A road network extraction method, system, device and medium based on remote sensing images

Abstract

The application belongs to the technical field of target classification, image fusion and remote sensing image processing, and discloses a road network extraction method, system, device and medium based on remote sensing images; the road network extraction method comprises the following steps: based on remote sensing images, pixel level features and object level features under multiple scales of each pixel point are acquired, and a training set is obtained by extracting the pixel points; the pixel level features and the object level features under multiple scales of each pixel point in the training set are respectively solved by using OCSVM, decision values are obtained, and the decision values are converted into probabilities of belonging to roads; the probabilities of belonging to roads are fused, a binary image of a road image is obtained, and a road center line network is extracted. The application combines the advantages of different level features and a one-class support vector machine which only focuses on positive sample features, and can extract a more complete and accurate road network.

Description

Technical Field

[0001] This invention relates to the fields of target classification, image fusion, and remote sensing image processing, and particularly to a method, system, device, and medium for road network extraction based on remote sensing images. Background Technology

[0002] Roads are not only basic geographic information, but also a fundamental condition and key support for many industries; among them, road network extraction based on remote sensing imagery is of great significance in urban planning, military operational deployment, and route planning.

[0003] Most existing traditional road network extraction methods rely on single features. For example, Yager et al. used edge features and Support Vector Machine (SVM) to extract road networks from remote sensing images; Hossain et al. proposed a road network extraction method based on mathematical morphology and spectral features, which has the advantages of simple steps and high efficiency, but cannot handle more complex road conditions. In summary, the above-mentioned traditional extraction methods based on single road features cannot make full use of road information and are difficult to achieve good extraction results.

[0004] The introduction of the fusion concept provides a new approach to improve the road network extraction effect. For example, Song and Civco et al. fused the results of support vector machine classification based on spectral features with the results of region growing segmentation based on shape features, achieving better extraction results. Shi's algorithm used pixel-level spectral spatial features and uniformity features, classified them separately using SVM, and then fused the results. The above methods use pixel-level features, and the extraction effect is not ideal for regions with complex contextual features.

[0005] The introduction of object-level features enhances the representation of contextual and spatial information beyond pixel-level features. For example, Runyi's algorithm combines K-means classification results obtained from spectral features derived through superpixel decomposition with classification results using GNDVI index features, achieving good results. Peng Yaqi's algorithm uses object-level spectral features and second-order moment features, employs random forest for classification, and uses DS evidence theory for fusion. However, due to the similarity between roads and surrounding features, methods using only object-level features are prone to exhibiting adhesion and incomplete block loss.

[0006] Based on the above analysis, it can be seen that although the existing methods have improved the shortcomings of traditional methods in describing roads using a single feature, they all start from a single level and cannot solve the hole phenomenon caused by using only pixel-level features and the adhesion problem caused by using only object-level features.

[0007] Cao Yungang et al. fed object-level and pixel-level features into an SVM classifier and fused the classification results at the decision level using methods such as DS evidence theory and multi-scale voting. This effectively alleviated the problems of holes and confluence in the extraction results caused by insufficient road description using a single-level feature. However, the existing methods use a large number of feature parameters, and parameter selection is a challenge. Inappropriate feature parameters may make it difficult for the features to properly represent the road attributes, resulting in poor classification performance of the trained classifier. In addition, the SVM used has poor applicability in road network extraction tasks. Interpretably, SVM is a common supervised binary classifier that requires a certain number of positive and negative samples for training. It performs well when the number of positive and negative samples is similar. However, in road network extraction tasks, there are many types of non-road features, which are difficult to describe with uniform features. Moreover, the sample size of road and non-road categories differs greatly. In this case, the classification surface generated by SVM will be significantly shifted to the side with fewer samples, resulting in lower classification accuracy. Furthermore, the classification results are greatly affected by sample selection, and the often-used random sampling may lead to poor classification. Summary of the Invention

[0008] The purpose of this invention is to provide a road network extraction method, system, device, and medium based on remote sensing imagery to solve one or more of the aforementioned technical problems. Specifically, the technical solution provided by this invention is a road network extraction scheme based on OCSVM (One-class Support Vector Machine) and multi-level feature fusion. It combines the advantages of features from different levels with a type of support vector machine that focuses only on positive sample features, enabling the extraction of a more complete and accurate road network.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] The first aspect of this invention provides a road network extraction method based on remote sensing imagery, comprising the following steps:

[0011] Based on remote sensing images, pixel-level features and object-level features at multiple scales are obtained for each pixel. Multiple pixels that meet the requirements are extracted from the remote sensing images to obtain a training set. Pixel-level features are formed by superimposing spectral features with gradient features, and object-level features are formed by superimposing spectral features with local Moran's index features.

[0012] Based on the pixel-level features and multi-scale object-level features of each pixel in the training set, OCSVM is used to solve the problem to obtain the pixel-level decision function and the object-level decision function. Based on the obtained pixel-level decision function and object-level decision function, the pixel-level decision value and object-level decision value of each pixel in the remote sensing image are obtained.

[0013] The obtained pixel-level decision values ​​and object-level decision values ​​of each pixel are converted into the probability that each pixel belongs to the road at the pixel level and object level; the obtained probabilities that each pixel belongs to the road at the pixel level and object level are fused to obtain a binary image of the road image.

[0014] Centerlines are extracted from the binary image of the road image to obtain a road centerline network.

[0015] A further improvement to the method of the present invention lies in that, in the step of acquiring pixel-level features and multi-scale object-level features based on remote sensing images,

[0016] The remote sensing image is an image obtained after bilateral filtering preprocessing based on the original remote sensing image.

[0017] A further improvement to the method of the present invention is that the step of extracting multiple qualified pixels based on the remote sensing image to obtain a training set includes:

[0018] Based on the remote sensing image, pixels are extracted using an adaptive domain-based method to obtain a training set;

[0019] Among them, the method of extracting pixels using adaptive neighborhood is described.

[0020] For a pixel i∈D, its associated adaptive domain GAN is defined as a subset contained in D. Adaptive domain GANs are built on a standard mapping h∈c, which is related to the uniformity tolerance m of the range on E; where,

[0021] The adaptive domain GAN is represented as follows:

[0022]

[0023]

[0024] In the formula, C X (i) represents the connected component containing pixel i∈D. E, c, and D represent the uniformity tolerance m, the standard mapping h, and the space where pixel i in the image resides, respectively. -1 ([h(i)-m],[h(i)+m]) represents pixel i that satisfies the standard mapping conditions;

[0025] Based on adaptive domain GANs, the selected training set is represented as follows:

[0026]

[0027] In the formula, g represents the grayscale feature of the image. i is the seed point, and i2 represents the index of a different seed point.

[0028] A further improvement to the method of the present invention lies in the step of using OCSVM to solve for the pixel-level decision function and the object-level decision function based on the pixel-level features and multi-scale object-level features of each pixel in the training set, respectively.

[0029] For the training set pixels in Pixel-level features and multi-scale object-level features are represented as follows:

[0030]

[0031]

[0032] In the formula, The pixel-level feature at pixel i3 This represents the object-level feature at pixel i3 at scale k, where k represents different scales. These represent the spectral features, gradient features, and local Moran index features at pixel i3, respectively.

[0033] Solve using OCSVM respectively, the solution expression is,

[0034]

[0035] In the formula, ω is the weight of the support vector; ρ is the threshold of the support vector. is a slack variable used to indicate the allowed existence of outliers; v∈(0,1) is a parameter used to control the upper limit of the number of outliers and the lower limit of the total number of support vectors; N represents the number of support vectors; φ represents the feature vector of pixel i3; φ is a mapping function used to map data samples to a higher-dimensional feature space, making them linearly separable.

[0036] During the solution process, the hyperplane parameters ω at different levels are obtained by using features at different levels. pixel ρ pixel ω object k ρ object k And construct the decision functions for each level, expressed as follows:

[0037]

[0038] In the formula, This represents the decision value at pixel level i. ω represents the decision value of pixel i at scale k for the object hierarchy; pixel ρ pixel ω represents the weights and thresholds of the support vectors at the pixel level, respectively; object k ρ object k φ represents the weights and thresholds of the support vectors in the object hierarchy at scale k, respectively; φ is the mapping function.

[0039] A further improvement of the method of the present invention lies in the step of converting the obtained pixel-level decision values ​​and object-level decision values ​​of each pixel into the probability that each pixel belongs to the road at the pixel level and object level.

[0040] The Platt scaling method is used to convert decision values ​​into probability values, thereby obtaining the probability that each pixel belongs to a road at each level, forming a coarse road extraction result at each level; wherein, the coarse road extraction result at each level includes the sub-road coarse extraction result at the object level and the road coarse extraction result at the pixel level at multiple scales.

[0041] A further improvement of the method of the present invention is that the step of fusing the probabilities of each obtained pixel point belonging to the road at the pixel level and the object level to obtain a binary image of the road image includes:

[0042] A mean fusion strategy is adopted to fuse the coarse extraction results of sub-roads at the object level under multiple scales to form a coarse extraction result at the object level.

[0043] The coarse extraction results at the pixel level and the object level will be merged to form the final coarse road extraction result;

[0044] Based on the final coarse road extraction result, fine processing is performed to obtain a binary image of the road image;

[0045] The step of fusing the coarse extraction results at the pixel level and the object level includes:

[0046] For the pixel-level probability matrix P pixel With the object-level probability matrix P object The fusion will adopt a decision fusion based on multi-threshold processing, and the fusion rule is as follows:

[0047]

[0048] In the formula, r represents the binary image value of the road image; when r = 1, it indicates that the pixel is a road, and when r = 0, it indicates that the pixel is not a road. These represent the high and low thresholds used at the object level and pixel level, respectively.

[0049] A further improvement of the method of the present invention is that the refinement process includes one or more of edge filtering, hole filling and shape filtering.

[0050] A second aspect of the present invention provides a road network extraction system based on remote sensing imagery, comprising:

[0051] The training set acquisition module is used to acquire pixel-level features and object-level features at multiple scales for each pixel based on remote sensing images; and to extract multiple pixels that meet the requirements based on the remote sensing images to obtain a training set; wherein, pixel-level features are formed by superimposing spectral features and gradient features, and object-level features are formed by superimposing spectral features and local Moran's index features.

[0052] The decision value acquisition module is used to solve OCSVM based on the pixel-level features and multi-scale object-level features of each pixel in the training set, respectively, to obtain the pixel-level decision function and the object-level decision function; based on the obtained pixel-level decision function and object-level decision function, the pixel-level decision value and the object-level decision value of each pixel in the remote sensing image are obtained.

[0053] The binary image acquisition module is used to convert the pixel-level decision value and object-level decision value of each pixel into the probability that each pixel belongs to the road at the pixel level and object level; and to fuse the probabilities that each pixel belongs to the road at the pixel level and object level to obtain a binary image of the road image.

[0054] The road network extraction module is used to extract the centerline based on the binary image of the road image to obtain the road centerline network.

[0055] A third aspect of the present invention provides an electronic device comprising:

[0056] At least one processor; and,

[0057] A memory communicatively connected to the at least one processor; wherein,

[0058] The memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the road network extraction method based on remote sensing imagery as described in any of the first aspects of the present invention.

[0059] The fourth aspect of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the road network extraction method based on remote sensing imagery as described in any one of the first aspects of the present invention.

[0060] Compared with the prior art, the present invention has the following beneficial effects:

[0061] The road network extraction method based on remote sensing imagery provided by this invention is a road network extraction scheme based on OCSVM and multi-level feature fusion. It combines the advantages of features at different levels with a type of support vector machine that focuses only on positive sample features, enabling the extraction of a more complete and accurate road network. Specifically, addressing the issue that a large number of feature parameters and poor feature robustness can easily lead to poor classification performance of the trained classifier, the technical solution of this invention comprehensively considers the characteristics of roads and the number of feature parameters, selecting spectral features, gradient features, and uniformity features to describe road attributes. These features do not require parameter adjustment and can be approached from both the road edge and the road interior, providing a better description of the road and fully utilizing the inherent characteristics of each feature. This approach effectively improves the classification accuracy of the trained model.

[0062] To address the shortcomings of SVM (Small Classifier) ​​training, which tends to favor the side with fewer samples when there is a large difference in the amount of positive and negative samples, resulting in poor classification performance and inapplicability to unlabeled datasets, this invention employs an improved OCSVM training technique. This technique uses a GAN (Generative Adversarial Network) approach for selecting the training set and a Platt (Platt) method to probabilize the classification results. This approach allows for convenient selection of the training set even without label information, focusing only on the characteristics of positive samples. The trained classifier only needs to form a boundary description of the target samples. Furthermore, this method converts the decision values ​​output by OCSVM into probability values, facilitating subsequent fusion and effectively improving classification accuracy and completeness to solve the aforementioned problems.

[0063] The fusion method also has a significant impact on the final road network extraction result. Traditional maximum value strategies and mean value strategies are prone to large errors, while methods such as DS evidence theory are computationally complex, and their uncertainty and confidence are difficult to obtain accurately. To address the shortcomings of multi-feature fusion using DS evidence theory, which suffers from computational complexity and difficulty in accurately describing uncertainty, leading to poor fusion results, the technical solution of this invention designs a multi-layer fusion strategy. Multiple thresholds are selected at each level to fuse the results, effectively improving the accuracy of the fusion results. Attached Figure Description

[0064] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art are briefly introduced below; obviously, the drawings described below are some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0065] Figure 1 This is a schematic flowchart of a road network extraction method based on remote sensing imagery provided in an embodiment of the present invention;

[0066] Figure 2 This is a schematic block diagram illustrating the overall process of road network extraction in an embodiment of the present invention;

[0067] Figure 3 This is a schematic diagram of feature extraction in an embodiment of the present invention;

[0068] Figure 4 This is a schematic diagram of the fusion process in an embodiment of the present invention;

[0069] Figure 5 This is a schematic diagram of the buffer in an embodiment of the present invention;

[0070] Figure 6 This is a schematic diagram of the experimental process of image 1 in an embodiment of the present invention; wherein, Figure 6 Image (a) is a schematic diagram of image 1. Figure 6 Image (b) is a schematic diagram of the bilateral filtering results for image 1. Figure 6 Image (c) is a schematic diagram of coarse road extraction binary data from image 1. Figure 6 Image 1 (d) is a refined binary representation of image 1. Figure 6 Image (e) is a schematic diagram of the road centerline in Image 1. Figure 6 Image (f) is a schematic diagram of the overlay of image 1;

[0071] Figure 7 This is a schematic diagram of the experimental process of image 2 in an embodiment of the present invention; wherein, Figure 7 Image (a) is a schematic diagram of image 2. Figure 7 Image (b) is a schematic diagram of the bilateral filtering results for image 2. Figure 7 Image (c) is a schematic diagram of coarse road extraction binary data from image 2. Figure 7 Image (d) is a detailed binary representation of image 2. Figure 7 Image (e) is a schematic diagram of the road centerline in Image 2. Figure 7 Image (f) is a schematic diagram of the overlay of image 2;

[0072] Figure 8 This is a schematic diagram of the experimental process of image 3 in an embodiment of the present invention; wherein, Figure 8 Image (a) is a schematic diagram of image 3. Figure 8 Image (b) is a schematic diagram of the bilateral filtering results for image 3. Figure 8 Image (c) is a schematic diagram of coarse road extraction binary data from image 3. Figure 8 Image 3 (d) is a detailed binary representation of image 3. Figure 8 Image (e) is a schematic diagram of the road centerline in image 3. Figure 8Image (f) is a schematic diagram of the overlay of image 3;

[0073] Figure 9 This is a schematic diagram comparing the road and centerline extraction results under different methods in an embodiment of the present invention; wherein, Figure 9 Image (a) is a schematic diagram of the SVM-LLKR method. Figure 9 (b) is a schematic diagram of the SLIC-Kmeans method. Figure 9 (c) is a schematic diagram of the RF-DS method. Figure 9 (d) is a schematic diagram of the method in an embodiment of the present invention;

[0074] Figure 10 This is a schematic diagram of a road network extraction system based on remote sensing images provided in an embodiment of the present invention. Detailed Implementation

[0075] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0076] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0077] The present invention will now be described in further detail with reference to the accompanying drawings:

[0078] Please see Figure 1 The present invention provides a road network extraction method based on remote sensing imagery, comprising the following steps:

[0079] Step 1: Based on the remote sensing image, obtain the pixel-level features and object-level features at multiple scales for each pixel; based on the remote sensing image, pixel points can be extracted using an adaptive domain method to obtain a training set; wherein, the training set includes multiple pixel points that meet the requirements;

[0080] Step 2: Based on the pixel-level features and multi-scale object-level features of each pixel in the training set, use OCSVM to solve the problems respectively, and obtain the pixel-level decision function and the object-level decision function accordingly; based on the obtained pixel-level decision function and object-level decision function, obtain the pixel-level decision value and the object-level decision value of each pixel in the remote sensing image.

[0081] Step 3: Convert the pixel-level decision value and object-level decision value of each pixel obtained in Step 2 into the probability that each pixel belongs to the road at the pixel level and object level; fuse the obtained probabilities of each pixel belonging to the road at the pixel level and object level to obtain a binary image of the road image.

[0082] Step 4: Based on the binary image of the road image obtained in Step 3, obtain the road centerline network.

[0083] In a further preferred embodiment of the present invention, in step 1, during the process of acquiring pixel-level features and multi-scale object-level features based on remote sensing images, the remote sensing images are images obtained after bilateral filtering preprocessing based on the original remote sensing images; interpretably, the purpose of the above preprocessing is to achieve the effect of noise reduction and edge preservation, providing higher quality images for subsequent road extraction.

[0084] Please see Figure 2 and Figure 3 In a further preferred embodiment of the present invention, during step 1, in the process of acquiring pixel-level features and multi-scale object-level features based on remote sensing images,

[0085] 1) Pixel-level features are formed by superimposing spectral features and gradient features. Further explanation is needed; for road extraction tasks, spectral features are the most basic and effective features. Therefore, the technical solution of this embodiment uses the original spectral features. Furthermore, the contact area between roads and non-roads often has small spectral differences, frequently causing adhesion phenomena. Therefore, enhancing road edges is crucial. The technical solution of this embodiment introduces gradient information as a discriminative feature to further distinguish road areas from non-road areas.

[0086] For example, for pixel i, its pixel-level feature vector is defined as follows: In the formula, S i G represents the spectral characteristics of pixel i. i It is the gradient feature of pixel i; where the gradient feature is obtained by using the difference approximation of the grayscale image.

[0087] Furthermore, the Sobel operator has advantages such as simplicity, efficiency, fast computation speed, and strong response to edge information, therefore it is used for edge extraction; the gradient calculation formula is as follows:

[0088] F x =g(x+1,y-1)+2*g(x+1,y)+g(x+1,y+1)

[0089] -g(x-1,y-1)-2*g(x-1,y)-g(x-1,y+1)

[0090] Fy=g(x-1,y+1)+2*g(x,y+1)+g(x+1,y+1)

[0091] -g(x-1,y-1)-2*g(x,y-1)-g(x+1,y-1)

[0092] In the formula, x and y represent the horizontal and vertical coordinates of a pixel, respectively, g(x,y) represents the gray value at (x,y) in the grayscale image, and F x F y Let represent the gradients in the horizontal and vertical directions, respectively. The 2-norm is used to synthesize the gradient information from both directions, forming the gradient features of the image, represented as .

[0093] 2) Object-level features are formed by superimposing spectral features and local Moran's index features. Further explanation is that since spectral features can clearly reflect the difference between roads and non-roads, spectral features are also selected for object-level features. Furthermore, considering that roads are continuous and generally homogeneous regions, this embodiment uses the Local Moran's Index to describe the local uniformity of the image. The Local Moran's Index is a statistical indicator used to measure the degree of clustering and spatial correlation of spatial data; it describes the degree of similarity between a pixel and other pixels, and its calculation formula is as follows. In the formula, i1 represents the center pixel, and j1 represents the other pixels within the neighborhood window of the center pixel. It is the pixel grayscale value at position i1. W is the average grayscale value of the pixels within the window. ij This is a spatial weight matrix, where n represents the number of pixels within the window. This represents the local Moran index at position i1.

[0094] The road network extraction method based on remote sensing imagery provided in this invention has discovered features that require fewer parameters and can better describe roads.

[0095] The principle of the invention in this embodiment is explained as follows: In high-resolution remote sensing images, roads are connected, and spatially adjacent pixels are more likely to belong to the same class. Furthermore, superpixel-based methods more easily incorporate multi-scale and structural information. Therefore, to reduce misclassification caused by occlusion and spectral variations, this embodiment uses superpixels as the processing object. Simple Linear Iterative Clustering (SLIC) has few parameters, runs quickly, and has good segmentation results; therefore, this embodiment selects the SLIC method to generate superpixels. In high-resolution remote sensing images, different land features vary in size, and analyzing using only a single scale is insufficient to comprehensively reflect land features at different scales. To more fully utilize the multi-scale information of high-resolution remote sensing images, this embodiment adjusts the hyperparameters in the SLIC method to obtain superpixel segmentation results at different scales. Specifically, this embodiment selects three scales, and the feature of each superpixel is the average of the features of all pixels, completing the object-level feature extraction at three scales. For example, for pixel i, its object-level feature vector is defined as... Where k takes values ​​of [1,2,3], representing different scales.

[0096] In a further preferred embodiment of the present invention, step 1, which involves extracting pixels from the remote sensing image using a General Adaptive Neighborhood (GAN) method to obtain a training set, may specifically include: employing an improved OCSVM (interpretive, innovatively adding a module for selecting the training set using GAN and the Platt method to convert decision values ​​into probability values, making OCSVM more suitable for unsupervised road extraction tasks).

[0097] For each pixel i∈D in an image, its associated GAN is defined as a subset contained in D. GANs are built on a standard mapping h∈c, which is related to the uniformity tolerance m over a range E, and is based on local measures of the image f, such as brightness and contrast. The mathematical definition of GAN is as follows:

[0098]

[0099] In the formula, C X (i) represents the connected component containing pixel i∈D. E, c, and D represent the uniformity tolerance m, the standard mapping h, and the space where pixel i in the image resides, respectively. -1 ([h(i)-m],[h(i)+m]) represents pixel i that satisfies the standard mapping conditions. Select seed point. The training set can be obtained using the above formula;

[0100] Therefore, the training set selected in this embodiment of the invention is represented as follows:

[0101]

[0102] In the formula, g represents the grayscale feature of the image. i is a seed point, and i2 represents the index of different seed points. The region formed by merging the various seed points through the training set generated by the above formula is called the training region R.

[0103] This invention proposes an improved OCSVM, which adds two modules to the traditional OCSVM: one for selecting the training set using GAN and the other for converting the output decision values ​​into probability values ​​using the Platt method. The former module enables the proposed method to be applied to unlabeled datasets, while the latter module generates probability values, providing the possibility of soft decision-making for subsequent fusion.

[0104] In a further preferred embodiment of the present invention, in step 2, OCSVM is used to solve for the pixel-level and object-level decision functions based on the pixel-level features and object-level features of each pixel in the training set, respectively.

[0105] For the pixels in the training set obtained above

[0106] Its pixel-level features and multi-scale object-level features are respectively represented as follows:

[0107]

[0108]

[0109] In the formula, The pixel-level feature at pixel i3 This represents the object-level feature at pixel i3 at scale k, where k represents different scales. These represent the spectral features, gradient features, and local Moran index features at pixel i3, respectively.

[0110] For each level of feature, OCSVM is used to solve the problem, and the solution expression is as follows:

[0111]

[0112] Where ω is the weight of the support vector; ρ is the threshold of the support vector; φ represents the feature of pixel i3, and φ is a mapping function that can map data samples to a higher-dimensional feature space, making them linearly separable. ω is a slack variable, indicating that outliers are allowed; N represents the number of support vectors; v∈(0,1) is a parameter controlling the upper limit of the number of outliers and the lower limit of the total number of support vectors. By using features at different levels, the hyperplane parameters ω at different levels can be obtained respectively. pixel ρ pixel ω object k ρ object k ;

[0113] Use the parameters obtained from the above formula to construct the decision functions for each level:

[0114]

[0115] In the formula, This represents the decision value at pixel level i. ω represents the decision value of pixel i at scale k for the object hierarchy; pixel ρ pixel ω represents the weights and thresholds of the support vectors at the pixel level, respectively; object k ρ object k φ represents the weights and thresholds of the support vectors in the object hierarchy at scale k, respectively; φ is the mapping function.

[0116] To further explain, the above formula can be used to obtain the decision values ​​for each level of the image's pixels.

[0117] In a further preferred embodiment of the present invention, in step 3, the pixel-level decision value and object-level decision value of each pixel obtained in step 2 are converted into the probability that each pixel belongs to the road at the pixel level and object level.

[0118] The Platt scaling method is used to transform decision values ​​into probability values; specifically, for the output decision value, it is transformed into a posterior probability value using the following formula:

[0119]

[0120] In the formula, This represents the probability that pixel i belongs to the road class at pixel level. This represents the probability that pixel i belongs to a road in the object hierarchy at scale k. This represents the decision value at pixel level i. Let represent the decision value of pixel i at scale k for the object level; A and B are the scale parameter and position parameter, respectively, which are fitted by maximum likelihood estimation, and their expression is as follows:

[0121]

[0122] In the formula, P i Let t be the probability value of pixel x. i This is called the target probability, and it is calculated using the following formula:

[0123]

[0124] In the formula, label i N is the label predicted by OCSVM for pixel i, where +1 indicates road class and -1 indicates non-road class; + N is the number of positive samples. - The number of negative samples is represented by these two parameters, which are statistically analyzed using the labels predicted by OCSVM.

[0125] Using the above method, the probability of each pixel belonging to a road at each level can be obtained, forming a coarse road extraction result at each level; including coarse sub-road extraction results at multiple object levels and coarse road extraction results at the pixel level at different scales.

[0126] Please see Figure 4 In a further preferred embodiment of the present invention, step 3, which involves fusing the probabilities of each pixel belonging to a road at both the pixel level and the object level to obtain a binary image of the road image, specifically includes:

[0127] First, the coarse extraction results of sub-roads at multiple scales are fused to form a coarse extraction result at the object level. The process of fusing the coarse extraction results of sub-roads at multiple scales involves:

[0128] For multi-scale object-level features, the classifier of this embodiment is used for training and prediction, and the decision values ​​are converted into probability values ​​to obtain three probability matrices P for the object level. object 1. P object 2. P object 3. These represent the probabilities of each pixel belonging to the road category at different scales. To combine the advantages of road network extraction results at different scales, a mean fusion strategy is used to integrate information from different scales. The probability that pixel i belongs to a road at the object level can be expressed as:

[0129]

[0130] The principle behind the invention's technical solution is as follows: To fully utilize the land cover information contained in superpixels at different scales, it is necessary to fuse the coarse road network extraction results obtained using superpixel features at different scales. The mean fusion strategy can reduce misclassification at a single scale, improve model stability, and is simple and effective.

[0131] Secondly, the coarse extraction results at the pixel level and the object level are fused to form the final coarse road extraction result; the process of fusing the coarse extraction results at the pixel level and the object level involves:

[0132] For the pixel-level probability matrix P pixel With the object-level probability matrix P object The fusion will employ a decision fusion based on multi-threshold processing; specifically, to make more accurate judgments on pixels from multiple perspectives, two thresholds, one high and one low, will be set for both pixel-level and object-level outputs.

[0133] Based on the probability of pixels at different levels, they are divided into three categories. If the posterior probability of a pixel is higher than a high threshold at any level or higher than a low threshold at both levels, it is classified as a road; otherwise, it is classified as a non-road. The pixel-level probability matrix and the object-level probability matrix are fused according to the following rules:

[0134]

[0135] In the formula, r represents the binary image value of the road image; when r = 1, it indicates that the pixel is a road, and when r = 0, it indicates that the pixel is not a road. These represent the high and low thresholds used at the object level and pixel level, respectively.

[0136] The principle behind the invention is as follows: when the posterior probability of a pixel is higher than a high threshold at any level, it indicates a very high probability that it belongs to the road class, and it can be directly classified as a road. If the posterior probability obtained at a certain level is between the high and low thresholds, then the posterior probability of that pixel in the probability matrix of another level needs to be judged. If it is also between the high and low thresholds, then the pixel is classified as a road. In all other cases, the pixel is considered not to be a road.

[0137] In this embodiment of the invention, the concept of hysteresis thresholding is introduced to select a threshold. Hysteresis thresholding finds two thresholds, high and low, by maximizing the inter-class variance, and classifies the image using these two thresholds.

[0138] The training region R is expanded to R through dilation in mathematical morphology. plus We introduce posterior probability values ​​for some non-road areas. Then we apply R... plus All probability values ​​within the image are processed using a hysteresis thresholding method to obtain pixel-level and object-level high and low thresholds. Then, using the aforementioned fusion rules, the final binary road image is formed.

[0139] A further preferred embodiment of the present invention includes: road refinement processing; specifically,

[0140] After decision-level fusion processing, misclassification was reduced to some extent. However, some errors still needed correction. Therefore, three operations were performed to refine the road binary map.

[0141] 1) Edge filtering, including:

[0142] To address the issue of road areas overlapping with surrounding non-road areas, an edge filtering method is employed to separate the road from the surrounding non-road features. First, edges are extracted from the original image, and then the edges are assigned a value of 0, which significantly reduces the overlap phenomenon.

[0143] 2) Hole filling, including:

[0144] Due to vehicle obstruction and noise interference, small holes exist in the road. To compensate for this, a hole-filling operation is performed. First, mathematical morphology is used to directly fill the holes in the image. Second, the original binary image is subtracted from the resulting image to obtain hole patches. Because this operation targets vehicle obstruction and noise interference, the hole area will not be very large. Hole patches with an area larger than a certain threshold are deleted; a recommended threshold is 30. The remaining holes are then added back to the original image to complete the hole-filling process.

[0145] 3) Shape filtering, including:

[0146] To further eliminate misclassified non-road areas, this embodiment of the invention performs shape filtering; specifically, the road binary map is first analyzed for connectivity, dividing it into connected regions, and then shape filtering is performed.

[0147] Shape filtering is performed using area A, fullness F, and linear feature index (LFI), calculated as follows:

[0148]

[0149] In the formula, num is the number of pixels in the connected region, L is the diagonal length of the smallest bounding rectangle of the connected region, and A MER This represents the area of ​​the minimum bounding rectangle, which is the number of pixels within the minimum bounding rectangle.

[0150] Analysis of the road areas reveals that they are relatively large, and for general road areas, their linear characteristic index is quite high; however, for intersecting road networks, the area fill factor is relatively low. Therefore, the following formula is used for shape filtering:

[0151]

[0152] Where, r shape r represents the value of the binary road image after shape filtering. shape When r = 1, it indicates that the pixel belongs to the road class. shape =0 indicates that the pixel is not a road; T A T F T LFI These are the area, region fullness, and linear feature index thresholds, respectively.

[0153] Please see Figure 5 In a further optional technical solution of the present invention, the step of extracting the road centerline specifically includes:

[0154] First, tensor encoding, tensor transfer, and tensor decomposition are performed on the refined binary road map. The non-maximum suppression algorithm is then used to extract road centerlines from the decomposed bar tensors, at which point the road centerlines, except for intersections, are well extracted. For the spherical tensors, a thresholding method is used to determine intersection locations, and then a linear fitting method is used to fill in the road centerlines at intersections, resulting in a complete and coherent road centerline network. Further optimized, a thresholding method is then used to delete line segments with lengths less than a threshold, yielding the final road centerlines.

[0155] To verify the effectiveness of the proposed method, three remote sensing images with different resolutions and scenes were used in the experiment. Images 1 and 2 were selected from Vplab, with an image resolution of 0.6 m / pixel and an image size of 512*512. Image 3 was selected from the dataset published by Tureken et al., with an image resolution of 1.2 m / pixel and an image size of 861*700. Three comparison algorithms were used, including SVM-LLKR, SLIC-Kmeans, and RF-DS.

[0156] The algorithm proposed in this embodiment of the invention first preprocesses the image using bilateral filtering, and then extracts pixel-level features and multi-scale object-level features respectively; then it selects a training set using the GAN method, and obtains the probability matrix of the road network using OCSVM and Plattscaling methods; then it fuses the results to form a coarse extraction result of the road network, and performs fine processing and centerline extraction on the coarse extraction result.

[0157] Table 1. Experimental Parameters

[0158]

[0159] For example, the experimental operating system was Windows 11 and the programming environment was MATLAB R2021b. Based on experimental experience, the uniformity tolerance m was selected as 15, the expanded training region was selected as a square structuring element with a size of 20, and the window size of the local Moran index was selected as 15. Other experimental parameters are shown in Table 1.

[0160] To better evaluate the effectiveness of the extraction results, this invention conducts experimental analysis on the fused image from both subjective and objective perspectives.

[0161] Please see Figures 6 to 9 The diagram shows the experimental process of each method on various images, as well as the final extraction results of each method. Visual observation reveals that the classifier proposed in this embodiment is well-suited for road network extraction tasks, achieving high accuracy and completeness. Experimental analysis of other comparative methods shows that applying only pixel-level features easily results in road surface holes; applying only object-level features easily leads to missing road blocks and adhesion between road and non-road areas. This paper fuses the results obtained at the pixel level and multi-scale object level, effectively solving these problems. Furthermore, the proposed method requires fewer parameter choices, making it easier to implement. In the centerline extraction stage, traditional morphological thinning operations produce many spikes, methods based on local linear kernel regression are ineffective for curved roads, and tensor voting algorithms sometimes easily produce spikes and breaks. The method used in this paper, however, can obtain complete and smooth centerlines.

[0162] Objective evaluation uses three commonly used metrics—completeness, precision, and detection quality—to quantitatively assess the effectiveness of different algorithms. The formulas for calculating these three metrics are as follows:

[0163]

[0164] In the formula, TP represents the number of road pixels identified as roads, FN represents the number of road pixels identified as non-road pixels, and FP represents the number of non-road pixels identified as roads. COM represents the completeness rate, reflecting the proportion of pixels identified as roads out of all pixels; COR represents the accuracy rate, reflecting the proportion of road pixels among those identified as roads; and Q represents the quality, a metric that balances completeness and accuracy. When quantitatively evaluating road centerlines, a buffer width is often set. If the centerline extracted by the algorithm is within the buffer radius, it is considered a correctly extracted road centerline; otherwise, it is considered an incorrectly extracted centerline. Experimental results are shown in Table 2.

[0165] Table 2. Experimental Results

[0166]

[0167] This invention discloses a road extraction method for remote sensing images based on OCSVM and multi-level feature fusion. First, bilateral filtering is applied to the image to reduce noise while preserving edge information. Second, pixel-level and object-level features are introduced to provide a more comprehensive description of the road. Next, the selection method and output of the OCSVM training set are optimized to make the training set cleaner and the output a soft classification. Then, multi-level fusion is performed to alleviate the problems of holes and concatenation in the extraction results caused by insufficient road description using a single level of features. Finally, post-processing is used to refine the road details and extract the road centerline. Experimental results show that this method can effectively improve the accuracy and completeness of road extraction.

[0168] The following are embodiments of the apparatus of the present invention, which can be used to execute embodiments of the method of the present invention. For details not disclosed in the apparatus embodiments, please refer to the embodiments of the method of the present invention.

[0169] Please see Figure 10 In another embodiment of the present invention, a road network extraction system based on remote sensing imagery is provided, comprising:

[0170] The training set acquisition module is used to acquire pixel-level features and object-level features at multiple scales for each pixel based on remote sensing images; and to extract multiple pixels that meet the requirements based on the remote sensing images to obtain a training set; wherein, pixel-level features are formed by superimposing spectral features and gradient features, and object-level features are formed by superimposing spectral features and local Moran's index features.

[0171] The decision value acquisition module is used to solve OCSVM based on the pixel-level features and multi-scale object-level features of each pixel in the training set, respectively, to obtain the pixel-level decision function and the object-level decision function; based on the obtained pixel-level decision function and object-level decision function, the pixel-level decision value and the object-level decision value of each pixel in the remote sensing image are obtained.

[0172] The binary image acquisition module is used to convert the pixel-level decision value and object-level decision value of each pixel into the probability that each pixel belongs to the road at the pixel level and object level; and to fuse the probabilities that each pixel belongs to the road at the pixel level and object level to obtain a binary image of the road image.

[0173] The road network extraction module is used to extract the centerline based on the binary image of the road image to obtain the road centerline network.

[0174] In another embodiment of the present invention, a computer device is provided, comprising a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions from the computer storage medium to achieve a corresponding method flow or corresponding function. The processor described in this embodiment of the present invention can be used for the operation of a road network extraction method based on remote sensing imagery.

[0175] In another embodiment of the present invention, a storage medium is provided, specifically a computer-readable storage medium (Memory), which is a memory device in a computer device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the road network extraction method based on remote sensing imagery in the above embodiments.

[0176] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0177] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0178] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0179] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0180] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A road network extraction method based on remote sensing imagery, characterized in that, Includes the following steps: Based on remote sensing images, obtain pixel-level features of each pixel and object-level features at multiple scales; Multiple qualified pixels are extracted from the remote sensing image to obtain a training set; wherein, pixel-level features are formed by superimposing spectral features and gradient features, and object-level features are formed by superimposing spectral features and local Moran index features. Based on the pixel-level features and multi-scale object-level features of each pixel in the training set, OCSVM is used to solve the problem to obtain the pixel-level decision function and the object-level decision function. Based on the obtained pixel-level decision function and object-level decision function, the pixel-level decision value and object-level decision value of each pixel in the remote sensing image are obtained. The obtained pixel-level decision values ​​and object-level decision values ​​of each pixel are converted into the probability that each pixel belongs to the road at the pixel level and object level; the obtained probabilities that each pixel belongs to the road at the pixel level and object level are fused to obtain a binary image of the road image. Centerlines are extracted from the binary image of the road image to obtain a road centerline network; in, In the step of obtaining the pixel-level decision function and the object-level decision function respectively by using OCSVM to solve for the pixel-level features and multi-scale object-level features of each pixel in the training set,... For the training set pixels in Pixel-level features and multi-scale object-level features are represented as follows: In the formula, The pixel-level feature at pixel i3 This represents the object-level feature at pixel i3 at scale k, where k represents different scales. These represent the spectral features, gradient features, and local Moran index features at pixel i3, respectively. Solve using OCSVM respectively, the solution expression is, In the formula, ω is the weight of the support vector; ρ is the threshold of the support vector. is a slack variable used to indicate the allowed existence of outliers; v∈(0,1) is a parameter used to control the upper limit of the number of outliers and the lower limit of the total number of support vectors; N represents the number of support vectors; φ represents the feature vector of pixel i3; φ is a mapping function used to map data samples to a higher-dimensional feature space, making them linearly separable. During the solution process, the hyperplane parameters ω at different levels are obtained by using features at different levels. pixel ρ pixel ω object k ρ object k And construct the decision functions for each level, expressed as follows: In the formula, f i pixel f represents the decision value at pixel level i; i object k ω represents the decision value of pixel i at scale k for the object hierarchy; pixel ρ pixel ω represents the weights and thresholds of the support vectors at the pixel level, respectively; object k ρ object k φ represents the weights and thresholds of the support vectors in the object hierarchy at scale k, respectively; φ is the mapping function.

2. The road network extraction method based on remote sensing imagery according to claim 1, characterized in that, The process involves acquiring pixel-level features for each pixel and object-level features at multiple scales based on remote sensing imagery. The remote sensing image is an image obtained after bilateral filtering preprocessing based on the original remote sensing image.

3. The road network extraction method based on remote sensing imagery according to claim 1, characterized in that, The step of extracting multiple qualified pixels from the remote sensing image to obtain a training set includes: Based on the remote sensing image, pixels are extracted using an adaptive domain-based method to obtain a training set; Among them, the method of extracting pixels using adaptive neighborhood is described. For a pixel i∈D, its associated adaptive domain GAN is defined as a subset contained in D. Adaptive domain GANs are built on a standard mapping h∈c, which is related to the uniformity tolerance m of the range on E; where, The adaptive domain GAN is represented as follows: In the formula, C X (i) represents the connected component containing pixel i∈D. E, c, and D represent the uniformity tolerance m, the standard mapping h, and the space where pixel i in the image resides, respectively. -1 ([h(i)-m],[h(i)+m]) represents pixel i that satisfies the standard mapping conditions; Based on adaptive domain GANs, the selected training set is represented as follows: In the formula, g represents the grayscale feature of the image. i is the seed point, and i2 represents the index of a different seed point.

4. The road network extraction method based on remote sensing imagery according to claim 1, characterized in that, In the step of converting the obtained pixel-level decision values ​​and object-level decision values ​​of each pixel into the probability that each pixel belongs to the road at the pixel level and object level, The Platt scaling method is used to convert decision values ​​into probability values, thereby obtaining the probability that each pixel belongs to a road at each level, forming a coarse road extraction result at each level; wherein, the coarse road extraction result at each level includes the sub-road coarse extraction result at the object level and the road coarse extraction result at the pixel level at multiple scales.

5. The road network extraction method based on remote sensing imagery according to claim 4, characterized in that, The step of fusing the probabilities of each pixel belonging to the road at the pixel level and the object level to obtain a binary image of the road image includes: A mean fusion strategy is adopted to fuse the coarse extraction results of sub-roads at the object level under multiple scales to form a coarse extraction result at the object level. The coarse extraction results at the pixel level and the object level will be merged to form the final coarse road extraction result; Based on the final coarse road extraction result, fine processing is performed to obtain a binary image of the road image; The step of fusing the coarse extraction results at the pixel level and the object level includes: For the pixel-level probability matrix P pixel With the object-level probability matrix P object The fusion will adopt a decision fusion based on multi-threshold processing, and the fusion rule is as follows: In the formula, r represents the binary image value of the road image; when r = 1, it indicates that the pixel is a road, and when r = 0, it indicates that the pixel is not a road. These represent the high and low thresholds used at the object level and pixel level, respectively.

6. The road network extraction method based on remote sensing imagery according to claim 5, characterized in that, The refinement process includes one or more of edge filtering, hole filling, and shape filtering.

7. A road network extraction system based on remote sensing imagery, characterized in that, include: The training set acquisition module is used to acquire pixel-level features and object-level features at multiple scales for each pixel based on remote sensing images. Multiple qualified pixels are extracted from the remote sensing image to obtain a training set; wherein, pixel-level features are formed by superimposing spectral features and gradient features, and object-level features are formed by superimposing spectral features and local Moran index features. The decision value acquisition module is used to solve OCSVM based on the pixel-level features and multi-scale object-level features of each pixel in the training set, respectively, to obtain the pixel-level decision function and the object-level decision function; based on the obtained pixel-level decision function and object-level decision function, the pixel-level decision value and the object-level decision value of each pixel in the remote sensing image are obtained. The binary image acquisition module is used to convert the pixel-level decision value and object-level decision value of each pixel into the probability that each pixel belongs to the road at the pixel level and object level; and to fuse the probabilities that each pixel belongs to the road at the pixel level and object level to obtain a binary image of the road image. The road network extraction module is used to extract centerlines from the binary image of the road image to obtain a road centerline network; in, In the step of obtaining the pixel-level decision function and the object-level decision function respectively by using OCSVM to solve for the pixel-level features and multi-scale object-level features of each pixel in the training set,... For the training set pixels in Pixel-level features and multi-scale object-level features are represented as follows: In the formula, The pixel-level feature at pixel i3 This represents the object-level feature at pixel i3 at scale k, where k represents different scales. These represent the spectral features, gradient features, and local Moran index features at pixel i3, respectively. Solve using OCSVM respectively, the solution expression is, In the formula, ω is the weight of the support vector; ρ is the threshold of the support vector. is a slack variable used to indicate the allowed existence of outliers; v∈(0,1) is a parameter used to control the upper limit of the number of outliers and the lower limit of the total number of support vectors; N represents the number of support vectors; φ represents the feature vector of pixel i3; φ is a mapping function used to map data samples to a higher-dimensional feature space, making them linearly separable. During the solution process, the hyperplane parameters ω at different levels are obtained by using features at different levels. pixel ρ pixel ω object k ρ object k And construct the decision functions for each level, expressed as follows: In the formula, f i pixel f represents the decision value at pixel level i; i object k ω represents the decision value of pixel i at scale k for the object hierarchy; pixel ρ pixel ω represents the weights and thresholds of the support vectors at the pixel level, respectively; object k ρ object k φ represents the weights and thresholds of the support vectors in the object hierarchy at scale k, respectively; φ is the mapping function.

8. An electronic device, characterized in that, include: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the road network extraction method based on remote sensing imagery as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the road network extraction method based on remote sensing imagery as described in any one of claims 1 to 6.

Images