A method for scene boundary recognition in remote sensing images consistent with human visual perception

Abstract

The present invention discloses a remote sensing image scene boundary recognition method that conforms to human visual perception. A remote sensing raster image is input, and the image is segmented at the object level and the scene level respectively to obtain the corresponding object spot raster image and the initial scene spot raster image, and then the initial scene spot is updated with the object spot to obtain the final scene spot raster image, and finally the final scene spot raster image is subjected to a raster-to-vector operation to obtain the remote sensing image scene boundary that conforms to human visual perception. The method is simple in principle, easy to implement, and highly automated, and has a wide range of application value in the fields of remote sensing and geography, and is not only helpful in identifying suspicious targets in various scenes of remote sensing images, but also helpful in determining ecological landscape boundaries, regional planning boundaries, and natural geographical zoning boundaries.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing image processing, and in particular to the field of remote sensing image segmentation. Background Art

[0002] Although remote sensing images can intuitively present landscapes such as forests, grasslands, farmlands, water bodies, and cities, it is difficult to distinguish clear boundaries between landscapes. Remote sensing image scenes that conform to human visual perception refer to landscape areas with specific functions that are geographically adjacent and dominated by a certain type of landform on remote sensing images. Identifying the boundaries of remote sensing image scenes that conform to human visual perception has broad application value in the fields of remote sensing and geography. It not only helps to identify suspicious targets in various scenes of remote sensing images, but also helps to determine ecological landscape boundaries, regional planning boundaries, and natural geographical division boundaries.

[0003] Remote sensing image scene boundary recognition is essentially image segmentation. Traditional image segmentation methods include threshold method, region-based method, edge-based method, clustering-based method, graph-based method and morphological watershed segmentation method. These traditional image segmentation methods are mainly based on the feature differences of a single image for segmentation, and it is often difficult to achieve ideal segmentation results on images with complex backgrounds. The collaborative segmentation method can extract common background areas from multiple images to obtain prior knowledge to assist image segmentation. The multi-scale segmentation method uses image information of different scales to obtain hierarchical segmentation results. Both the collaborative segmentation method and the multi-scale segmentation method have improved the segmentation effect of complex background images to a certain extent. With the continuous development of image acquisition equipment, the complexity of image details has greatly increased. Low-level image features or image features extracted by heuristic rules cannot meet the complex needs of current image segmentation. Deep learning models are widely used in the field of image segmentation to build semantic segmentation classifiers to complete more complex image segmentation and classification tasks.

[0004] However, whether it is the traditional image segmentation method, the collaborative segmentation method, the multi-scale segmentation method or the semantic segmentation of deep learning, their main task is to segment specific objects from the image, rather than the scene in the image; although the multi-scale segmentation method under large-scale parameters can segment the scene, the segmented scene does not conform to human visual perception. Therefore, the existing segmentation methods have not yet achieved scene boundary recognition that conforms to human visual perception.

[0005] When human vision perceives an image, it first perceives the overall average state of the image, then perceives the local details of the image, and gives priority to continuous areas in the image during the perception process. Inspired by the human visual perception process, the basic idea of ​​remote sensing image scene boundary recognition that conforms to human visual perception is to first calculate the average size of the image patch from a macroscopic perspective as the basic unit of computer visual perception, and then use the average size of the image patch to guide the image upscaling scene-level segmentation to obtain the initial scene patch, and then update the initial scene patch with the object patch obtained by combining the original image object-level segmentation, and finally perform a raster-to-vector operation on the final scene patch raster image to obtain the remote sensing image scene boundary that conforms to human visual perception. Summary of the invention

[0006] In view of the shortcomings of existing image segmentation methods in identifying scene boundaries of remote sensing images, the present invention proposes a remote sensing image scene boundary recognition method that conforms to human visual perception, aiming to identify scene boundaries on remote sensing images that conform to human visual perception. The method has a simple principle, is easy to implement, has a high degree of automation, and has a wide range of applications.

[0007] To achieve this object, the present invention adopts the following technical solutions:

[0008] A remote sensing image scene boundary recognition method that conforms to human visual perception comprises the following steps:

[0009] A. Image input: input a remote sensing raster image.

[0010] B. Image object level segmentation: Perform object level segmentation on the image to obtain an object patch grid image.

[0011] When performing object-level segmentation on an image, a variety of segmentation methods can be used, such as multi-scale segmentation methods, superpixel segmentation methods, etc.

[0012] C. Image upscaling and scene-level segmentation: Calculate the optimal spatial resolution for upscaling that matches human visual perception of the image scene, smoothly upscale the image to the optimal spatial resolution that matches human visual perception of the image scene, perform scene-level segmentation on the upscaled image, and obtain the initial scene patch grid image.

[0013] When calculating the optimal spatial resolution for upscaling that meets human visual perception of image scenes, it is necessary to cluster the image according to the number of main land cover types in the image to obtain cluster patches, then calculate the mean area of ​​each cluster patch in the image, and finally take the square root of the mean area of ​​the cluster patches as the optimal spatial resolution for upscaling that meets human visual perception of image scenes. Land cover types can be summarized into five categories: vegetation, water bodies, bare land, artificial surfaces, and ice and snow. Therefore, when automating this step, it is preferred to set the number of clusters to no more than 5. When clustering images, a clustering method with a predetermined number of categories can be used, such as the K-means clustering method.

[0014] When upscaling an image, it must be smoothed, otherwise it is impossible to obtain an ideal upscaled image. Image smoothing can be done by using methods such as mean filtering and Gaussian smoothing filtering.

[0015] When performing scene-level segmentation on the upscaled image, a clustering method with a settable cluster number range can be used for segmentation, such as the ISODATA clustering method. Since the surface cover types can be summarized into five categories: vegetation, water bodies, bare land, artificial surfaces, and ice and snow, it is preferred to set the minimum cluster number to 1 and the maximum cluster number to 5 when automating this step.

[0016] D. Scene patch update: Use the object patch to update the initial scene patch to obtain the final scene patch raster image.

[0017] Based on the area ratio of the object patch in the initial scene patch, the object patch will be incorporated into the scene patch in which its area ratio is larger.

[0018] E. Scene boundary extraction: perform raster-to-vector operations on the final scene patch raster image to obtain the remote sensing image scene boundary that conforms to human visual perception.

[0019] The present invention has the following characteristics:

[0020] (1) The principle is simple and easy to implement.

[0021] (2) High degree of automation.

[0022] (3) It has a wide range of applications and can be widely used for boundary extraction of various scenes in remote sensing and geographic fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a technical process for identifying the boundaries of remote sensing image scenes that conform to human visual perception.

[0024] Figure 2 This is the remote sensing image used in the case.

[0025] Figure 3 is the object patch raster image.

[0026] Figure 4 is the image upscaling result.

[0027] Figure 5 is the initial scene patch raster image.

[0028] Figure 6 is the final scene patch raster image.

[0029] Figure 7 It is the remote sensing image scene boundary recognition result that conforms to human visual perception. DETAILED DESCRIPTION

[0030] The technical implementation scheme of the present invention is further described below in conjunction with the accompanying drawings.

[0031] Based on a remote sensing raster image, we now perform remote sensing image scene boundary recognition that conforms to human visual perception. The overall technical process is shown in Figure 1 The image is a three-band image of red, green and blue, with a spatial resolution of 30m, a total of 2048 columns × 2048 rows of pixels, and the value range of each band of red, green and blue is 0-255, byte data.

[0032] A. Image input: input a remote sensing raster image.

[0033] In this case, a remote sensing raster image is input, namely Figure 2 .

[0034] B. Image object level segmentation: Perform object level segmentation on the image to obtain an object patch grid image.

[0035] In this case, the Felzenszwalb superpixel segmentation method is used to perform object-level segmentation on the image, and the object patch grid image is obtained. Figure 3 .

[0036] C. Image upscaling and scene-level segmentation: Calculate the optimal spatial resolution for upscaling that matches human visual perception of the image scene, smoothly upscale the image to the optimal spatial resolution that matches human visual perception of the image scene, perform scene-level segmentation on the upscaled image, and obtain the initial scene patch grid image.

[0037] This case uses the K-means clustering method and sets the number of clusters to 5. The cluster patches are obtained and the mean area of ​​the cluster patches is 69066.091m 2 , the square root of this number, 262.8 m, is used as the optimal spatial resolution for image upscaling.

[0038] In this case, the Gaussian pyramid method is used to smooth and upscale the image. First, the image is smoothed with a Gaussian convolution kernel, and then the even rows and even columns of the smoothed image are replaced with 0. Finally, the deconvolution kernel of the smoothed convolution kernel is used to traverse the image to obtain the image upscaling result. Figure 4 .

[0039] In this case, the ISODATA automatic clustering algorithm is used to perform scene-level clustering segmentation on the upscaled image. The input features used for clustering are the red, green, and blue bands of the image. The minimum number of clusters is set to 1 and the maximum number of clusters is set to 5. The initial scene patch grid image is obtained. Figure 5 .

[0040] D. Scene patch update: Use the object patch to update the initial scene patch to obtain the final scene patch raster image.

[0041] In this case, for each object spot in the object spot grid image, all pixels at the corresponding position in the initial scene spot grid image are extracted one by one, and the mode of these pixel scene labels is used as the new label of the object spot. After traversing and processing all object spots, the final scene spot grid image is obtained. Figure 6 .

[0042] E. Scene boundary extraction: perform raster-to-vector operations on the final scene patch raster image to obtain the remote sensing image scene boundary that conforms to human visual perception.

[0043] In this case, QGIS software is used to convert the final scene raster image into a vector image to obtain the remote sensing image scene boundary that conforms to human visual perception. Figure 7 .

[0044] The above is only a specific implementation of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by anyone familiar with the technology within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A remote sensing image scene boundary recognition method that conforms to human visual perception, characterized in that: The following steps are involved: A. Image input: input a remote sensing raster image; B. Image object level segmentation: perform object level segmentation on the image to obtain an object patch grid image; C. Image upscaling and scene-level segmentation: Calculate the optimal spatial resolution of upscaling that meets human visual perception of the image scene, smoothly upscale the image to the optimal spatial resolution of human visual perception of the image scene, perform scene-level segmentation on the upscaled image, and obtain the initial scene patch grid image; D. scene patch update: using the object patch to update the initial scene patch to obtain the final scene patch grid image; E. Scene boundary extraction: perform raster-to-vector operations on the final scene patch raster image to obtain the remote sensing image scene boundary that conforms to human visual perception.

2. According to the method for identifying the boundary of a remote sensing image scene that conforms to human visual perception in claim 1, in said step C, the optimal spatial resolution of the upscaling of the image scene that conforms to human visual perception is calculated, characterized in that: The images are clustered according to the number of main land cover types in the images to obtain cluster patches, the mean area of ​​the cluster patches in the images is calculated, and the square root of the mean area of ​​the cluster patches is taken as the optimal spatial resolution for upscaling that conforms to human visual perception of the image scene.

3. According to the method for identifying scene boundaries of remote sensing images in accordance with human visual perception in claim 1, in step C, the image is smoothly upscaled, characterized in that: The image must be smoothed before upscaling.

4. According to the method for recognizing scene boundaries of remote sensing images in accordance with human visual perception in claim 1, in step D, the initial scene spots are updated using the object spots, characterized in that: Based on the area ratio of the object patch in the initial scene patch, the object patch will be incorporated into the scene patch in which its area ratio is larger.

Images