A method for extracting buildings from large remote sensing images supported by scene classification

Abstract

The present invention provides a large-scale remote sensing image building extraction method supported by scene classification, including: S1: obtaining a large-scale remote sensing image; S2: setting a sliding window, reading the entire remote sensing image, and obtaining sliding window data; S3: using a scene classification network model to classify the sliding window data into two categories: building scenes and non-building scenes; S4: judging whether the scene classification prediction result of the sliding window data is a building scene category according to the classification result of the scene classification network model, if so, inputting the sliding window data into the building extraction network model to obtain a building prediction result, if not, obtaining a prediction result of all non-buildings; S5: judging whether to traverse the entire remote sensing image, if so, outputting the large-scale image building prediction result, if not, returning to step S2. The beneficial effects of the present invention are: saving the time of large-scale remote sensing image building extraction and improving the efficiency of large-scale remote sensing image building extraction.

Description

Technical Field

[0001] The present invention relates to the technical field of image recognition, and in particular to a method for extracting buildings from large-scale remote sensing images supported by scene classification. Background Art

[0002] Buildings are an important carrier for human survival and one of the most important elements in building smart cities. With the rapid development of satellite remote sensing technology, building extraction based on high-resolution remote sensing image data has become one of the most important research directions in remote sensing image interpretation. The building extraction method has also changed from a method based on manual features to a convolutional neural network based on high-level features to automatically extract from high-resolution remote sensing images. The method for extracting buildings from large-scale remote sensing images generally adopts prediction based on a semantic segmentation network model. However, since buildings are concentratedly distributed, there are a large number of non-building areas in large-scale remote sensing images. Extracting buildings from non-building areas will reduce the image processing efficiency. Based on the above problems, the present invention proposes a method for extracting buildings from large-scale remote sensing images supported by scene classification. Compared with traditional full-image semantic segmentation building extraction, this method can improve the efficiency of building extraction. Summary of the invention

[0003] In order to solve the above problems, the present invention provides a method for extracting buildings from large remote sensing images supported by scene classification, which mainly includes the following steps:

[0004] S1: Acquire large remote sensing images;

[0005] S2: Set the sliding window, read the entire remote sensing image, and obtain the sliding window data;

[0006] S3: using the scene classification network model to classify the sliding window data obtained in step S2 into two categories: building scenes and non-building scenes;

[0007] S4: According to the classification result of the scene classification network model, determine whether the scene classification prediction result of the sliding window data is a building scene category. If so, input the sliding window data into the building extraction network model to obtain a building prediction result. If not, the sliding window data does not contain a building and obtains a prediction result of all non-buildings;

[0008] S5: Determine whether to traverse the entire remote sensing image. If so, output the large image building prediction result. If not, return to step S2.

[0009] Furthermore, the sliding window reads the large image in sequence with a fixed-size window and step size, starts reading the window data from the upper left corner of the input large remote sensing image, slides the window row by row, and reads the next row after one row is read until the entire image is traversed.

[0010] Furthermore, the scene classification network model F scene-class (·) is a binary classification, then

[0011] pre category =F scene-calss (X i )

[0012] Among them, pre category For buildings or non-buildings.

[0013] Furthermore, for the sliding window data judged as the building scene category, a building extraction network model based on semantic segmentation is used to extract buildings from the sliding window data.

[0014] Furthermore, the building extraction network model is F building (·),but:

[0015] pre Building =F building (x i-Building )

[0016] Among them, pre Building For window data X i-Building Building prediction results, pre Building If the pixel value is predicted to be a building, it is assigned a value of 1, and if it is predicted to be a non-building, it is assigned a value of 0.

[0017] Furthermore, through the judgment in step S4, if the scene classification prediction result is non-building, then X i-nonBuilding The building prediction result of the data window is pre nonBuilding , pre nonBuilding All pixel values ​​in X are assigned to 0; if the scene classification prediction result is a building, then i-Building The building prediction result of the data window is pre Building ; will pre nonBuilding and pre Building The combined output is the final result Y, which is the building prediction result of the large-scale remote sensing image.

[0018] The technical solution provided by the present invention has the beneficial effects of saving the time for extracting buildings from large-scale remote sensing images and improving the efficiency of extracting buildings from large-scale remote sensing images. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:

[0020] Figure 1 It is a flow chart of a method for extracting buildings from large remote sensing images supported by scene classification in an embodiment of the present invention. DETAILED DESCRIPTION

[0021] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, specific embodiments of the present invention are now described in detail with reference to the accompanying drawings.

[0022] An embodiment of the present invention provides a method for extracting buildings from large remote sensing images supported by scene classification.

[0023] Please refer to Figure 1 , Figure 1 1 is a flow chart of a method for extracting buildings from large remote sensing images supported by scene classification in an embodiment of the present invention, which specifically includes the following steps:

[0024] (1) Obtain a large remote sensing image and set a sliding window: The sliding window is a process of sequentially reading large image data using a fixed-size window (e.g., 512×512) and a step size. The window data is read starting from the upper left corner of the input large remote sensing image X, and the window is slid row by row. After one row is read, the next row is read until the entire image is traversed. Suppose the read sliding window data is X i .

[0025] (2) Scene classification network model: Scene classification is to assign one or more meaningful scene category labels with high-level semantic information to the image based on the image data content. The scene classification network model in the present invention is a binary classification, that is, it is divided into building scene and non-building scene categories. Let the scene classification network model be F scene-class (·),but:

[0026] pre category =F scene-calss (X i )

[0027] Among them, pre category For buildings or non-buildings.

[0028] (3) Determine whether it is a building: Determine whether the sliding window data X i The scene classification prediction results pre category Is it a building? If so, then X i Marked with X i-Building And execute step (4). If it is not a building, then X i Marked with X i-nonBuilding And skip step (4) and execute step (5). Since the scene classification network model has fewer decoding layers than the building extraction network model, the scene classification network model saves processing time than the building extraction network model under the same conditions. In this way, the judgment method avoids i Perform building extraction operations to improve the efficiency of building extraction from large remote sensing images.

[0029] (4) Building extraction network model: The building extraction network model adopts the semantic segmentation method and consists of an encoder and a decoder to classify and predict each pixel value of the image data. Let the building extraction network model be F building (·),but:

[0030] pre Building =F building (x i-Building )

[0031] Among them, pre Building For window data X i-Building Building prediction results. Building If the pixel value is predicted to be a building, it is assigned a value of 1, and if it is not a building, it is assigned a value of 0.

[0032] (5) Result synthesis and output: According to the judgment in step (3), if it is not a building, then X i-nonBuilding The building prediction result of the data window is pre nonBuilding , pre nonBuilding All pixel values ​​in the X are assigned to 0; if it is a building, then i-Building The building prediction result of the data window is pre Building . nonBuilding and pre Building The combined output is the final result Y. Y is the building prediction result of the large remote sensing image.

[0033] In the above steps, the scene classification network model and the building extraction network model belong to the existing technology. Before the building extraction network model is used to predict the large-scale remote sensing image, the scene classification network model is used to determine whether it is a building scene category. If it is a non-building scene category, the building extraction network model is skipped. If it is a building scene category, the building extraction network model prediction is put in, so as to avoid the building extraction process of non-building scenes and increase the building extraction efficiency of large-scale remote sensing images.

[0034] Compared with the prior art, the invention has the following beneficial effects: by using a sliding window method, end-to-end building extraction can be achieved in large remote sensing images. By classifying scenes to determine whether they are building scenes, the building extraction process in non-building scene areas is avoided, thereby improving the building extraction speed of large remote sensing images.

[0035] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for extracting buildings from large remote sensing images supported by scene classification, characterized by: The following steps are involved: S1: Acquire large remote sensing images; S2: Set a sliding window, read the entire remote sensing image, and obtain sliding window data; the sliding window reads large images in sequence with a fixed-size window and step size, starts reading window data from the upper left corner of the input large remote sensing image, slides the window row by row, and reads the next row after one row is read, until the entire image is traversed; S3: Using a scene classification network model to classify the sliding window data obtained in step S2 into two categories: building scenes and non-building scenes; the scene classification network model F scene-class (·) is a binary classification, then pre category =F scene-calss (X i ) Among them, pre category be a building or non-building; S4: According to the classification result of the scene classification network model, determine whether the scene classification prediction result of the sliding window data is a building scene category. If so, input the sliding window data into the building extraction network model to obtain a building prediction result. If not, the sliding window data does not contain a building and obtains a prediction result of all non-buildings. For the sliding window data judged as a building scene category, use the building extraction network model based on semantic segmentation to extract buildings from the sliding window data. The building extraction network model is F building (·),but: pre Building =F building (X i-Building ) Among them, pre Building For window data X i-Building Building prediction results, pre Building If the pixel value is predicted to be a building, it is assigned a value of 1, and if it is predicted to be a non-building, it is assigned a value of 0; S5: Determine whether to traverse the entire remote sensing image, if so, output the large image building prediction result, if not, return to step S2; According to the judgment of step S4, if the scene classification prediction result is non-building, then X i-nonBuilding The building prediction result of the data window is pre nonBuilding , pre nonBuilding All pixel values ​​in X are assigned to 0; if the scene classification prediction result is a building, then i-Building The building prediction result of the data window is pre Building ; will pre nonBuilding and pre Building The combined output is the final result Y, which is the building prediction result of the large-scale remote sensing image.

Images