A large image vectorization method for remote sensing thematic information

By chunking remote sensing images and ensuring seamless vector connections at boundaries, the problems of fine polygons and overlap in large images vectorization are solved, and the seamless vectorization results of large areas are achieved, which are suitable for GIS spatial analysis of super-large images.

CN114419191BActive Publication Date: 2024-12-06SOUTHEAST DIGITAL ECONOMY DEV INST
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202210031085.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-01-06
Publication Date
2024-12-06
Estimated Expiration
2042-01-06

AI Technical Summary

Technical Problem

During the vectorization process of large images, traditional methods are difficult to deal with super-large images, which makes the images unable to be read into memory at one time, and the vector results extracted after chunking are prone to fine polygons and overlap problems, making seamless splicing impossible.

Method used

Block the image into blocks of the same size, and by ensuring seamless vector connection at the boundaries, the block vectorization method is adopted to ensure that there is pixel overlap between each block, realize seamless synthesis of polygons, and obtain remote sensing images. Vectorization results for large regions.

Benefits of technology

It realizes seamless vectorization of large images, avoids fine polygons and overlap problems, ensures the integrity and consistency of vectorization results, and is suitable for GIS spatial analysis of super-large images.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114419191B_ABST
    Figure CN114419191B_ABST
Patent Text Reader

Abstract

In order to solve the problem that the image exceeds the computer memory when vectorizing remote sensing thematic information, the present invention proposes a large image vectorization method for remote sensing thematic information. By dividing the image into image blocks of the same size and then vectorizing the blocks, seamless synthesis of the images at the boundaries is achieved, and the vectorization result of the large area of ​​the remote sensing image is obtained. The characteristic of this method is that each block overlaps one pixel, so that the blocks can be connected without the appearance of fragmented polygons. By comparing with the vectorization results of small images, the method in this paper can obtain the same vectorization results as those without block division.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention proposes a large image vectorization method for remote sensing thematic information, and specifically solves the large image vectorization problem by seamlessly splicing block-by-block vectorized ground objects. Background Art

[0002] Remote sensing images record the true state of the surface through surface reflection or backscattering information, and represent the spatial structure of objects through the arrangement of objects. However, this information is indirect and implicit, and cannot be directly applied to all walks of life. Remote sensing information extraction is the process of extracting information reflecting surface characteristics from remote sensing images, and is the process of transforming data into information and knowledge. In the past two decades, with the improvement of spatial resolution of remote sensing images, the size of images covering the same study area has grown rapidly, reaching several GB or even more than ten GB. These images are difficult to read into memory at one time, which brings difficulties to image processing. To address the above problems, large images are generally divided into image blocks of a certain size, and processed in blocks to break through the limitations of computer memory.

[0003] According to the different ways of determining the classification results of each pixel of the remote sensing image, the current remote sensing information extraction algorithms can be divided into two categories: the first category is pixel-based algorithms, that is, the result of each pixel is determined by the pixel value at the corresponding position, which can be expressed as:

[0004] y r,c =f(x r,c ) (1)

[0005] Among them, x r,c Represents the value of the input pixel, y r,c represents the result of information processing. Since the result is only affected by x r,c The second type is based on convolution algorithms: that is, the result value of each pixel is not only affected by the input pixel value at the corresponding position, but also by the pixel value in the convolution window around the pixel. The process can be expressed as:

[0006] y r,c =f(N(x r,c )) (2)

[0007] Among them, N(x r,c ) represents x r,c The convolution window centered on . Since the pixels at the edge of the block are affected by the pixels in the convolution window, whether or not to block the result and how to block the result will lead to differences in the result. In this way, the result after the block results are spliced ​​together cannot be aligned, that is, the edges cannot be accurately aligned.

[0008] In deep learning, the feature extraction of all pixels is based on the convolution window, which leads to differences in the results obtained by whether or not to divide the image into blocks and different ways of dividing the image into blocks. To address the above problems, deep learning generally overlaps and crops the images to a certain extent, so that there is a certain degree of overlap between different images, and only selects the middle part, that is, the result that is not affected by the edge, as the result. This method can extract information from the large image obtained.

[0009] However, the thematic information obtained from remote sensing images is generally represented by raster, that is, the surface area characteristics are simulated and expressed through the description of pixels. This expression method is simple and intuitive, but it is difficult to apply to GIS spatial analysis. Vectorization is the process of converting thematic information expressed in raster into vector data, which can improve the efficiency of thematic information analysis and is more suitable for GIS spatial analysis. It is the basis for remote sensing thematic information to provide spatial services (Zhang Xiaocan, Pan Yunhe. Raster data vectorization technology based on "raster technology" in GIS [J]. Journal of Computer-Aided Design and Graphics, 2001 (10); Song Yanru, Liu Qi, Fu Yuying, et al. Research on the vectorization method of remote sensing interpretation results [J]. Science and Technology and Innovation, 2021 (11): 82-85.).

[0010] Traditional vectorization methods are based on boundary tracing based on remote sensing images. Commercial software such as ArcGIS and open source tools such as GDAL provide vectorization methods (www.gdal.org), but these methods all read the image into memory at one time and vectorize the entire image at one time. This is a more appropriate method when the image size is small. However, in practice, the image containing the entire study area is too large to be read into memory at one time, making the vectorization algorithm difficult to implement (Rong Yuecheng. Dynamic vectorization algorithm for large-scale remote sensing classification images [J]. Acta Geodaetica et Cartographica Sinica, 2012, 41(06): 898-903.).

[0011] For the problem of vectorization of remote sensing thematic information of large images, an intuitive idea is to divide the remote sensing thematic information image into blocks and then vectorize it. This method can break through the image limitation. However, the vector results extracted after the image is divided into blocks are prone to fragmented polygons (holes or overlaps). If a certain feature is divided into different image blocks, the feature is divided into multiple polygons.

[0012] In view of the above problems, the present invention proposes a large image vectorization method for remote sensing thematic information. The method divides the image into image blocks of the same size and then vectorizes the blocks. By ensuring that the vectors at the boundaries are seamlessly connected, seamless synthesis of the images at the boundaries is achieved, thereby obtaining the result of large-area vectorization of remote sensing images. Summary of the invention

[0013] The present invention inputs remote sensing thematic information in raster format and uses a single-pixel width boundary E to represent different ground objects and targets. Since the raster image size exceeds the memory limit, it cannot be read in at one time. The present invention will use a block-by-block approach to convert each ground object target into a vector format object. The specific implementation method is as follows:

[0014] Step 1: Image segmentation

[0015] Assuming that the length and width of the original image are W and H respectively, it is divided into blocks of size w and h. The number of image blocks N(x) and N(y) can be expressed as:

[0016]

[0017]

[0018] in, Indicates rounding up.

[0019] According to the above process, each block B of the image can be expressed as:

[0020] B= i ,b j >,i=0,1,...,N(x)-1,j=0,1,...,N(y)-1 (3)

[0021] In order to ensure that a ground object crosses different image blocks and that there is one pixel overlap between each block, the image blocks i ,b j >Starting coordinate x(b i ,b j ) and y(b i ,b j )for:

[0022]

[0023]

[0024] The above process causes the actual processed size of each image block to increase by 1 relative to its original size.

[0025] Step 2: Block vectorization

[0026] The blocks are vectorized sequentially, and the vectorization algorithm uses any classic vectorization algorithm.

[0027] This step will obtain the polygonal area list A of each image block i,j , which can be expressed as:

[0028] ​​

[0029] Step 3: Extract the polygons to be merged

[0030] Extracting Segment Blocks i ,b j >The polygons on the right and bottom edges are added to the set M as new elements, and these polygons are merged with the polygons on the adjacent image blocks.

[0031] M is used to represent the list of polygons to be merged, and the element m∈M represents a group of polygons to be merged.

[0032] In order to determine the merging of polygons, four cases can be divided into:

[0033] (1) When i = 0, j = 0, no special treatment is required;

[0034] (2) When i>0, j=0, the image block i ,b j The polygon on the left edge needs to be aligned with the image block i-1 ,b j >Merge the polygons on the right. Extract image blocks i ,b j >Polygon on the left edge Determine whether it is consistent with the image block i-1 ,b j >Right polygon Is there an intersection of more than 2 pixels? If yes, Add to Afterwards, it indicates that the two polygons need to be merged.

[0035] (3) When i=0,j>0, the image block i ,b j >The polygon on the upper edge needs to be aligned with the image block i ,b j-1 >Merge the polygons at the lower edge of the image. Extract image blocks i ,b j >Polygon on the upper edge Determine whether it is consistent with the image block i ,b j-1 > Bottom polygon Is there an intersection of more than 2 pixels? If yes, Add to Afterwards, it indicates that the two polygons need to be merged.

[0036] (4) When i>0, j>0, add the left polygon and the right polygon to the merge list according to cases (2) and (3).

[0037] ​​​​​​​​​The above process maintains a list that records the polygons that need to be merged in the vectorization result.

[0038] Step 4: Update the merge list

[0039] For a polygon a∈A that appears in m i ∈M,m j ∈M, i≠j, then merge them, that is, m j Add the elements in m to i and m j Removed from M.

[0040] Step 5: Polygon merging

[0041] m i ∈M, merge the polygons therein to obtain a new polygon, and delete the original polygon.

[0042] The object of the present invention is to keep one pixel overlap between the left and right sides so that polygons can be connected together to obtain a seamless vectorization result. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 Flowchart of the present invention

[0044] Figure 2 Input remote sensing thematic information map of the present invention

[0045] Figure 3 The block vectorization result of the present invention is

[0046] Figure 4 Schematic diagram of polygon merging of the present invention

[0047] Figure 5 Schematic diagram of concave polygon merging

[0048] Figure 6 The overall vectorization result of the present invention is

[0049] Figure 7 Vectorize results for large areas DETAILED DESCRIPTION

[0050] The specific embodiments of the present invention are combined with the attached Figure 1 To explain:

[0051] The remote sensing thematic information extraction algorithm assigns each pixel on the remote sensing image to a predefined category according to a certain method. Remote sensing information extraction can use instance segmentation, semantic segmentation and other methods. Different types of objects have different labels. It is worth noting that category labels are not necessary in the thematic information vector process. Different objects are represented by single-pixel-wide boundaries. Single pixels are used to ensure the uniqueness of the path when tracing the boundaries. The input of the present invention is as follows Figure 2 , white represents the boundary of the object, and the length and width of the image are both 2000×2000 (pixels). It should be noted that the present invention can theoretically be applied to images of any size. 2000×2000 is selected here for the convenience of illustrating the problem.

[0052] Step 1: Image segmentation

[0053] The image is divided into 750×750 blocks, and the image will be divided into 9 blocks with 3 rows and 3 columns. It should be noted that each block overlaps one pixel, which can ensure that different blocks can be connected and there are no broken polygons caused by inconsistent vectorization between blocks.

[0054] The block size is a configurable parameter. On a machine with a memory size of 16G, it can also run if it is set to 10000×10000.

[0055] Step 2: Block vectorization

[0056] For a certain block, start from any unprocessed point and trace in a clockwise or counterclockwise direction. When the traced path returns to the starting point, a closed polygon is obtained. When all boundary pixels on the boundary are traced, the tracing of all boundaries on the block is completed.

[0057] Complete the vector tracking of all blocks from left to right and from top to bottom.

[0058] The vectorization results of the above results are as follows Figure 3 shown.

[0059] Step 3: Extract polygons to be merged

[0060] from Figure 3 It can be seen that there are dividing edge lines between the upper and lower, and left and right adjacent blocks. For the left and right adjacent blocks, the left edge of the right block is adjacent to the right edge of the left block, and the relevant polygons need to be merged. For the upper and lower adjacent blocks, the upper edge of the lower block is adjacent to the lower edge of the upper block.

[0061] Figure 4The left and right boundaries are seamlessly connected, but the same feature is divided into two pieces by the edge line, and the polygons marked as L and R in the figure need to be merged. Based on the above analysis, the present invention marks the polygons with more than two overlapping pixels on the left and right sides as polygons to be merged.

[0062] Step 4: Update the merge list

[0063] For a "concave" shaped polygon, it may be split into three parts, for example Figure 5 L 1 , L 2 and R, because when the left polygon is merged, it is not known that L 1 , L 2 The polygon R will be connected at the same time, so after the merge list is completed, the two sets with the same polygon are merged, that is, the two sets are merged together and only one identical set is retained.

[0064] Step 5: Polygon Merging

[0065] The merge list contains two or more polygons, and a seamless new polygon is obtained. The polygon obtained by the method of the present invention is as follows: Figure 6 shown.

[0066] Since the 2000×2000 polygon can be directly vectorized, the block vector results of this paper are compared with the direct one-time vector results. All points, lines and surfaces are consistent, indicating that the use of block vectorization can achieve the same results as the overall vectorization, which illustrates the effectiveness of this method.

[0067] In order to further verify the applicability of the proposed method to large images, the plot data of Zhenhai District, Ningbo City, Zhejiang Province was selected for the experiment. The image size is 61952×39424. The tif format image occupies 2.32G of hard disk space and cannot be read into the memory for processing at one time. The above-mentioned block division method is used for vectorization, and a total of 47888 patches are obtained. The results are shown in the figure. Figure 7 As shown, it is explained that the method of the present invention can be directly vectorized.

[0068] The above is only a description of an embodiment of the present invention, but the protection scope of the present invention should not be regarded as limited to the specific forms described in the embodiment. The protection scope of the present invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the concept of the present invention.

Claims

1. A large image vectorization method for remote sensing thematic information, characterized in that: Input remote sensing thematic information in grid format, use single-pixel width boundaries to represent different ground objects, and this method uses a block-by-block approach to convert each ground object into a vector format object V, including the following steps: Step 1: Image segmentation Assuming that the length and width of the original image are W and H respectively, it is divided into blocks of size w and h. The number of image blocks N(x) and N(y) can be expressed as: in, Indicates rounding up; According to the above process, each block B of the image can be expressed as: B=<b i ,b j >,i=0,1,…,N(x)-1,j=0,1,...,N(y)-1 (3) In order to ensure that a ground object target has intersections in different image blocks and that there is one pixel overlap between each block, image block < b i , b j >Starting coordinate x(b i , b j ) and y(b i , b j )for: The above process increases the actual processing size of each image block by 1 relative to the original size. In the above image segmentation method, each segment overlaps only one pixel. Step 2: Block vectorization Vectorize the blocks one by one, This step will obtain the polygonal area list A of each image block i,j It is expressed as: Step 3: Extract the polygons to be merged Extract segment block <b i , b j > The polygons on the right edge and the bottom edge are added to the set M as new elements. These polygons will be merged with the polygons on the adjacent image blocks. M is used to represent the list of polygons to be merged. The element m∈M represents a set of polygons to be merged. In order to determine the merging of polygons, it is divided into four cases: (1) When i=0, j=0, no special treatment is required; (2) When i>0, j=0, image block < b i , b j >The polygon on the left edge needs to be aligned with the image block <b i-1 , b j >Merge the polygons on the right and extract the image block <b i , b j >Polygon on the left edge Determine whether it is consistent with the image block <b i-1 , b j >Right polygon Is there an intersection of more than 2 pixels? If yes, Add to Afterwards, it indicates that the two polygons need to be merged; (3) When i=0, j>0, image block < b i , b j > The polygon on the upper edge needs to be aligned with the image block i , b j-1 >The polygons at the lower edge of the image are merged to extract the image block <b i , b j >Polygon on the upper edge Determine whether it is consistent with the image block <b i , b j-1 >Lower polygon Is there an intersection of more than 2 pixels? If yes, Add to Afterwards, it indicates that the two polygons need to be merged;​ (4) When i>0, j>0, add the left polygon and the right polygon to the merge list according to cases (2) and (3); The above process maintains a list to record the polygons that need to be merged in the vectorization result; Step 4: Update the merge list For a polygon a∈A that appears in m i ∈M,m j ∈M, i≠j, then merge them, that is, m j Add the elements in m to i and m j Delete from M; Step 5: Polygon merging m i ∈M, merge the polygons therein to obtain a new polygon, and delete the original polygon.

2. The large image vectorization method for remote sensing thematic information according to claim 1 is characterized in that: In step three, each polygon is merged with a polygon that has two adjacent pixels, and then the polygon merging and updating in step four is performed to ensure that each ground object has a unique polygon corresponding to it.

Citation Information

Cited By

  • A remote sensing image processing method supporting dynamic geographic alignment and metadata fusion

    CN122676321A