Image ground feature attribute labeling method and device

Abstract

The invention discloses an image ground feature attribute labeling method and device. The method comprises the following steps: selecting a first initial growth point from an image area of a target ground object in a to-be-labeled image; based on a selection region growth algorithm and a similarity measurement criterion, performing first region growth from the first initial growth point to obtain a primary growth result of the target ground object image region; if it is determined that an area which is not accurately recognized exists in the to-be-labeled image according to the primary growth result, selecting a second initial growth point from the area which is not accurately recognized to perform secondary growth, and obtaining a target ground object image area correction growth result; integrating the initial growth result and the corrected growth result to obtain a target ground object image area; and filling the holes in the target ground object image area, obtaining the contour of the filled target ground object image area, and obtaining a target ground object marking area. According to the method and the device, interactive quick marking of the ground feature attributes can be realized, and the marking efficiency of the sample data is improved.

Description

Technical Field

[0001] The present invention relates to the field of seismic exploration acquisition, and particularly to a method and device for annotating image ground object attributes. Background Art

[0002] In oil and gas exploration tasks, the calibration of the main ground object attribute information of obstacles based on satellite image data, such as roads, water bodies, houses, etc., is an important link in seismic acquisition design and construction, and is the basic data for subsequent observation system design and shot point layout. Therefore, accurately and efficiently identifying road data in remote sensing images is of great significance for field data acquisition, namely, realizing physical point layout, obstacle avoidance, and acquisition operations, and is the basis for realizing path planning tasks and the intelligence of oil and gas exploration.

[0003] Currently, the road recognition technology based on remote sensing image data is mainly the deep neural network extraction method, and the performance of the deep neural network highly depends on the quantity and accuracy of training sample data. The ground object attribute training sample data of remote sensing images directly determines the accuracy of identifying obstacle information in the exploration area, so it has an important impact on seismic exploration acquisition construction. In addition, the current deep neural network method has poor generalization ability, that is, different data sets need to be used for training when identifying roads in different spatial scenarios. These two factors lead to the need to produce a large number of labeled images before using the deep neural network method for training. A large number of labeled remote sensing images can train a more accurate deep neural network with more reasonable segmentation results, and a better-performing deep neural network can quickly and accurately segment a large number of remote sensing images, thereby enabling the construction team to have an overall understanding of the road network structure of the entire area, which is the basic data for specific tasks such as construction path planning and shot point layout in obstacle areas in the acquisition design and construction links.

[0004] Currently, in the oil and gas exploration acquisition design link, the methods for making image labels include manually selecting the annotation area and automatically annotating the image through an annotation algorithm model. Summary of the Invention

[0005] When using the method of manually selecting the annotation area of remote sensing images to make image labels, there are problems such as high work repetition, low manual annotation efficiency, and the annotation accuracy cannot meet the project requirements. When using the method of annotating images through an annotation algorithm model, a large number of already annotated image samples often need to be collected, which increases the workload, and there is a problem that the automatic annotation method cannot accurately locate the annotation target in the image, which will cause an error between the final obtained annotation result and the actual result, resulting in inaccurate annotation results. All in all, the existing image annotation methods have problems such as low annotation accuracy, large workload, and low work efficiency.

[0006] In view of the above problems, the present invention is proposed to provide an image object attribute annotation method and device that overcome the above problems or at least partially solve the above problems.

[0007] In a first aspect, an embodiment of the present invention provides an image object attribute annotation method, including:

[0008] Select a first starting growth point from the image area of the target object in the image to be annotated;

[0009] Based on the region growing algorithm of selective regions and the similarity measurement criterion, perform the first region growing starting from the first starting growth point to obtain the initial growth result of the target object image area;

[0010] If it is determined according to the initial growth result that there are unrecognized regions in the image to be annotated, select a second starting growth point from the unrecognized regions for secondary growth to obtain the corrected growth result of the target object image area;

[0011] Integrate the initial growth result and the corrected growth result to obtain the target object image area;

[0012] Fill the holes in the target object image area, obtain the contour of the filled target object image area, and obtain the target object annotation area.

[0013] In some optional embodiments, based on the region growing algorithm of selective regions and the similarity measurement criterion, performing the first region growing starting from the first starting growth point to obtain the initial growth result of the target object image area includes:

[0014] Traverse the points to be grown in the growth neighborhood of the first starting growth point;

[0015] Determine the similarity between the points to be grown and the first starting growth point according to the similarity measurement criterion, add the points to be grown that meet the growth criterion to the first growth domain set, use the newly added points to be grown as the new first starting growth point, and return to continue executing the step of traversing the points to be grown in the growth neighborhood of the first starting growth point until there are no points to be grown that meet the growth criterion, and obtain the initial growth result of the target object image area.

[0016] In some optional embodiments, traversing the points to be grown in the growth neighborhood of the first starting growth point; determining the similarity between the points to be grown and the first starting growth point according to the similarity measurement criterion, and adding the points to be grown that meet the growth criterion to the first growth domain set includes:

[0017] Taking the first starting growth point as the center, determine the neighborhood of the first starting growth point;

[0018] Determine the similarity Dissimilarity between the to-be-grown point in the neighborhood and the first starting growth point according to the following similarity measurement rules:

[0019] Dissimilarity(p1, p2) = λ color dis color (p1, p2) + λ local dis local (p1, p2),

[0020] where λ color is the weight for adjusting the spatial balance degree, and λ local is the weight for adjusting the color balance degree. dis color represents the modulus of the difference of the vectors in the form of 1×1×3 corresponding to points p1 and p2, and dis local represents the distance between the coordinates of the two points in the image;

[0021] Traverse the to-be-grown points in the growth neighborhood of the first starting growth point. If the similarity between the to-be-grown point and the first starting growth point is less than the pre-determined growth threshold of the first starting growth point, add the to-be-grown point to the first growth domain set S.

[0022] In some optional embodiments, determining the growth threshold of the first starting growth point includes:

[0023] Based on the digital characteristics of the three-dimensional matrix corresponding to the to-be-annotated image, determine the growth threshold MaxSim of the first starting growth point according to the following formula: where Std image represents the standard deviation of the three-dimensional matrix corresponding to the image after mean shift.

[0024] In some optional embodiments, if it is determined according to the initial growth result that there are regions in the to-be-annotated image that are not accurately recognized, select a second starting growth point from the regions that are not accurately recognized for secondary growth to obtain the corrected growth result of the target ground object image region, including:

[0025] Judge whether there are regions in the to-be-annotated image that are not accurately recognized according to the initial growth result; the regions that are not accurately recognized include regions of missed recognition and / or over-recognition;

[0026] If so, select a second starting growth point from each region that is not accurately recognized, and based on the selected region growth algorithm and similarity measurement criteria, start the second region growth from the second starting growth point to obtain the corrected growth result of the target ground object image region.

[0027] In some alternative embodiments, the region that is not accurately recognized is an unrecognized region. A second starting growth point is selected from each region that is not accurately recognized. Based on the selective region growing algorithm and the similarity metric criterion, a second region growing is performed starting from the second starting growth point to obtain the corrected growth result of the target ground object image region, including:

[0028] A second starting growth point is selected from the unrecognized region, and the points to be grown in the growth neighborhood of the second starting growth point are traversed. If the dissimilarity between the point to be grown in the growth neighborhood of the second starting growth point and the second starting growth point is less than the growth threshold MaxSim of the second starting growth point in the unrecognized region, the point to be grown is added to the second growth domain set;

[0029] The newly added point to be grown is used as the new second starting growth point, and the step of traversing the points to be grown in the growth neighborhood of the second starting growth point is continued until there are no new points to be grown in the second growth domain set, and the corrected growth process ends, obtaining the corrected growth result of the target ground object image region.

[0030] In some alternative embodiments, the region that is not accurately recognized is an over-recognized region. A second starting growth point is selected from each region that is not accurately recognized. Based on the selective region growing algorithm and the similarity metric criterion, a second region growing is performed starting from the second starting growth point to obtain the corrected growth result of the target ground object image region, including:

[0031] A second starting growth point is selected in the over-recognized region, and the points to be grown in the growth neighborhood of the second starting growth point are traversed. If the dissimilarity between the point to be grown in the growth neighborhood of the second starting growth point and the second starting growth point is less than the growth threshold newMaxSim of the second starting growth point in the over-recognized region, the point to be grown is added to the third growth domain set;

[0032] The newly added point to be grown is used as the new second starting growth point, and the step of traversing the points to be grown in the growth neighborhood of the second starting growth point is continued until there are no new points to be grown in the third growth domain set, and the corrected growth process ends, obtaining the corrected growth result of the target ground object image region;

[0033] The growth threshold newMaxSim of the second starting growth point in the over-recognized region is determined according to the following formula: newMaxSim = MaxSim * α times , where α is a correction coefficient.

[0034] In some alternative embodiments, the region that is not accurately recognized includes an unrecognized region and / or an over-recognized region. The initial growth result and the corrected growth result are integrated to obtain the target ground object image region, including:

[0035] For the missed recognition region, add the initial growth result to the corrected growth result of the missed recognition region;

[0036] For the over-recognition region, subtract the corrected growth result of the over-recognition region from the initial growth result;

[0037] Based on the result after addition and / or subtraction, obtain the target ground object image region.

[0038] In some alternative embodiments, filling the holes in the target ground object image region, obtaining the contour of the filled target ground object image region, and obtaining the target ground object annotation region, including:

[0039] According to the connected region area of the target ground object image region and a preset filling threshold, determine the connected regions to be filled, perform hole filling based on the connected regions to be filled, perform edge extraction on the filled target ground object image region, extract the contour of the target ground object image region, and obtain the target ground object annotation region according to the contour of the target ground object image region.

[0040] In some alternative embodiments, according to the connected region area of the target ground object image region and a preset filling threshold, determining the connected regions to be filled, including:

[0041] According to the closing operation method, perform contour processing on the image to be annotated and construct a structure body of the target ground object image region with a size of k×k, perform connected region search on the processed image and calculate the area of each connected region;

[0042] Determine the preset filling threshold according to the size of the image to be annotated and the size of the structure body of the target ground object image region;

[0043] Determine the connected regions with an area smaller than the preset filling threshold in the connected regions as the connected regions to be filled.

[0044] In some alternative embodiments, determining the preset filling threshold according to the size of the image to be annotated and the size of the structure body of the target ground object image region, including:

[0045] Determine the preset filling threshold limit according to the following formula: where k is the structure body size, H is the height of the input image, W is the width of the input image, and λ fill is the weight coefficient for adjusting the image size and the structure body size.

[0046] In some alternative embodiments, performing edge extraction on the filled target ground object image region and extracting the contour of the target ground object image region, including:

[0047] Perform Gaussian filtering on the filled target ground object image region;

[0048] Determine the gradient of the target ground object image region after Gaussian filtering according to the edge detection algorithm, and construct the edge gradient and corrected gradient direction of the image region;

[0049] After correcting the gradient direction of the image region according to the corrected gradient direction, perform non-maximum suppression on the gradient of the target ground object image region to extract the edge information of the target ground object image region;

[0050] Perform double-threshold selection on the edge information of the extracted target ground object image region according to the edge gradient to extract the contour of the target ground object image region.

[0051] In some alternative embodiments, constructing the edge gradient and corrected gradient direction of the image includes:

[0052] According to the gradient in the horizontal direction and the gradient in the vertical direction of the image, construct the edge gradient Edge Gradient of the image based on the following formula: where G x represents the gradient in the horizontal direction of the image, and G y represents the gradient in the vertical direction of the image;

[0053] According to the gradient in the horizontal direction and the gradient in the vertical direction of the image, construct the corrected gradient direction Edge Angle of the image based on the following formula: where G x represents the gradient in the horizontal direction of the image, and G y represents the gradient in the vertical direction of the image;

[0054] Performing non-maximum suppression on the gradient of the target ground object image region includes:

[0055] For each pixel point in the image region, search for the gradient values of its adjacent pixel points in the corresponding gradient direction and compare them with its own gradient value;

[0056] If the pixel point gradient value is greater than or equal to the adjacent pixel point gradient value, retain the pixel point gradient value, otherwise change the pixel point gradient value to 0;

[0057] Performing double-threshold selection on the edge information of the extracted target ground object image region according to the edge gradient includes:

[0058] If the edge gradient value of the pixel point is higher than the maximum gradient Max Gradient, regard the pixel point as a true boundary point

[0059] If the edge gradient of the pixel point is lower than the preset minimum gradient Min Gradient, regard the pixel point as a non-boundary point;

[0060] If the edge gradient of a pixel is between the preset maximum gradient Max Gradient and the preset minimum gradient Min Gradient, the pixel is regarded as a false boundary point;

[0061] Search for the pixels in the neighborhood of the false boundary point. If the false boundary point is connected to a true boundary point, the false boundary point is regarded as a true boundary point; if the false boundary point is not connected to a true boundary point, the false boundary point is discarded.

[0062] In some optional embodiments, obtaining the target object annotation area according to the contour of the target object image area includes:

[0063] Use the Douglas-Peucker algorithm to perform polygon fitting on the contour of the extracted target object image area to obtain the polygon fitting result, and perform region growing on the polygon fitting result to generate the target object annotation area.

[0064] In a second aspect, an image object attribute annotation device provided by an embodiment of the present invention includes:

[0065] A region growing module, configured to select a first starting growth point from the image area of the target object in the image to be annotated; based on the selected region growing algorithm and the similarity metric criterion, perform the first region growing starting from the first starting growth point to obtain the initial growth result of the target object image area; if it is determined according to the initial growth result that there is an area in the image to be annotated that is not accurately recognized, select a second starting growth point from the area that is not accurately recognized for secondary growth to obtain the corrected growth result of the target object image area;

[0066] A region determining module, configured to integrate the initial growth result and the corrected growth result to obtain the target object image area; fill the holes in the target object image area to obtain the contour of the filled target object image area, and obtain the target object annotation area.

[0067] An embodiment of the present invention provides a computer storage medium, in which computer-executable instructions are stored, and when the computer-executable instructions are executed by a processor, the above-mentioned image object attribute annotation method is implemented.

[0068] An embodiment of the present invention provides a computer device, including: a memory, a processor, and a computer program stored on the memory and executable on the processor, and when the processor executes the program, the above-mentioned image object attribute annotation method is implemented.

[0069] The beneficial effects of the above technical solutions provided by the embodiments of the present invention at least include:

[0070] This method selects a first starting growth point from the image region of the target ground object in the image to be labeled; based on the region growing algorithm and the similarity measurement criterion, the first region growing is carried out starting from the first starting growth point to obtain the initial growth result of the target ground object image region; the adopted similarity measurement criterion re-measures the similarity between pixel points according to factors such as the actual size of the remote sensing image and the pixel distribution of the image itself, improving the accuracy of image region recognition.

[0071] After the first growth is completed, if it is determined according to the initial growth result that there are regions in the image to be labeled that are not accurately recognized, a second starting growth point is selected from the regions that are not accurately recognized for secondary growth to obtain the corrected growth result of the target ground object image region; the initial growth result and the corrected growth result are integrated to obtain the target ground object image region. Compared with the method that ends after only one region growth, the method of secondary growth can accurately recognize the regions that are not accurately recognized, making the finally recognized image region more complete and accurate. When the region growing method is used for recognition, due to the existence of noise points and obstacles in the image, there may be some holes in the recognition result, that is, some unrecognized regions. In addition, the region growing method does not perform special processing on the contour edges, so the contour edges are often not smooth enough. To solve these problems, after the growth process is completed, the holes in the target ground object image region are filled, the contour of the filled target ground object image region is obtained, and the target ground object labeling region is obtained. After the hole filling process, the edge of the recognition result can be made smoother, and finally a region segmentation result with high accuracy, fast labeling rate, and strong interpretability is obtained.

[0072] Other features and advantages of the present invention will be described in the following specification, and, in part, will be obvious from the specification, or will be understood by implementing the present invention. The objectives and other advantages of the present invention can be achieved and obtained by the structures specifically pointed out in the written specification, claims, and drawings.

[0073] The technical solutions of the present invention will be further described in detail below with reference to the drawings and embodiments. Description of the Drawings

[0074] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention, and do not constitute a limitation to the present invention. In the drawings:

[0075] Figure 1 is the flowchart of the method for labeling the attributes of image ground objects in the embodiments of the present invention;

[0076] Figure 2 is the original remote sensing image in the embodiments of the present invention;

[0077] Figure 3 This is a sample diagram of a remote sensing image in an embodiment of the present invention;

[0078] Figure 4 This is a local sample diagram before mean shift in an embodiment of the present invention;

[0079] Figure 5 This is the mean shift result diagram of h s = h r = 12 in an embodiment of the present invention;

[0080] Figure 6 This is the mean shift result diagram of h s = h r = 30 in an embodiment of the present invention;

[0081] Figure 7 This is the initial growth result diagram of the image without mean shift in an embodiment of the present invention;

[0082] Figure 8 This is the initial growth result diagram after mean shift with h s = h r = 12 in an embodiment of the present invention;

[0083] Figure 9 This is the initial growth result diagram after mean shift with h s = h r = 30 in an embodiment of the present invention;

[0084] Figure 10 This is the result diagram of the first region growth before performing closing operation on the original image in an embodiment of the present invention;

[0085] Figure 11 This is the result diagram of the first region growth after performing closing operation on the original image in an embodiment of the present invention;

[0086] Figure 12 This is the region growth result after hole filling in an embodiment of the present invention;

[0087] Figure 13 This is the result of edge extraction on the filled image in an embodiment of the present invention;

[0088] Figure 14 This is the result diagram of re-growing by fitting polygons to the extracted edges in an embodiment of the present invention;

[0089] Figure 15 This is the final ground object attribute annotation area in an embodiment of the present invention;

[0090] Figure 16 This is the structural schematic diagram of the image ground object attribute annotation device in an embodiment of the present invention. Detailed implementation manners

[0091] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be completely conveyed to those skilled in the art.

[0092] When using the method of manually selecting the annotation area of remote sensing images to make image labels, there are problems such as high work repetition, low manual annotation efficiency, and the annotation accuracy cannot reach the required accuracy of the project; when using the annotation algorithm model to annotate images, a large number of annotated image samples often need to be collected, which increases the workload, and there is a problem that the automatic annotation method cannot accurately locate the annotation targets in the images, which will cause errors between the final obtained annotation results and the actual results. To solve the problems of low annotation efficiency and inaccurate annotation area recognition in the prior art, an image ground object attribute annotation method is provided in an embodiment of the present invention.

[0093] An embodiment of the present invention provides an image ground object attribute annotation method, and its process is as Figure 1 shown, including the following steps:

[0094] Step S101: Select a first starting growth point from the image area of the target ground object in the image to be annotated;

[0095] Step S102: Based on the region growing algorithm and the similarity measurement criterion, start the first region growing from the first starting growth point to obtain the initial growth result of the target ground object image area;

[0096] Step S103: If it is determined according to the initial growth result that there are unaccurately recognized areas in the image to be annotated, select a second starting growth point from the unaccurately recognized areas for secondary growth to obtain the corrected growth result of the target ground object image area;

[0097] Step S104: Integrate the initial growth result and the corrected growth result to obtain the target ground object image area;

[0098] Step S105: Fill the holes in the target ground object image area, obtain the contour of the filled target ground object image area, and obtain the target ground object annotation area.

[0099] Preferably, before performing the above step S101, the collected image data needs to be processed by mean shift, and the specific processing process is as follows:

[0100] For the data sample X of the image to be annotated from n independent and identically distributed data sets (t), define the radially symmetric kernel function Gaussian kernel function K(x):

[0101] K(x) = c k k(||x|| 2 ), (1)

[0102] where: c k is the normalization parameter, k is the standard Gaussian kernel function, x represents the vector from the data point to the kernel center, here it is a D-dimensional vector, representing a point in the data space.

[0103] The data sample X of the above labeled image (t) ∈R d , X (t) represents the image state at the t-th step of the iteration. In the mean shift algorithm, t usually refers to the number of iterations. For image processing, during each iteration, some features of the image, such as pixel values, color distributions, etc., will change according to the algorithm rules. R d represents a d-dimensional real space, where d represents the number of features of each considered data point (usually pixels or groups of pixels in the case of image processing). Each dimension usually corresponds to a feature.

[0104] The kernel density function estimate corresponding to the data sample is

[0105]

[0106] Substitute formula (1) into formula (2), that is

[0107] where: is the kernel density function, h is the Gaussian kernel function bandwidth parameter, d is a d-dimensional vector, and X i is the i-th data point in the dataset.

[0108] The local mode of the kernel density function is the local maximum. Therefore, take the partial derivative of formula (3), denote the derivative of k(x) as k’(x), and denote g(x) = -k’(x) to obtain the gradient expression at this time:

[0109]

[0110]

[0111] Equation (5) represents the gradient of the kernel density estimate at X(t), and this gradient points in the direction of increasing probability density, which is used to guide the movement of data points in the mean shift algorithm.

[0112] At this time, the last term in equation (5) is the mean shift term, denoted as m(X (t) ). That is:

[0113]

[0114] The iterative process of offsetting the center point is as follows:

[0115] X (t+1) = X (t) + m(X (t) ). (7)

[0116] When the iterative process stops and X (t+1) = X (t) i.e., when the drift point no longer changes, the operation process ends.

[0117] Specifically, for visible light images, the original image can be regarded as H×W sample points. Therefore, the mean shift method can also be used for clustering. The method of performing mean shift on visible light images is to use the values (x, y) corresponding to the first two dimensions of each pixel point in the matrix of the original RGB image H×W×3 as the physical position information of the point, which is represented in the spatial domain. The value corresponding to the third dimension is used as the color information of the point, which is represented in the range domain. For these two domains, the Euclidean distance is used to measure the distance between different points. Therefore, when considering the position vector and the color vector together in the 5D spatial-color joint domain (the number of dimensions is equal to the number of channels plus 2), a reasonable normalization method is needed to compensate for the different distance information brought by different dimensions of space and color. Its canonical form is:

[0118]

[0119] In the formula: X s is the spatial part of the feature vector, X r is the color part of the feature vector, k(x) is the common profile used in both domains, h s and h r are the kernel bandwidths used, and C is the normalization constant.

[0120] As the color kernel bandwidth h r increases, the detailed information in the image is continuously lost. However, if the mean shift is not performed on the image, the segmentation result will be significantly missing. Although mean shift can cluster adjacent regions with similar colors, the number of segmentation categories after clustering is too large because mean shift cannot specify the number of final clustering categories. Therefore, the mean shift algorithm cannot be used as the road recognition result, but it can assist in the design of a semi-automatic annotation system. The images in this embodiment are all based on the images after mean shift for growth and recognition.

[0121] In this embodiment, taking the recognition of the road part in the remote sensing image as an example, the accuracy of this method is verified. Figure 2 is the original remote sensing image, Figure 3 is from Figure 2 the selected sample image, Figure 4 is the local sample image before mean shift, Figure 5 is h s = h r = 12 of the mean shift result image, Figure 6 is h s = h r = 30 of the mean shift result image. It can be seen from the figure that the smaller the bandwidth h, the less smooth the obtained probability density function is and the more noise it contains; if the selected width is too large, the generated probability density function will be overly smooth and contain less details.

[0122] Preferably, in the above step S101, based on the region growing algorithm of selection and the similarity measurement criterion, the first region growing is carried out starting from the first starting growth point to obtain the initial growth result of the target ground object image region, including:

[0123] Traverse the points to be grown in the growth neighborhood of the first starting growth point;

[0124] According to the similarity measurement criterion, determine the similarity between the points to be grown and the first starting growth point, add the points to be grown that meet the growth criterion to the first growth domain set, take the newly added points to be grown as the new first starting growth point, and return to continue to execute the step of traversing the points to be grown in the growth neighborhood of the first starting growth point until there are no points to be grown that meet the growth criterion, and obtain the initial growth result of the target ground object image region.

[0125] Preferably, traverse the points to be grown in the growth neighborhood of the first starting growth point; according to the similarity measurement criterion, determine the similarity between the points to be grown and the first starting growth point, and add the points to be grown that meet the growth criterion to the first growth domain set, including:

[0126] Taking the first starting growth point as the center, determine the neighborhood of the first starting growth point;

[0127] Determine the similarity Dissimilarity between the points to be grown in the neighborhood and the first starting growth point according to the following similarity measurement rule:

[0128] Dissimilarity(p1,p2) = λ color dis color (p1,p2) + λ local dis local (p1,p2),

[0129] where λcolor The weight for adjusting the spatial balance degree, λ local The weight for adjusting the color balance degree, dis color Represents the modulus of the difference between the vectors in the form of 1×1×3 corresponding to two points p1 and p2, dis local Represents the distance between the coordinates of two points in the image;

[0130] Traverse the points to be grown in the growth neighborhood of the first starting growth point. If the similarity between the point to be grown and the first starting growth point is less than the pre-determined growth threshold of the first starting growth point, add the point to be grown to the first growth domain set S.

[0131] For the region growing method, the growth result is only related to three parts: the starting growth point, the similarity judgment criterion, and the size of the search range. Compared with the traditional annotation method that requires careful marking of the road contour, for the region growing method, the operator only needs to arbitrarily select a pixel point belonging to the road in a large connected domain after mean shift, significantly reducing the annotation workload.

[0132] The basic region growing method only analyzes grayscale images and does not consider the distance relationship between the points in the search neighborhood and the starting point, only considering the pixel value difference. In actual application scenarios, this selection method is too rigid, and there are many cases of missed recognition and over-recognition. Therefore, for this problem, this patent proposes a new similarity measurement method:

[0133] Dissimilarity(p1,p2)=λ color dis color (p1,p2)+λ local dis local (p1,p2),

[0134] After linearly combining the two points as the final similarity evaluation index Dissimilarity, the closer this index is to 0, the higher the similarity between p1 and p2. It can be seen from the formula that for the points to be classified in the neighborhood that are closer to the starting point, their dis local is smaller, so a slightly larger color distance dis color is allowed. In this case, it helps to resist the influence of noise. Even if there are some noise points around the starting point, they can be included in the subsequent growth point set S through this similarity measurement. For the points to be classified in the neighborhood that are farther from the starting point, dis local is larger, so the allowed color distance dis color is smaller, that is, a higher requirement is imposed on the color values of the farther points. They must be close enough to the starting point to be included in the first growth domain set S. This similarity measurement criterion helps to improve the robustness of the region growing algorithm.

[0135] The above-mentioned first set of growth regions is represented as S = {p1|Dissimilarity(p, p1) < MaxSim, (x1, y1) ∈ U}, where p1 is the point to be grown, p is the starting growth point, x1 represents the x-direction coordinate of the point p1 to be grown, y1 represents the y-direction coordinate of p1, and U represents the neighborhood of the starting growth point.

[0136] Preferably, determining the growth threshold of the first starting growth point includes:

[0137] Based on the digital characteristics of the three-dimensional matrix corresponding to the image to be labeled, determine the growth threshold MaxSim of the first starting growth point according to the following formula: where Std image represents the standard deviation of the three-dimensional matrix corresponding to the image after mean shift.

[0138] The selection of the maximum distance MaxSim can be judged using a fixed value, or it can be dynamically measured in combination with the digital characteristics of the H×W×3 matrix corresponding to the image. This selection criterion allows more deviation of pixels when the standard deviation of the image is large, that is, there are obvious color differences in different regions of the image. When the overall standard deviation of the image is small, that is, the similarity between different regions is high, the selection of the growth point is more cautious.

[0139] To solve the problems of noise points or road markings in the image that may occur in the initial road segmentation, the search range of the growth region in this paper is extended from the basic four-neighborhood to the twenty-four-neighborhood, that is, a 5×5 range centered on the starting point. Each point in the range is used as the point to be grown, and its Dissimilarity with the starting point is calculated respectively and compared with the MaxSim corresponding to the given image. If the selected growth range is a (2k + 1)×(2k + 1) matrix centered on the starting point, then the point dis local with the farthest distance from the growth point in the region to be grown is At this time, if the standard deviation of the image is too small, the situation of MaxSim < dis local may occur, so λ local is added to constrain and adjust it.

[0140] Figure 7 For Figure 4 the initial growth result diagram of the image without mean shift, the result is displayed in binary. The light-colored part in the figure is the grown road part. Figure 8 is the initial growth result diagram after mean shift with h s = h r = 12 displayed in binary, Figure 9 For the initial growth result diagram after mean shift with h s = h rThe initial growth result map after the mean shift of the value = 30 is shown in binary. The white part in the figure is the initially grown road part.

[0141] Preferably, in the above step S103, if it is determined from the initial growth result that there are areas in the image to be labeled that are not accurately recognized, then a second starting growth point is selected from the areas that are not accurately recognized for secondary growth to obtain a corrected growth result of the target ground object image area, including:

[0142] Judging whether there are areas in the image to be labeled that are not accurately recognized according to the initial growth result; the areas that are not accurately recognized include missed recognition areas and / or over-recognition areas;

[0143] If so, a second starting growth point is selected from each area that is not accurately recognized, and based on the selective region growing algorithm and the similarity metric criterion, secondary region growth is started from the second starting growth point to obtain a corrected growth result of the target ground object image area.

[0144] Preferably, if the area that is not accurately recognized is a missed recognition area, a second starting growth point is selected from each area that is not accurately recognized, and based on the selective region growing algorithm and the similarity metric criterion, secondary region growth is started from the second starting growth point to obtain a corrected growth result of the target ground object image area, including:

[0145] A second starting growth point is selected from the missed recognition area, and the points to be grown in the growth neighborhood of the second starting growth point are traversed. If the similarity Dissimilarity between the point to be grown in the growth neighborhood of the second starting growth point and the second starting growth point is less than the growth threshold MaxSim of the second starting growth point in the missed recognition area, the point to be grown is added to the second growth domain set;

[0146] The newly added point to be grown is used as the new second starting growth point, and the step of traversing the points to be grown in the growth neighborhood of the second starting growth point is continued until there are no new points to be grown in the second growth domain set, and the corrected growth process ends to obtain a corrected growth result of the target ground object image area.

[0147] Preferably, if the area that is not accurately recognized is an over-recognition area, a second starting growth point is selected from each area that is not accurately recognized, and based on the selective region growing algorithm and the similarity metric criterion, secondary region growth is started from the second starting growth point to obtain a corrected growth result of the target ground object image area, including:

[0148] Select a second starting growth point in the over-identification area, traverse the points to be grown in the growth neighborhood of the second starting growth point. If the dissimilarity between the point to be grown in the growth neighborhood of the second starting growth point and the second starting growth point is less than the growth threshold newMaxSim of the second starting growth point in the over-identification area, add the point to be grown to the third growth domain set;

[0149] Take the newly added point to be grown as the new second starting growth point, and continue to execute the step of traversing the points to be grown in the growth neighborhood of the second starting growth point until there are no new points to be grown in the third growth domain set, and the correction growth process ends, obtaining the correction growth result of the target ground object image area;

[0150] The growth threshold newMaxSim of the second starting growth point in the over-identification area is determined according to the following formula: newMaxSim = MaxSim * α times , where α is the correction coefficient. In this embodiment, the correction coefficient is selected as 0.9, and the correction coefficient can be adjusted dynamically.

[0151] Preferably, in the above step S104, the areas that are not accurately identified include missed identification areas and / or over-identification areas. Integrate the initial growth result and the correction growth result to obtain the target ground object image area, including:

[0152] For the missed identification area, add the initial growth result and the correction growth result of the missed identification area;

[0153] For the over-identification area, subtract the initial growth result from the correction growth result of the over-identification area;

[0154] Based on the result after addition and / or subtraction, obtain the target ground object image area.

[0155] If only one regional growth is performed, the growth result obtained by the one-time regional growth method may not be able to completely identify the target area. In the actual scenario, the initial growth result obtained when the image is first grown may have missed identification areas, may have over-identification areas, or may have both over-identification areas and missed identification areas. According to the actual situation, corresponding processing is performed for different phenomena of inaccurate identification.

[0156] Although mean shift performs preliminary processing on the image at the color level, it cannot completely solve the problems of missed recognition and over-recognition. Missed recognition means that the recognition result is missing, and over-recognition means that non-road parts are recognized as roads. To solve these two types of problems, this paper selects to perform regional regrowth on the areas that are not accurately recognized. For the problem of missed recognition, the second regional growth is to select another growth point in the missed recognition area in the image after the initial growth is completed and perform a new round of regional growth operations on it. The final result is the sum of the initial growth result and the second growth result. For the problem of over-recognition, regional regrowth selects a point in the area to be removed and performs reverse growth on it. The final result is the initial growth result minus the regrowth result. In actual operation, if the growth thresholds adopted in the first growth process and the second growth process of the over-recognition area are the same, the initial growth and regrowth are completely reversible operations and cannot solve the over-recognition problem. Therefore, the growth threshold needs to be reduced during reverse growth. If there are areas that are not accurately recognized in the initial growth result, after obtaining the initial growth result and before performing contour processing, the second regional growth needs to be performed according to the method of secondary growth. Correspondingly, if there are areas that are not accurately recognized, secondary growth needs to be performed on the basis of Figure 8 and Figure 9 to obtain a more complete image area.

[0157] Preferably, in the above step S105, filling the holes in the target ground object image area, obtaining the contour of the filled target ground object image area, and obtaining the target ground object annotation area, including:

[0158] Determining the connected regions to be filled according to the area of the connected regions of the target ground object image area and a preset filling threshold, performing hole filling based on the connected regions to be filled, extracting the edge of the filled target ground object image area, extracting the contour of the target ground object image area, and obtaining the target ground object annotation area according to the contour of the target ground object image area.

[0159] Preferably, determining the connected regions to be filled according to the area of the connected regions of the target ground object image area and a preset filling threshold includes:

[0160] Performing contour processing on the image to be annotated according to the closing operation method and constructing a structure body of the target ground object image area with a size of k×k, searching for connected regions in the processed image and calculating the area of each connected region;

[0161] Determining the preset filling threshold according to the size of the image to be annotated and the size of the structure body of the target ground object image area;

[0162] Determining the connected regions with an area smaller than the preset filling threshold in the connected regions as the connected regions to be filled.

[0163] When using the region growing method for recognition, due to the presence of noise and obstacles in the image, there may be some holes in the recognition result, that is, some unrecognized regions. In addition, the region growing method does not specifically process the contour edges, so the contour edges are often not smooth enough. In the actual scenario, the road, as the object to be detected, usually appears in a strip shape and the edges should be relatively coherent. However, when using the region growing method in this embodiment, the similarity metric criterion measures the color information and spatial information and does not process the contour information. Therefore, a closing operation method is adopted to preliminarily process the contour problem. During the closing operation, the regeneration result is converted into a binary image, and a structuring element with a size of k×k is constructed, and the recognition result is first dilated and then eroded. The closing operation can make the edges of the recognition result smoother. Figure 10 is the result graph of the first region growth before the closing operation on the Figure 2 original image and is shown in binary, Figure 11 is the result graph of the first region growth after the closing operation on the Figure 2 original image and is shown in binary. The white part is the grown road region. Comparing Figure 10 and Figure 11 , it can be seen that when the closing operation is not performed, there are many noise points and holes in the figure, and the contour edges are not smooth enough. Performing the closing operation can well solve the above problems.

[0164] Preferably, a preset filling threshold is determined according to the size of the image to be labeled and the size of the structuring element of the target object image region, including:

[0165] The preset filling threshold limit is determined according to the following formula: where k is the size of the structuring element, H is the height of the input image, W is the width of the input image, and λ fill is the weight coefficient for adjusting the image size and the structuring element size.

[0166] Although the closing operation can process some holes smaller than the structuring element, for holes larger than the 2k×2k region, there will still be holes after the closing operation. If the size k of the structuring element is increased in order to fill the result using the closing operation, it will cause excessive dilation of the foreground part of the road during the dilation operation, resulting in the loss of a large amount of detailed information. In addition, it is extremely laborious and highly repetitive to manually select points for each hole in the first growth result and then perform growth. In the actual scenario, the road, as the object to be detected, mostly appears continuously, that is, there should be no small part of the road foreground in the background region, and there should be no small part of the background in the road foreground. To solve this problem, the method searches for connected regions in the image after the closing operation, calculates the area of each connected region, and reverses the foreground and background of the regions with an area smaller than the filling threshold in the connected regions, that is, converting small-area foreground pixels to background and small-area background pixels to foreground, to achieve the purpose of hole filling.

[0167] Annotating an image means classifying each pixel in the image. In the task of identifying road parts in a remote sensing image, all pixel points of the road parts are marked as the foreground class 1, and the remaining pixel points are marked as the background class 0. Therefore, the task of image annotation can be converted into the task of classifying all pixel points in the image.

[0168] Figure 12 It is the result of region growing after hole filling and has been binarized. Figure 11 In, there is a black area in the road part. This area is the background area and should not appear in the road part. After foreground-background inversion, this background area has been filled. Figure 12 The white part in is the grown road area.

[0169] Preferably, edge extraction is performed on the filled target object image region to extract the contour of the target object image region, including:

[0170] Performing Gaussian filtering on the filled target object image region;

[0171] Determining the gradient of the target object image region after Gaussian filtering according to the edge detection algorithm and constructing the edge gradient and corrected gradient direction of the image region;

[0172] After correcting the gradient direction of the image region according to the corrected gradient direction, non-maximum suppression is performed according to the gradient of the target object image region to extract the edge information of the target object image region;

[0173] Performing double-threshold selection on the edge information of the target object image region extracted according to the edge gradient to extract the contour of the target object image region.

[0174] Preferably, constructing the edge gradient and corrected gradient direction of the image includes:

[0175] Based on the gradient in the horizontal direction and the gradient in the vertical direction of the image, constructing the edge gradient Edge Gradient of the image according to the following formula: where G x represents the gradient in the horizontal direction of the image, and G y represents the gradient in the vertical direction of the image;

[0176] Based on the gradient in the horizontal direction and the gradient in the vertical direction of the image, constructing the corrected gradient direction Edge Angle of the image according to the following formula: where G x represents the gradient in the horizontal direction of the image, and G y represents the gradient in the vertical direction of the image;

[0177] Performing non-maximum suppression according to the gradient of the target object image region includes:

[0178] For each pixel point in the image region, search for the gradient values of its adjacent pixel points in the corresponding gradient direction and compare them with its own gradient value;

[0179] If the pixel point gradient value is greater than or equal to the adjacent pixel point gradient value, retain the pixel point gradient value; otherwise, change the pixel point gradient value to 0.

[0180] Performing double-threshold selection on the edge information of the extracted target object image region according to the edge gradient includes:

[0181] If the edge gradient value of a pixel point is higher than the maximum gradient Max Gradient, regard the pixel point as a true boundary point

[0182] If the edge gradient of a pixel point is lower than the preset minimum gradient Min Gradient, regard the pixel point as a non-boundary point;

[0183] If the edge gradient of a pixel point is between the preset maximum gradient Max Gradient and the preset minimum gradient Min Gradient, regard the pixel point as a false boundary point;

[0184] Search for the pixel points in the neighborhood of the false boundary point. If the false boundary point is connected to a true boundary point, regard the false boundary point as a true boundary point; if the false boundary point is not connected to a true boundary point, discard the false boundary point.

[0185] After filling the holes, the image already has a relatively complete contour, but there is still a slight difference from the actual road shape. In actual projects, the shape of the edge area of the road part is relatively coherent. Therefore, it is necessary to correct the contour of the road part. The Canny edge detection algorithm can be used to extract the edge part of the filled image, or other edge extraction algorithms can also be adopted. The Canny edge detection algorithm needs to first perform Gaussian filtering on the image to eliminate the noise in the image. The Sobel operator can be used for calculating the gradient information, or other operators can also be selected for gradient calculation.

[0186] The constructed edge gradient is used to measure the edge strength. It is calculated by combining the magnitudes of the horizontal and vertical gradients. The larger the edge gradient, the more obvious the edge feature at that location. The constructed gradient direction can obtain the direction of the edge, that is, the orientation of the edge in the image space. This information is used in the non-maximum suppression step to determine which pixels are the true pixels on the edge. Since the original gradient direction has too many values, the obtained gradient direction is approximately simplified, and the original gradient direction is approximately transformed into the vertical, horizontal, main diagonal, and secondary diagonal directions. Non-maximum suppression of the image can effectively remove the non-maximum values in the image after gradient calculation and non-maximum suppression processing, facilitating the extraction of the true edge information. For the image after non-maximum suppression, if the gradient value of a pixel point is higher than the maximum gradient Max Gradient, this point is marked as a true boundary point; if the gradient of the pixel point is lower than the minimum gradient Min Gradient, this point is regarded as a non-boundary point. The pixel points with gradient values between the maximum gradient Max Gradient and the minimum gradient Min Gradient are marked as false boundary points. The maximum gradient and the minimum gradient can be preset with a value according to the gradient information or can be adjusted dynamically.

[0187] Figure 13 The result of edge extraction for the filled image, where the white lines are the extracted road edges.

[0188] Preferably, in the above step S105, obtaining the target object annotation area according to the contour of the target object image area includes:

[0189] Using the Douglas-Peucker algorithm to perform polygon fitting on the contour of the extracted target object image area to obtain the polygon fitting result, and performing region growing on the polygon fitting result to generate the target object annotation area.

[0190] The specific operation steps of the Douglas-Peucker algorithm are as follows:

[0191] A straight line AB is drawn between the starting point A and the ending point B of the extracted edge curve, that is, a chord of the curve is taken. Next, the point C on the curve that is farthest from this straight line is found, and its distance d from AB is calculated. The distance d is compared with a preset threshold. If the maximum distance is less than the preset threshold, this straight-line segment is used as an approximation of the curve. If the maximum distance d is greater than the set threshold, the curve AB is divided into two segments AC and BC at point C, and the above steps are respectively performed on these two segments. When all curves are processed, the polygon fitting result can be obtained. The steps of performing polygon fitting using the Douglas-Peucker algorithm are steps that can be achieved by the public and will not be elaborated here too much. It should be noted that if the preset threshold is too large, the contour will be completely flattened, and the recognition of curved roads will be poor. If the preset threshold is too small, the contour fitting effect is not obvious, and the purpose of correcting the edge still cannot be achieved. Therefore, the threshold needs to be adjusted in combination with the actual scenario. The result of polygon fitting is the contour rather than the complete annotation result. Finally, the polygon fitting result is region-grown with the positive growth points given during the growth process as the starting growth points to convert the road contour fitted by the polygon into a binary marked map with the road area highlighted. The positive growth points refer to the first starting growth points and the growth points during the missed recognition process.

[0192] Figure 14 It is a result map regrown for polygon fitting of the extracted edge. At this time, it can be seen that the white part is a relatively complete target area, and the contour part has become very smooth. Figure 15 It is the final ground feature attribute annotation area. Figure 15 The light-colored area marked by the frame line in it is the annotation area, that is, the road area to be annotated in this embodiment.

[0193] Based on the same inventive concept, an embodiment of the present invention further provides an image ground feature attribute annotation device. This device can be set in a device capable of processing computer instructions and having an arithmetic function. The structure of this device is as Figure 16 shown, including:

[0194] A region growth module 11, configured to select a first starting growth point from the image region of the target ground feature in the image to be annotated; based on the selected region growth algorithm and similarity measurement criterion, perform the first region growth starting from the first starting growth point to obtain the initial growth result of the target ground feature image region; if it is determined according to the initial growth result that there is an area in the image to be annotated that is not accurately recognized, select a second starting growth point from the area that is not accurately recognized for secondary growth to obtain the corrected growth result of the target ground feature image region;

[0195] The region determination module 12 is configured to integrate the initial growth result and the corrected growth result to obtain a target ground object image region; fill the holes in the target ground object image region, and obtain the contour of the filled target ground object image region to obtain a target ground object annotation region.

[0196] Since for the image ground object attribute annotation device in the above embodiments, the specific manners in which each module performs operations have been described in detail in the embodiments related to the method, no detailed description will be given here.

[0197] The above method and device in the embodiments of the present invention use mean shift to cluster the image by distance and color, and then make full use of the diversity of the similarity measurement criteria in the region growing method and the optimization effect of morphological post-processing on the segmentation result. It can grow along the pixel points with high spatial similarity and distance similarity around the starting growth point and correct the growth result. Finally, an accurate image annotation region is obtained. Compared with the traditional manual annotation method and the annotation algorithm method, the workload is reduced, the annotation efficiency is improved, and the marked target image region is more accurate and complete. Using this method for annotation enables the construction team to have an overall understanding of the road network structure of the entire area, providing basic data for specific tasks such as construction path planning and obstacle area excitation point layout in the acquisition design and construction links.

[0198] The embodiments of the present invention further provide a computer storage medium, in which computer-executable instructions are stored. When the computer-executable instructions are executed by a processor, the above-mentioned image ground object attribute annotation method is implemented.

[0199] The embodiments of the present invention further provide a computer device, including: a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the above-mentioned image ground object attribute annotation method is implemented.

[0200] Unless otherwise specifically stated, terms such as processing, calculating, computing, determining, displaying, etc. may refer to the actions and / or processes of one or more processing or computing systems, or similar devices, which operate on and transform data represented as physical (such as electronic) quantities in the registers or memories of the processing system into other data represented as physical quantities in the memories, registers, or other such information storage, transmission, or display devices of the processing system. Information and signals can be represented using any of a variety of different technologies and methods. For example, the data, instructions, commands, information, signals, bits, symbols, and chips mentioned throughout the above description can be represented by voltages, currents, electromagnetic waves, magnetic fields or particles, optical fields or particles, or any combination thereof.

[0201] It should be understood that the specific order or hierarchy of steps in the disclosed process is an example of an exemplary method. Based on design preferences, it should be understood that the specific order or hierarchy of steps in the process can be rearranged without departing from the scope of the present disclosure. The appended method claims present the elements of the various steps in an exemplary order and are not intended to be limited to the specific order or hierarchy recited.

[0202] In the above detailed description, various features are combined in a single embodiment to simplify the present disclosure. This method of disclosure should not be interpreted as reflecting an intention that the embodiments of the claimed subject matter require more features than are expressly recited in each claim. Rather, as reflected in the appended claims, the invention lies in less than the full scope of the features of the single disclosed embodiment. Accordingly, the appended claims are hereby expressly incorporated into the detailed description, with each claim standing on its own as a separate preferred embodiment of the invention.

[0203] Those skilled in the art should also understand that the various illustrative logical blocks, modules, circuits, and algorithmic steps described in connection with the embodiments herein can be implemented as electronic hardware, computer software, or combinations thereof. To clearly illustrate the interchangeability of hardware and software, the various illustrative components, blocks, modules, circuits, and steps have been generally described in terms of their functionality. Whether such functionality is implemented as hardware or software depends upon the particular application and the design constraints imposed on the overall system. Skilled artisans may implement the described functionality in a variable manner for each particular application, but such implementation decisions should not be construed as departing from the scope of the present disclosure.

[0204] The steps of a method or algorithm described in connection with the embodiments herein can be embodied directly in hardware, in a software module executed by a processor, or in a combination thereof. The software module can be located in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, hard disk, a removable disk, a CD-ROM, or any other form of storage medium well known in the art. An exemplary storage medium is coupled to the processor such that the processor can read information from, and write information to, the storage medium. Of course, the storage medium can also be integral to the processor. The processor and the storage medium can be located in an ASIC. The ASIC can be located in a user terminal. Of course, the processor and the storage medium can also exist as discrete components in a user terminal.

[0205] For software implementation, the techniques described in this application can be implemented by modules (e.g., procedures, functions, etc.) that perform the functions described in this application. These software codes can be stored in a memory unit and executed by a processor. The memory unit can be implemented within the processor or outside the processor. In the latter case, it is communicatively coupled to the processor via various means, which are well known in the art.

[0206] The above description includes examples of one or more embodiments. Of course, it is not possible to describe all possible combinations of components or methods for the purpose of describing the above embodiments, but those of ordinary skill in the art should recognize that the various embodiments can be further combined and arranged. Accordingly, the embodiments described herein are intended to cover all such changes, modifications, and variations that fall within the scope of the appended claims. In addition, with respect to the term "comprising" used in the specification or claims, the word is intended to be construed in a manner similar to the term "including" as interpreted when used as a transitional word in a claim. Further, any use of the term "or" in the specification or claims is intended to mean "non-exclusive or".

Claims

1. An image feature attribute annotation method, characterized in that, Including: Selecting a first starting growth point from the image region of the target feature in the image to be annotated; Based on the selective region growing algorithm and the similarity metric criterion, performing the first region growing starting from the first starting growth point to obtain the initial growth result of the target feature image region; If it is determined according to the initial growth result that there are regions in the image to be annotated that are not accurately recognized, then selecting a second starting growth point from the regions that are not accurately recognized for secondary growth to obtain the corrected growth result of the target feature image region; Integrating the initial growth result and the corrected growth result to obtain the target feature image region; Filling the holes in the target feature image region, obtaining the contour of the filled target feature image region, and obtaining the target feature annotation region.

2. The method according to claim 1, wherein Based on the selective region growing algorithm and the similarity metric criterion, performing the first region growing starting from the first starting growth point to obtain the initial growth result of the target feature image region, including: Traversing the points to be grown in the growth neighborhood of the first starting growth point; Determining the similarity between the point to be grown and the first starting growth point according to the similarity metric criterion, adding the points to be grown that meet the growth criterion to the first growth domain set, using the newly added point to be grown as the new first starting growth point, and returning to continue executing the step of traversing the points to be grown in the growth neighborhood of the first starting growth point until there are no points to be grown that meet the growth criterion, obtaining the initial growth result of the target feature image region.

3. The method according to claim 2, wherein The step of traversing the points to be grown in the growth neighborhood of the first starting growth point; determining the similarity between the point to be grown and the first starting growth point according to the similarity metric criterion, and adding the points to be grown that meet the growth criterion to the first growth domain set includes: Taking the first starting growth point as the center to determine the neighborhood of the first starting growth point; Determining the similarity Dissimilarity between the points to be grown in the neighborhood and the first starting growth point according to the following similarity metric rule: Dissimilarity(p1,p2)=λ color dis color (p1,p2)+λ local dis local (p1,p2), Among them, λ color is the weight for adjusting the spatial balance degree, and λ local is the weight for adjusting the color balance degree. dis color represents the modulus of the difference between the vectors in the form of 1×1×3 corresponding to points p1 and p2, and dis local represents the distance between the coordinates of the two points in the image; Traversing the points to be grown in the growth neighborhood of the first starting growth point, if the similarity between the point to be grown and the first starting growth point is less than the pre-determined growth threshold of the first starting growth point, then adding the point to be grown to the first growth domain set S.

4. The method according to claim 2, wherein Determining the growth threshold of the first starting growth point, including: Based on the digital features of the three-dimensional matrix corresponding to the image to be labeled, determine the growth threshold MaxSim of the first starting growth point according to the following formula: where Std image represents the standard deviation of the three-dimensional matrix corresponding to the image after mean shift.

5. The method according to claim 1, characterized in that, If it is determined according to the initial growth result that there are regions in the image to be annotated that are not accurately recognized, then selecting a second starting growth point from the regions that are not accurately recognized for secondary growth to obtain the corrected growth result of the target feature image region, including: Judging whether there are regions in the image to be annotated that are not accurately recognized according to the initial growth result; the regions that are not accurately recognized include regions of missed recognition and / or over-recognition; If so, selecting a second starting growth point from each region that is not accurately recognized, and based on the selective region growing algorithm and the similarity metric criterion, performing the second region growing starting from the second starting growth point to obtain the corrected growth result of the target feature image region.

6. The method according to claim 5, characterized in that, If the unaccurately recognized region is a missed recognition region, select a second starting growth point from each unaccurately recognized region, and based on the selective region growing algorithm and the similarity metric criterion, perform a second region growing starting from the second starting growth point to obtain the corrected growth result of the target ground object image region, including: Select a second starting growth point from the missed recognition region, traverse the points to be grown in the growth neighborhood of the second starting growth point. If the dissimilarity between the point to be grown in the growth neighborhood of the second starting growth point and the second starting growth point is less than the growth threshold MaxSim of the second starting growth point in the missed recognition region, add the point to be grown to the second growth domain set; Take the newly added point to be grown as the new second starting growth point, and continue to execute the step of traversing the points to be grown in the growth neighborhood of the second starting growth point until there are no new points to be grown in the second growth domain set, and the corrected growth process ends, obtaining the corrected growth result of the target ground object image region.

7. The method according to claim 5, wherein If the unaccurately recognized region is an over-recognition region, select a second starting growth point from each unaccurately recognized region, and based on the selective region growing algorithm and the similarity metric criterion, perform a second region growing starting from the second starting growth point to obtain the corrected growth result of the target ground object image region, including: Select a second starting growth point in the over-recognition region, traverse the points to be grown in the growth neighborhood of the second starting growth point. If the dissimilarity between the point to be grown in the growth neighborhood of the second starting growth point and the second starting growth point is less than the growth threshold newMaxSim of the second starting growth point in the over-recognition region, add the point to be grown to the third growth domain set; Take the newly added point to be grown as the new second starting growth point, and continue to execute the step of traversing the points to be grown in the growth neighborhood of the second starting growth point until there are no new points to be grown in the third growth domain set, and the corrected growth process ends, obtaining the corrected growth result of the target ground object image region; The growth threshold newMaxSim for the second starting growth point of the recognition area is determined according to the following formula: newMaxSim = MaxSim * α times , where α is a correction factor.

8. The method according to claim 1, wherein The unaccurately recognized region includes a missed recognition region and / or an over-recognition region. Integrate the initial growth result and the corrected growth result to obtain the target ground object image region, including: For the missed recognition region, add the initial growth result and the corrected growth result of the missed recognition region; For the over-recognition region, subtract the initial growth result from the corrected growth result of the over-recognition region; Based on the result after addition and / or subtraction, obtain the target ground object image region.

9. The method according to claim 1, characterized in that, Fill the holes in the target ground object image region, obtain the contour of the filled target ground object image region, and obtain the target ground object annotation region, including: According to the connected region area of the target ground object image region and a preset filling threshold, determine the connected region to be filled, perform hole filling based on the connected region to be filled, perform edge extraction on the filled target ground object image region, extract the contour of the target ground object image region, and obtain the target ground object annotation region according to the contour of the target ground object image region.

10. The method according to claim 9, characterized in that, Determine the connected regions to be filled according to the area of the connected regions in the target object image region and a preset filling threshold, including: According to the closing operation method, perform contour processing on the to-be-annotated image and construct a structure body of the target object image region with a size of k×k, search for connected regions in the processed image and calculate the area of each connected region; Determine the preset filling threshold according to the size of the to-be-annotated image and the size of the structure body of the target object image region; Determine the connected regions with an area smaller than the preset filling threshold in the connected regions as the connected regions to be filled.

11. The method according to claim 10, characterized in that, Determine the preset filling threshold according to the size of the to-be-annotated image and the size of the structure body of the target object image region, including: Determine the preset filling threshold limit according to the following formula: where k is the structure size, H is the height of the input image, W is the width of the input image, and λ fill is the weight coefficient for adjusting the image size and the structure size.

12. The method according to claim 9, wherein Perform edge extraction on the filled target object image region to extract the contour of the target object image region, including: Perform Gaussian filtering on the filled target object image region; Determine the gradient of the target object image region after Gaussian filtering according to the edge detection algorithm and construct the edge gradient and corrected gradient direction of the image region; After correcting the gradient direction of the image region according to the corrected gradient direction, perform non-maximum suppression according to the gradient of the target object image region to extract the edge information of the target object image region; Perform double-threshold selection on the edge information of the extracted target object image region according to the edge gradient to extract the contour of the target object image region.

13. The method according to claim 12, characterized in that, Construct the edge gradient and corrected gradient direction of the image, including: Based on the gradients in the horizontal and vertical directions of the image, construct the edge gradient EdgeGradient of the image according to the following formula: where G x represents the gradient in the horizontal direction of the image, and G y represents the gradient in the vertical direction of the image; Based on the gradients in the horizontal and vertical directions of the image, construct the corrected gradient direction Edge Angle of the image according to the following formula: where G x represents the gradient in the horizontal direction of the image, and G y represents the gradient in the vertical direction of the image; The non-maximum suppression according to the gradient of the target object image region includes: For each pixel point in the image region, search for the gradient values of its adjacent pixel points in the corresponding gradient direction and compare them with its own gradient value; If the pixel point gradient value is greater than or equal to the adjacent pixel point gradient value, retain the pixel point gradient value, otherwise change the pixel point gradient value to 0; The double-threshold selection of the edge information of the extracted target object image region according to the edge gradient includes: If the edge gradient value of a pixel point is higher than the maximum gradient Max Gradient, regard the pixel point as a true boundary point If the edge gradient of a pixel point is lower than the preset minimum gradient Min Gradient, regard the pixel point as a non-boundary point; If the edge gradient of a pixel point is between the preset maximum gradient Max Gradient and the preset minimum gradient Min Gradient, regard the pixel point as a false boundary point; Search for the pixel points in the neighborhood of the false boundary point. If the false boundary point is connected to a true boundary point, regard the false boundary point as a true boundary point; if the false boundary point is not connected to a true boundary point, discard the false boundary point.

14. The method according to claim 1, characterized in that Obtain the target object annotation region according to the contour of the target object image region, including: Use the Douglas-Peucker algorithm to perform polygon fitting on the contour of the extracted target object image region to obtain the polygon fitting result, and perform region growing on the polygon fitting result to generate the target object annotation region.

15. An image ground object attribute annotation device, characterized in that, Including: A region growth module, configured to select a first starting growth point from within the image region of the target ground object in the image to be annotated; based on the selective region growth algorithm and the similarity metric criterion, perform the first region growth starting from the first starting growth point to obtain the initial growth result of the target ground object image region; if it is determined from the initial growth result that there are regions in the image to be annotated that are not accurately recognized, select a second starting growth point from the regions that are not accurately recognized for secondary growth to obtain the corrected growth result of the target ground object image region; A region determination module, configured to integrate the initial growth result and the corrected growth result to obtain the target ground object image region; Fill the holes in the target ground object image region, obtain the contour of the filled target ground object image region, and obtain the target ground object annotation region.

16. A computer storage medium, characterized in that, The computer storage medium stores computer-executable instructions, and when the computer-executable instructions are executed by a processor, the image ground object attribute annotation method according to any one of claims 1-14 is implemented.

17. A computer device, characterized in that, Comprising: A memory, a processor, and a computer program stored on the memory and executable on the processor, and when the processor executes the program, the image ground object attribute annotation method according to any one of claims 1-14 is implemented.

Images