixels. The image is mainly a typical urban scene area under a complex background, including rich types of ground objects such as roads, playgrounds, water bodies, and man-made objects with complex structures, as shown in Figure 3.

In the method of the present invention The curve is shown in Figure 4, which describes that as the scale parameter M continues to increase, the index changes. Among them, the vertical dotted line corresponds to the best SP. These scales together with the finest scale constitute the best multi-scale J-image sequence. The corresponding scale parameter is M∈[5,13,18,28]. The final segmentation result is shown in Figure 5 shown.

In method two, set the Scale Parameter to 77, the Shape Parameter to 50, and the CompactnessParameter to 40. In Document 2, the proportion of each band is set to be the same, the parameter "Scale Parameter" is 30, the parameter "Shape Heterogeneous Degree" is 0.4, and the parameters "Compactness Parameter" and "SmoothnessParameter" are both 0.5. The experimental results of the two methods are shown in Figure 6 and Figure 7, respectively.

In order to compare the experimental results of different methods, the present invention marks the typical objects or positions in the scene. Among them, positions A and B are sports fields, positions C, D and F are buildings, position E is a road, and position G is an artificial lake. Through visual analysis, it can be seen that the three methods can effectively extract the sports field area at position A. Among them, this method and method two are obviously more accurate than method three in locating the edge of the lawn, and method three has a certain over-segmentation phenomenon; for position A In the playground area of B, the second method does not extract the outline of the lawn, and the third method confuses part of the runway area with the lawn; the structure of the building at position C is complex, and only this method not only maintains the integrity of the object outline, but also accurately locates it. For the edge of the object, methods 2 and 3 have mis-segmentation and over-segmentation respectively; the shape of the building at positions D and F is regular, and this method and method 2 are more accurate in locating the edge of the roof, but there is under-segmentation at position F and method 2 ; Only method 3 effectively extracts the contour information of the artificial lake at position G. In general, the method of the present invention and method 2 are obviously better than method 3 in the ability to locate the edge detail information of the object, and the method of the present invention maintains a more complete outline of a large homogeneous area. Method 3 can effectively identify adjacent ground objects belonging to different types but with similar spectral characteristics, but there are also prominent problems of low positioning accuracy and over-segmentation.

Experiment 2 Results and Analysis

In Experiment 2, three-band high-resolution aerial remote sensing DOM (Digital Orthophoto Map) images were selected. The data collection time was March 2009, the spatial resolution was 0.6m, and the size was 512×512 pixels. The location was Nanjing, China, as shown in Figure 8. Show. It can be seen from the comparison with Figure 3 that the background complexity of the data used in Experiment 2 has been reduced, but the details of spectrum, texture, edge and other details of different types of ground objects are more significant, and the objects include large-area homogeneous areas with regular shapes and Man-made buildings with complex structures and rich texture features put forward higher requirements for segmentation algorithms to accurately locate object edges and maintain the integrity of object outlines.

In this method The curve is shown in Figure 9. according to It can be seen from the curve that the scale parameter corresponding to the optimal multi-scale sequence is M∈[5,10,12,21,25,27], and the segmentation results are shown in Figure 10.

In method 2, set the Scale Parameter to 100, the Shape Parameter to 50, and the CompactnessParameter to 50. In Document 14, the proportion of each band is set to be the same, the Scale Parameter is 50, the ShapeHeterogeneous Degree is 0.3, and the Compactness Parameter and Smoothness Parameter are both 0.5. The experimental results of the two methods are shown in Figure 11 and Figure 12, respectively.

Same as Experiment 1, the present invention marks typical objects or positions in the scene. Can find out by visual analysis to different method experimental results, all three kinds of methods have accurately segmented out the playground lawn of position A and the playground runway of position B, but only the method of the present invention has effectively kept the integrality of playground outline, method Second, there are some long and narrow spurious units between two large segmentation areas, such as the area adjacent to the lawn on the outside of the runway; for the small lawn at position C, there is under-segmentation in method 2; for the roof of the building at position D, The segmentation effects of the three methods are similar. Methods 2 and 3 further extract the texture information inside the roof, but method 3 has over-segmentation and inaccurate edge positioning; the playground stands at positions E and F have complex structures and rich texture information. The inventive method maintains the integrity of the ceiling area of the stand and effectively extracts the details of the auxiliary buildings on both sides; the inventive method and the second method accurately extract the road areas located at positions G and H, while the third method has the problem of mis-segmentation; for the position For the badminton court at position I, the tennis court at position J, and the large vegetation area at position K, the three methods are more accurate in locating the edge of the object, but in the second method, there are narrow and long false units in the division of the tennis court. Based on the above analysis, a conclusion similar to that of Experiment 1 can be drawn, which further verifies the reliability of the proposed algorithm.

Accuracy evaluation

The above mainly evaluates the segmentation effect of different methods through visual analysis. Here, the accuracy index will be used to further quantitatively analyze the experimental results. In the method proposed by Deng et al. (Document 3), the index J-value is not only used to calculate the multi-scale J-image sequence, but also used as an evaluation index of segmentation accuracy, which is defined as follows:

${\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; J</span>}}}_{\mathrm{<span\; class="notranslate">\; a</span>}\mathrm{<span\; class="notranslate">\; v</span>}\mathrm{<span\; class="notranslate">\; g</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; u</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; R</span>}}{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}}{\mathrm{<span\; class="notranslate">\; J</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; r</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; 2...</span>}\mathrm{<span\; class="notranslate">\; R</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 6</span>}<span\; class="notranslate">\; )</span>$

Among them, U is the total number of pixels in the image, R is the total number of regions in the segmentation result, W _r and J _r are the total number of pixels inside the rth region and its corresponding J-value. When the accuracy index corresponding to the segmentation result The smaller the value, the higher the average spectral homogeneity within the object in the segmentation result, and the better the segmentation effect. Due to the large variety of ground objects in remote sensing images, the evaluation angles and standards for segmentation accuracy are not the same in different applications. However, Deng et al. It has good versatility to evaluate the spectral distribution of the segmentation results as a whole ^[4] . Therefore, the present invention adopts The accuracy of the experimental results of the three methods was evaluated.

Analyze the distribution of J _r in the segmentation results, and uniformly quantify the J-value corresponding to all objects to 20 units in the interval [0,1], then the J _r distribution curve is shown in Figure 13 and Figure 14:

In the figure, the solid lines of different colors represent the proportions of the objects belonging to different J-value intervals in the segmentation results of the three methods, while the dotted lines represent the proportions of the objects corresponding to the three methods. index. Can find out by comparing Fig. 13, 14: in two groups of experiments, the present invention proposes method accuracy index Both are significantly better than the other two methods, consistent with the results of visual decomposition. The trend of the J _r curves in the two experiments is roughly the same, and the difference is mainly reflected in the intervals of typical feature concentration in the scene, such as the [0.2,0.55] interval in Experiment 1 and the [0.1,0.4] interval in Experiment 2. In addition, in Experiment 2, the segmentation accuracy of the three algorithms was significantly improved compared with Experiment 1. The main reason is that the spatial resolution of remote sensing images used in Experiment 2 is higher, the background is relatively simple, and the details of ground objects are more prominent. Objects and their boundaries are closer to the types of ground objects in the actual scene.

Aiming at the automatic segmentation of high-resolution remote sensing images, the present invention proposes a novel non-supervised multi-scale segmentation method. This method makes comprehensive use of the spectral and texture features of the object, and proposes an adaptive scale parameter (SP) selection strategy based on the J-value based on the local homogeneity index, so that the typical object types in the scene can be matched with it. Segmentation in the best scale J-image. On this basis, the proposed multi-scale segmentation strategy makes the region segmentation subject to the constraints of the object boundaries extracted at the previous scale, and corrects these boundaries at the current scale to avoid the accumulation of errors between scales. The regional merging strategy based on multi-features can effectively distinguish different types of ground objects with similar spectral characteristics, and avoid the phenomenon of false merging. Experiments show that compared with eCognition and traditional supervised segmentation methods, the proposed method locates the object edge accurately and extracts the object outline more completely, which has higher segmentation accuracy and can realize automatic high-resolution remote sensing image segmentation without manual intervention in the whole process , is a generic and efficient unsupervised solution.

Claims (4)

1. a kind of non-supervisory dividing method of high score remote sensing image based on scale parameter Automatic Optimal is it is characterised in that main wrap Include three steps: 1) self adaptation sp of the j-value based on local homogeneity index selects, so that it is determined that optimal multiple dimensioned j- Image sequence；2) between based on yardstick, the image segmentation of object bounds constraints policy is it is achieved that arrive smart multi-scale division by thick； 3) region merging technique based on multiple features, to tackle over-segmentation phenomenon that may be present in segmentation result.

2. the non-supervisory dividing method of high score remote sensing image based on scale parameter Automatic Optimal as claimed in claim 1, it is special Levy and be, during self adaptation sp of the j-value based on local homogeneity index selects:

The calculating process of multiple dimensioned j-image is as follows: carries out color quantizing to raw video in luv space first；Quantifying shadow In picture, set the window z of a size of m × m centered on pixel z (m is sp) pixel, and the seat by each pixel in window Mark z (x, y) is as its pixel value, and z (x, y) ∈ z；Angle point in window is removed simultaneously；

If gray level sum is p in quantification image, make z_pFor belonging to the set of all pixels of gray level p, m in window z_pFor institute There is the pixel belonging to gray level p corresponding pixel average, then belong to the variance of same gray-level pixels in window z and can represent For:

$s_w=\underset{p=1}{\overset{p}{\mathrm{\&sigma;}}}\underset{q\&element;z_p}{\mathrm{\&sigma;}}\left|\right|q-m_p|{|}^2---\left(1\right)$

In window z, the population variance of all pixels is represented by:

$s_t=\underset{z\&element;z}{\mathrm{\&sigma;}}\left|\right|z-m|{|}^2---\left(2\right)$

Then homogeneity index j-value in local may be defined as:

J-value=(s_t-s_w)/s_w(3)

Now, using corresponding for pixel z j-value as the pixel value of this pixel, travel through view picture quantification image, can obtain sp is m When j-image, by change sp can obtain multiple dimensioned j-image sequence；

The step of the self adaptation sp selection strategy based on j-value is:

Step1: calculating sp is j-image sequence during m (m=5,6....n), and wherein m=5 is the min window that j-image allows Mouth size, n represents the j-image of coarse scale；

Step2: calculate the pixel j-value average under all yardstick j-imageAnd buildCurve.

Step3: numerousIn point of inflexion on a curve, only select some flex points the most prominent, these flex points should meet

3. the non-supervisory dividing method of high score remote sensing image based on scale parameter Automatic Optimal as claimed in claim 2, it is special Levy and be, in the segmentation stage, the segmentation strategy of object bounds constraint between proposing based on yardstick；If optimal j-image sequence comprises l Individual yardstick, is represented by s_k(k=1,2...l), implements process as follows:

Step1: first to coarse scale s₁Split.According to formula (4) threshold value t₁Carry out seed region extraction, its Middle μ_kAnd σ_kRepresent yardstick s respectively_kThe j-value average of middle all pixels and variance；

t_k=μ_k-0.2σ_k, (k=1,2...l) (4)

In s₁In, all j-value values are less than t₁Pixel adopt four methods of attachment to constitute UNICOM regions, as seed one by one Region；With seed region as starting point, according to four direction up and down, with j-value value, order from small to large carries out region increasing Long, border when adjacent area crosses just constitutes s₁Under segmentation result；

Step2: the object bounds of a upper yardstick are mapped to current scale and are modified；Current scale j-image is converted For a width bianry image, only retain the object bounds by mapping extraction between yardstick, and carry out morphological dilation.Expand knot Structure unit size is set as m × m pixel, and m is the sp of current scale；Using the border after expanding, current scale j-image is drawn It is divided into seed region independent one by one, and according to j-value value, region growth, phase are carried out from small to large to these seed regions Border when neighbouring region crosses is the result that border is revised；

The higher object of internal homogenizing degree is thought that it is mated with actual type of ground objects, no longer enters under current scale Row segmentation；Judgment rule is that the j-value average within this object should be less than corresponding threshold value t of current scale_k；Here basis On, according to threshold value t_kRemaining object is split, cutting procedure is identical with yardstick 1, and segmentation result is mapped to next chi Degree；

Step3: repeat the cutting procedure of step2, until yardstick l segmentation finishes, thus obtaining preliminary segmentation result.

4. the non-supervisory dividing method of high score remote sensing image based on scale parameter Automatic Optimal as claimed in claim 3, it is special Levy and be, raw video each wave band correspondence multiple dimensioned j-image sequence is calculated according to optimal sp；If raw video comprises f Individual wave band, for any one object q, defined feature vector j_qf=(j_q1,j_q2..., j_qf), wherein each component represent right As j-value average under l yardstick j-image of each wave band for the q, then for any wave band f, (f=1,2 ... f), hasBy judging adjacent object q_aAnd q_bCharacteristic vector between Euclidean distance judging its similar journey Degree, such as shown in formula (5).

$d(q_a,q_b)=\left|\right|j_{q_a}-j_{q_b}\left|\right|---\left(5\right)$

Region merging technique is carried out using rag, detailed process is as follows:

Step1: according to the segmentation result in yardstick l, generate the rag of all adjacent object；

Step2: select and any object q_aAdjacent all objects, calculate Euclidean distance according to formula (5)；

Step3: if there is object q_bMeet d (q_a,q_b)≤0.1 is then it is assumed that q_aAnd q_bBelong to same target, merge q_aAnd q_bAnd it is raw The rag of Cheng Xin；Otherwise, return step2；

Step4: repeat step2 to step3, travel through all objects, obtain final segmentation result.