CN110443822A - A kind of high score remote sensing target fine extracting method of semanteme edge auxiliary - Google Patents

Abstract

A kind of fine extracting method of high score remote sensing target based on edge auxiliary is extracted task design depth convolutional neural networks according to remote sensing target first, is maked sample and trains to obtain target detection model and Model for Edge Detection.Then remote sensing image is input in target detection model and carries out the extraction of atural object outsourcing frame, so that it is determined that ground object target type and position range；Remote sensing image is input to progress atural object edge extracting in Model for Edge Detection, to obtain ground object target edge strength figure.Finally for the target in each outsourcing frame, edge strength figure with corresponding position is that instruction carries out object boundary extraction operation, refinement is directly carried out to the higher limbus of intensity as boundary, it with boundary is completely that target is repaired to fuzzy especially interrupted edges, it is final to determine target fine boundary, object boundary vector can be turned to polygon element by mission requirements.The fine extraction of high score remote sensing target can be achieved in the present invention.

Description

A Semantic Edge Aided Method for Fine Extraction of High-score Remote Sensing Targets

technical field

The invention belongs to the field of remote sensing and target detection and edge extraction in computer vision, and proposes a two-stage remote sensing target fine extraction method based on deep learning combined with target detection and edge detection.

technical background

Target extraction is an important method to extract information from remote sensing data. With the continuous improvement of the spatial resolution of remote sensing images, the requirements for observable targets and their extraction accuracy are also increasing accordingly. Although instance segmentation in general computer vision can be used for some parts However, the requirements for the actual boundary of objects are far from each other. For example, the two-stage method of target detection and mask segmentation (such as Mask-RCNN, etc.) commonly used at present is difficult to be directly applied to the fine target extraction of high-resolution remote sensing images. In fact, the target detection based on the deep learning method only needs to locate the target and determine the outer frame, so the accuracy is generally high; the actual effect of the target extraction effect mainly lies in the fine determination of the actual boundary of the ground object in the mask segmentation stage. Due to the complex changes of the target in terms of image spectral performance, spatial shape scale, etc., it is still difficult to achieve good accuracy by detecting local data in the frame for boundary recognition due to the current depth segmentation model. This is also the main improvement idea of this method. By introducing semantics The edge improves the extraction accuracy of the target boundary, so as to realize the fine extraction of the high-scoring remote sensing target as a whole, and achieve the practical ability.

Since the introduction of deep learning, the current mainstream target detection algorithms can be roughly divided into two categories, namely, the target detection convolutional neural network based on region proposal and the end-to-end integrated target detection convolutional neural network. In order to pursue faster detection speed, the end-to-end target detection method adopts the recognition result of a single detection as the final target detection result, and its effect is not ideal compared to the multi-stage training-based target detection method based on region proposal. In the field of target detection based on region proposal, R-CNN is a pioneering proposal. It changes the previous tedious and exhaustive sliding window form, extracts a small number of suitable region candidate boxes through selective search, and then normalizes the size of the candidate region. and use deep convolutional neural network to extract features, and finally use SVM for classification and recognition to obtain the final target detection result. The detection algorithms such as Fast-RCNN, Faster-RCNN, FPN, and Mask-RCNN proposed later are all targeted improvements on this basis. Since the target detection algorithm only needs to locate and classify the target, it has a good detection accuracy.

The edge detection algorithm requires accurate determination of the target boundary, which is more technically difficult than simply locating the target. Among edge detection algorithms based on deep learning, RCF is a more effective convolutional neural network model. The main idea is to use the VGG model for fine-tuning, reduce the dimensionality of multiple convolutional layers in VGG to obtain a feature fusion map, then deconvolve and calculate the loss value of each stage separately, and finally synthesize 5 at multiple scales. The feature fusion map of each stage gets the final feature fusion map and calculates the loss value. Using the features of all convolutional layers can improve the effect compared to using only the features of the last convolutional layer of each stage in HED. In the case of strong boundaries, because the edges contain relatively rich semantic information, and then global information is obtained through convolution, it has a good edge detection effect. However, in the case of weak boundaries, it is difficult for the target edge to provide effective semantic information, and even due to the shadow occlusion of tall objects, the distinction between target and non-target edges is completely lost, so that the neural network cannot identify the target edge, resulting in edge extraction. line condition. In actual requirements, the edge of the broken line cannot be regarded as a complete boundary extraction, so the boundary repair work aiming at boundary integrity can be regarded as a very important content in the target boundary extraction task.

This method attempts to combine the advantages of target detection and edge extraction. Under the guidance of the high-precision target detection outer frame, the edge detection results are used to finely locate the target boundary. For occluded or disturbed edges, it is based on the target detection results. Actively infer and connect Completion, so as to achieve fine extraction of remote sensing targets.

SUMMARY OF THE INVENTION

In order to overcome the shortcomings of inaccurate and incomplete edge extraction of objects in high-resolution remote sensing images due to blurred object boundaries and mutual occlusion, the invention proposes a method for finely extracting high-resolution remote sensing objects based on edge assistance.

The invention can overcome the defect of inaccurate mask boundary in general instance segmentation through edge assistance, and at the same time, through a two-stage process with the assistance of the outer frame, the target boundary can be guaranteed to be complete, overcome the defect of easy-to-break lines in general edge extraction, and finally achieve high Fine extraction of remote sensing targets.

For achieving the above object, the technical scheme proposed by the present invention is as follows:

A method for finely extracting high-scoring remote sensing targets based on edge assistance, the method comprises the following steps:

Step 1: Make a remote sensing image target extraction sample, draw the fine boundary of the target against the image and determine the type; generate two different labels for the same image sample, that is, generate the target outer box label according to the requirements of the target detection model, and according to the requirements of the edge detection model. Generate target boundary annotations; specifically:

Step 1.1: Obtain high-resolution remote sensing images: use optical satellite remote sensing data with visible light-near-infrared sensors or aerial remote sensing data equipped with general optical cameras. According to the resolution requirements, multi-spectral images or fusion panchromatic images can be used directly;

Step 1.2: Crop the remote sensing image: Select the typical target range in the production area and crop the remote sensing image according to uniform pixels;

Step 1.3: Make deep learning training samples: Use ArcGIS or other GIS software to draw deep learning training samples, mark the boundaries of each feature target, get the corresponding .shp file, and generate target outer frame labels according to the requirements of the target detection model; The detection model requires the generation of target boundary annotations;

Step 1.4: Collect more than 200 images and corresponding annotations as training samples according to task requirements, and prepare test samples separately for detection accuracy;

Step 2: Use the prepared image samples to train the target detection model and the edge detection model; train the FasterR-CNN neural network to obtain the target detection model, train the RCF neural network to obtain the edge detection model, and replace and modify the network accordingly according to the production target; specifically including :

Step 2.1: Design a deep convolutional neural network: In order to train the target detection model and the edge detection model, two neural networks, RCF and Faster R-CNN, are selected, and the networks can be replaced and modified according to the production target;

Step 2.2: Initialize weights: Use the VGG pre-training model to initialize the RCF network weights, and use the pre-training model on the COCO dataset to initialize the Faster R-CNN weights;

Step 2.3: Set training hyperparameters: configure hyperparameters, and set specific values after model tuning;

RCF training parameter settings: number of iterations = 8000, batch_size = 4, learning rate update strategy = step, learning rate update step = [3200, 4800, 6400, 8000], initial learning rate = 0.001, learning rate update coefficient = 0.1;

Faster R-CNN training parameter settings: number of training stages = 3, number of iterations in each stage = [40, 120, 40], number of iterations per round = 1000, verification interval = 50, batch_size = 4, learning rate update strategy = step , learning rate update step=1000, initial learning rate=0.001, learning rate update momentum=0.9, weight decay=0.0001;

Step 2.4: Input samples and train the model: Input the training samples into the RCF model, and train according to the hyperparameters described in Step 2.3 to obtain an edge detection model that can extract the edge contours of the objects; input the training samples into Faster R -In CNN, train according to the hyperparameters described in step 2.3 to obtain a target detection model that can extract the outer frame of the object target;

Step 3: Input the image samples for testing into the target detection model and the edge detection model, and obtain the outer frame of the object target and the edge intensity map; the edge intensity map of the object object can indicate the possibility that the corresponding position in the remote sensing image is the edge of the target The outer frame of the object target is a rectangular bounding box that indicates the location range and target type of the target; it includes:

Step 3.1: Input the high-resolution remote sensing image into the target detection model to obtain the rectangular outer box of the object target; input the high-resolution remote sensing image into the edge detection model to obtain the edge intensity map of the object target;

Step 3.2: Parametric outer box: convert the parameters of the rectangular outer box output by the target detection model into the coordinates of the lower left vertex and the upper right vertex; specifically:

x1=x-w y1=y-h

x2=x+w y2=y+h

Among them, x, y, w, h represent the center abscissa, center ordinate, width and height of the rectangular outer frame, respectively;

Step 4: Refine the edge intensity map in Step 3 into a single-pixel-wide boundary; the details include:

Step 4.1: According to the set threshold, the edge intensity map of the object object is converted from a grayscale image to a binary image. Use binary_image(x, y) to represent the edge judgment of the object at the image (x, y) (1 means it is an edge, 0 means it is not an edge), which can be expressed as:

The threshold is a real number in the interval [0,1], which can be set by the user, and the initial default value is 0.5. x, y are the horizontal and vertical coordinates of the image.

Step 4.2: For the thick edge lines with multiple pixel widths in the binary image, the erosion operation is performed continuously with the center of each edge line until all the edge lines are only one pixel wide, so as to achieve the purpose of skeleton extraction.

Step 5: Using the outer frame of the object target in step 3 as a constraint, repair the single-pixel-wide boundary in step 4 to obtain a complete and fine polygon object boundary.

Specifically include:

Step 5.1: Find out the unclosed part of the boundary line of the object target, and divide it into three types, namely the boundary breakline at the edge of the image (the boundary at the edge of the image is incomplete due to the skeleton extraction algorithm), the inner part of the image. The boundary line breaks (the part not correctly identified by the boundary detection model), the boundary line head inside the image (the redundant boundary line on the target boundary due to the skeleton extraction algorithm).

Step 5.2: Three types of boundary breaks are handled in three different ways.

Step 5.2.1: For the boundary broken line at the edge of the image, the edge intensity map at the corresponding position is used as an indicator, and the pixels with higher edge intensity values are selected to fill the target boundary broken line. If there is still a gap with the image boundary, use A straight line perpendicular to the image boundary, which is disconnected from the target boundary and connected to the image boundary to form a closed target boundary.

Step 5.2.2: For the boundary broken line inside the image, reset the threshold, and use the edge intensity map at the corresponding position as an indication, map the pixels whose intensity value is higher than the new threshold in the edge intensity map to the binary map, If the target boundary break line is still not closed, connect the two break points with the original geometric properties. Among them, the steps of repairing the breakpoint with the original geometric characteristics are as follows.

Step 5.2.2.1: Take out the innermost circle boundary line connected with the two breakpoints.

Step 5.2.2.2: Divide the boundary line of the innermost circle into several parts according to the change of the slope, and determine the rough geometric shape of the boundary line.

Step 5.2.2.3: Repair the broken point into a closed figure according to the geometric shape, so that the repaired place maintains the original geometric characteristics.

Step 5.2.3: Delete the isolated boundary lines inside the image, that is, the isolated boundary lines that are far away from other broken lines.

Step 5.3: Perform skeleton extraction again, and refine the part filled in step 5.2 into a single-pixel wide target boundary line.

Step 5.4: Re-traverse the image, delete all unclosed target boundary lines, and obtain a complete and fine polygonal object boundary.

Owing to adopting the above-mentioned technical scheme, the present invention has the following advantages and beneficial effects:

1. The present invention adopts the method of combining the target detection model and the edge detection model, which not only avoids the defects of inaccurate mask boundaries in the general instance segmentation algorithm, but also solves the problems of easily broken lines and thick outlines in the general edge extraction algorithm, It can ensure the integrity and refinement of the target boundary.

2. Compared with the traditional manual recognition of remote sensing targets, the present invention has a faster recognition speed. For a remote sensing image with hundreds of targets, the present invention can recognize less than one second, while manual drawing often takes several hours; The invention has a high accuracy rate, and because the end-to-end design method is adopted, the situation that the accuracy is reduced due to the misoperation of manual drawing is avoided.

3. The present invention adopts the deep neural network algorithm in machine learning, and only needs to change the sample training to identify various targets and extract the target boundary, without the need to redesign the algorithm, and has strong reusability and robustness .

Description of drawings

1 is a schematic diagram of an edge-assisted high-resolution remote sensing target fine extraction method of the present invention;

2 is a sample diagram of a target detection model in an embodiment of the present invention;

3 is a sample diagram of an edge detection model in an embodiment of the present invention;

4 is an example diagram of a prediction result of a target detection model in an embodiment of the present invention;

5 is an example diagram of an edge detection model prediction result in an embodiment of the present invention;

FIG. 6 is a diagram of an example comparison of final boundary results in an embodiment of the present invention.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention, that is, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. The components of the embodiments of the invention generally described in the drawings herein may be arranged and designed in a variety of different configurations.

Thus, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the invention as claimed, but is merely representative of selected embodiments of the invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the present invention.

Figure 1 is a schematic diagram of the edge-assisted high-resolution remote sensing target fine extraction method.

Referring to Figure 1, a preferred embodiment is provided for the present invention, comprising the following steps:

Step 1: Make a remote sensing image target extraction sample, draw the fine boundary of the target against the image and determine the type;

Step 2: Use the prepared image samples to train the target detection model and the edge detection model;

Step 3: Input the image samples used for testing into the target detection model and the edge detection model, and obtain the outer frame of the object target and the edge intensity map;

Step 4: Refine the edge intensity map in Step 3 into a single-pixel-wide boundary;

Step 5: Using the outer frame of the object target in step 3 as a constraint, repair the single-pixel-wide boundary in step 4 to obtain a complete and fine polygon object boundary.

According to the above embodiment, the detailed steps of step 1 are as follows:

Step 1.1: Obtain high-resolution remote sensing images: use optical satellite remote sensing data with visible light-near-infrared sensors or aerial remote sensing data equipped with general optical cameras. According to the resolution requirements, multispectral images or fusion panchromatic images can be used directly.

Step 1.2: Crop the remote sensing image: Select the range of typical targets in the production area and uniformly crop the remote sensing image to a size of 512*512 pixels (the cropping range should be random and widely covered according to the size of the production area).

Step 1.3: Make deep learning training samples: Use ArcGIS or other GIS software to draw deep learning training samples. In this embodiment, it is necessary to mark the boundaries of each feature target, obtain the corresponding .shp file, and generate targets according to the requirements of the target detection model The outer box annotation is shown in Figure 2; the target boundary annotation is generated according to the requirements of the edge detection model, as shown in Figure 3.

Step 1.4: Generally, more than 200 images and corresponding annotations are collected as training samples according to the task requirements. If the detection accuracy is required, test samples can be prepared separately.

The detailed steps of step 2 are as follows:

According to the above-mentioned embodiment, step 2.1: Design a deep convolutional neural network: in order to train the target detection model and the edge detection model, two kinds of neural networks, RCF and Faster R-CNN, are selected in the present invention. According to the production target network, the corresponding replace, modify.

Step 2.2: Initialize weights: Use the VGG pretrained model to initialize the RCF network weights, and use the pretrained model on the COCO dataset to initialize the Faster R-CNN weights.

Step 2.3: Set training hyperparameters: Configure hyperparameters. The specific numerical settings after model tuning are as follows.

RCF training parameter settings: number of iterations = 8000, batch_size = 4, learning rate update strategy = step, learning rate update step = [3200, 4800, 6400, 8000], initial learning rate = 0.001, learning rate update coefficient = 0.1.

Faster R-CNN training parameter settings: number of training stages = 3, number of iterations in each stage = [40, 120, 40], number of iterations per round = 1000, verification interval = 50, batch_size = 4, learning rate update strategy = step , learning rate update step=1000, initial learning rate=0.001, learning rate update momentum=0.9, weight decay=0.0001.

Step 2.4: Input samples and train the model: Input the training samples into the RCF model, and train according to the hyperparameters described in Step 2.3 to obtain an edge detection model that can extract the edge contours of the objects; input the training samples into Faster R -In the CNN, train according to the hyperparameters described in step 2.3 to obtain a target detection model that can extract the outer frame of the object target.

According to the above embodiment, the detailed steps of step 3 are as follows:

Step 3.1: Input the high-resolution remote sensing image into the target detection model to obtain the rectangular outer box of the object target, as shown in Figure 4; input the high-resolution remote sensing image into the edge detection model to obtain the edge intensity map of the object target, as shown in the figure 5 shown.

Step 3.2: Parametric outer box: Convert the parameters of the rectangular outer box output by the target detection model into the coordinates of the lower left vertex and the upper right vertex. specifically:

x1=x-w y1=y-h

x2=x+w y2=y+h

Among them, x, y, w, h represent the center abscissa, center ordinate, width and height of the rectangular outer box, respectively.

According to the above embodiment, the detailed steps of step 4 are as follows:

Step 4.1: According to the set threshold, the edge intensity map of the object object is converted from a grayscale image to a binary image. Use binary_image(x, y) to represent the edge judgment of the object at the image (x, y) (1 means it is an edge, 0 means it is not an edge), which can be expressed as:

The threshold is a real number in the interval [0,1], which can be set by the user, and the initial default value is 0.5. x, y are the horizontal and vertical coordinates of the image.

Step 4.2: For the thick edge lines with multiple pixel widths in the binary image, the erosion operation is performed continuously with the center of each edge line until all the edge lines are only one pixel wide, so as to achieve the purpose of skeleton extraction.

According to the above embodiment, the detailed steps of step 5 are as follows:

Step 5.1: Find out the unclosed part of the boundary line of the object target, and divide it into three types, namely the boundary breakline at the edge of the image (the boundary at the edge of the image is incomplete due to the skeleton extraction algorithm), the inner part of the image. The boundary line breaks (the part not correctly identified by the boundary detection model), the boundary line head inside the image (the redundant boundary line on the target boundary due to the skeleton extraction algorithm).

Step 5.2: Three types of boundary breaks are handled in three different ways.

Step 5.2.1: For the boundary broken line at the edge of the image, the edge intensity map at the corresponding position is used as an indicator, and the pixels with higher edge intensity values are selected to fill the target boundary broken line. If there is still a gap with the image boundary, use A straight line perpendicular to the image boundary, which is disconnected from the target boundary and connected to the image boundary to form a closed target boundary.

Step 5.2.2: For the boundary broken line inside the image, reset the threshold, and use the edge intensity map at the corresponding position as an indication, map the pixels whose intensity value is higher than the new threshold in the edge intensity map to the binary map, If the target boundary break line is still not closed, connect the two break points with the original geometric properties. Among them, the steps of repairing the breakpoint with the original geometric characteristics are as follows.

Step 5.2.2.1: Take out the innermost circle boundary line connected with the two breakpoints.

Step 5.2.2.2: Divide the boundary line of the innermost circle into several parts according to the change of the slope, and determine the rough geometric shape of the boundary line.

Step 5.2.2.3: Repair the broken point into a closed figure according to the geometric shape, so that the repaired place maintains the original geometric characteristics.

Step 5.2.3: Delete the isolated boundary lines inside the image, that is, the isolated boundary lines that are far away from other broken lines.

Step 5.3: Perform skeleton extraction again, and refine the part filled in step 5.2 into a single-pixel wide target boundary line.

Step 5.4: Re-traverse the image, delete all unclosed target boundary lines, and obtain the complete and fine object boundary as shown in Figure 6-A, of which Figure 6-B is the predicted boundary result of Mask-RCNN. By comparison, it can be found that the present invention has a relatively obvious precision advantage.

The content described in the embodiments of the present specification is only an enumeration of the realization forms of the inventive concept, and the protection scope of the present invention should not be regarded as limited to the specific forms stated in the embodiments, and the protection scope of the present invention also extends to those skilled in the art. Equivalent technical means that can be conceived by a person based on the inventive concept.

Claims (1)

1. a kind of fine extracting method of high score remote sensing target based on edge auxiliary, includes the following steps:

Step 1: production remote sensing image Objective extraction sample, control image draw the fine boundary of target and determine type；To same One width image sample generates two different marks, i.e., generates target outsourcing collimation mark note according to target detection model needs, according to Model for Edge Detection requires to generate object boundary mark；It specifically includes:

Step 1.1: obtain high score remote sensing image: using have visible light-near infrared sensor optical satellite remotely-sensed data or Multispectral image or the panchromatic shadow of fusion can be used directly according to resolution requirement in the Airborne Data Classification for carrying general optical camera Picture；

Step 1.2: cutting remote sensing image: in production area selection typical target location by remote sensing image according to unified pixel It is cut；

Step 1.3: production deep learning training sample: drawing deep learning training sample using ArcGIS or other GIS softwares, The boundary for marking each ground object target obtains corresponding .shp file, generates target outsourcing frame according to target detection model needs Mark；It is required to generate object boundary mark according to Model for Edge Detection；

Step 1.4: acquiring 200 or more images according to mission requirements and corresponding mark is used as training sample, be detection accuracy list Only setup test sample；

Step 2: using ready image sample training target detection model and Model for Edge Detection；Training Faster R-CNN Neural network obtains target detection model, and training RCF neural network obtains Model for Edge Detection, can accordingly be replaced according to productive target It changes, modify network；It specifically includes:

Step 2.1: projected depth convolutional neural networks: for training objective detection model and Model for Edge Detection, selection is Both neural networks of RCF and Faster R-CNN can be replaced accordingly according to productive target, modify network；

Step 2.2: initialization weight: RCF network weight is initialized using VGG pre-training model, using on COCO data set Pre-training model initialize the weight of Faster R-CNN；

Step 2.3: training hyper parameter: configuration hyper parameter being set, numerical value is specifically set after model tuning；

The training parameter of RCF is set: the number of iterations=8000, batch_size=4, learning rate more new strategy=step, study Rate updates step-length=[3200,4800,6400,8000], initial learning rate=0.001, and learning rate updates coefficient=0.1；

The training parameter of Faster R-CNN is set: training stage number=3, each stage iteration wheel number=[40,120,40], every wheel The number of iterations=1000, validation interval=50, batch_size=4, learning rate more new strategy=step, learning rate update step-length =1000, initial learning rate=0.001, learning rate updates momentum=0.9, weight decay=0.0001；

Step 2.4: input sample, training pattern: training sample is input in RCF model, according to super described in step 2.3 Parameter is trained, and obtains the Model for Edge Detection that can extract ground object target edge contour；Training sample is input to It in Faster R-CNN, is trained according to the hyper parameter described in step 2.3, obtains that ground object target outsourcing frame can be extracted Target detection model；

Step 3: the image sample for being used to test being input in target detection model and Model for Edge Detection, ground object target is obtained Outsourcing frame and edge strength figure；Corresponding position is the possibility of object edge in the bright remote sensing image of ground object target edge strength chart Property；Ground object target outsourcing frame is the rectangle framing mask for indicating target present position range and target type；It specifically includes:

Step 3.1: high score remote sensing image being input to target detection model, obtains the rectangle outsourcing frame of ground object target；By high score Remote sensing image is input to Model for Edge Detection, obtains ground object target edge strength figure；

Step 3.2: parametrization outsourcing frame: converting its bottom left vertex for the rectangle outsourcing frame parameter that target detection model exports and sit Mark and right vertices coordinate；Specifically:

X1=x-w y1=y-h

X2=x+w y2=y+h

Wherein, x, y, w, h respectively represent the center abscissa of rectangle outsourcing frame, and center ordinate is wide and high；

Step 4: edge strength figure in step 3 is refined into the boundary of single pixel wide；It specifically includes:

Step 4.1: according to the threshold value of setting, ground object target edge strength figure being converted to binary map by grayscale image；Use binary_ Image (x, y) indicates that (1 indicates to be edge edge estimate of situation of the ground object target at image (x, y), and 0 indicates not to be side Edge), it may be expressed as:

Wherein threshold is the real number on section [0,1], can be by user setting, initial default value 0.5；X, y are image Transverse and longitudinal coordinate；

Step 4.2: the edge thick line wide for pixels multiple in binary map is constantly corrosion behaviour with the center of each of the edges line Make, until all edge lines only remain a pixel it is wide, achieve the purpose that skeletal extraction；

Step 5: being constraint with ground object target outsourcing frame in step 3, single pixel edge in step 4 is repaired, to have obtained Whole fine polygon ground object target boundary；Detailed step is as follows:

Step 5.1: finding out part not closed in ground object target boundary line, be classified as three types, respectively image border The boundary at place is broken, the boundary broken string inside image, boundary the end of a thread inside image, and wherein the boundary broken string at image border is Cause boundary at image border imperfect due to skeletal extraction algorithm, boundary inside image broken string be border detection model not The part correctly identified, boundary the end of a thread inside image is to cause object boundary excess edge occur due to skeletal extraction algorithm Line；

Step 5.2: three kinds of different methods of boundary broken string of three types are handled；

Step 5.2.1: breaking for the boundary at image border, and the edge strength figure with corresponding position is instruction, chooses side The higher pixel of edge intensity value fills up object boundary broken string, if still having gap with image boundary, hangs down using with image boundary Straight straight line makes it be connected with object boundary broken string, and connect the object boundary line to form closure with image boundary；

Step 5.2.2: break for the boundary inside image, reset threshold value, and with the edge strength figure of corresponding position For instruction, the pixel that intensity value in edge strength figure is higher than new threshold value is mapped in binary map, if object boundary broken string is still It is not closed, then two breakpoints are connected with original geometrical property；Wherein, the step of repairing breakpoint with original geometrical property is such as Under；

Step 5.2.2.1: the innermost circle boundary line being connected with two breakpoints is taken out；

Step 5.2.2.2: it if innermost circle boundary line is divided into stem portion with the situation of change of slope, defines boundaries substantially Geometry；

Step 5.2.2.3: repairing from breakpoint into closed figure according to geometry, and mend is made to keep original geometry special Property；

Step 5.2.3: for the isolated boundary line inside image, i.e., with other broken strings all apart from farther away isolated boundary line, then It is deleted；

Step 5.3: skeletal extraction is carried out again, by the fractional refinement filled up in step 5.2 at the object boundary line of single pixel wide；

Step 5.4: traversing image again, delete all inc object boundary lines, with obtaining complete fine polygon Object object boundary.