Building roof type remote sensing identification method using residual error capsule network

Abstract

The invention relates to the technical field of image recognition, in particular to a building roof type remote sensing recognition method utilizing a residual error capsule network, and the method comprises the steps: obtaining sample roof remote sensing images, and obtaining a roof sample library; the number of roof remote sensing images in the roof sample library is M, and the roof remote sensing images are respectively recorded as first to Mth images; the actual number of eave lines, the actual number of ridge lines and the actual number of waist lines of the corresponding roof and the actual included angles between every two eave lines, ridge lines and waist lines are obtained; training the initial residual capsule neural network parameters to obtain a first residual capsule neural network; obtaining a to-be-measured roof remote sensing image; inputting the first residual capsule neural network, and outputting to obtain the number of eave lines, the number of ridge lines, the number of waist lines and included angles between every two eave lines, ridge lines and waist lines of the to-be-detected roof; and matching to obtain the to-be-measured roof type. According to the invention, building roof type identification is realized through image identification.

Description

Technical Field

[0001] The present invention relates to the technical field of image recognition, and in particular to a remote sensing recognition method for building roof types using a residual capsule network. Background Art

[0002] In the field of building remote sensing identification, accurately acquiring roof type information is crucial for applications such as urban planning, solar energy potential assessment, and low-altitude aircraft route planning. Building roof types are primarily distinguished based on structural features, such as the number of eaves, ridges, and waistlines, as well as their spatial relationships (e.g., angles between each pair). Currently, using high-resolution cameras mounted on drones to acquire remote sensing images and automatically identify roof types based on deep learning techniques has become a major research direction. Existing technologies primarily employ convolutional neural networks for image recognition.

[0003] Using convolutional neural networks to identify roof features is a feasible approach. However, while convolutional neural networks excel at identifying local features, such as the eaves, ridges, and waistlines of a roof, this approach struggles to capture the spatial relationship between these lines, resulting in insufficient recognition accuracy for complex roofs.

[0004] With the advancement of image recognition technology, capsule neural network models are gaining increasing application. The advantage of capsule neural network models lies in their ability to accurately identify the spatial relationships between the entire image and its local components. Therefore, they can be used to overcome the limited ability of convolutional neural networks to recognize the spatial relationships between eaves, ridges, and waistlines. Therefore, there is an urgent need to combine convolutional neural networks with capsule neural networks to accurately identify building roof types. Summary of the Invention

[0005] (1) Technical problems to be solved The purpose of the present invention is to provide a remote sensing recognition method for building roof types using a residual capsule network to achieve detection of complex roof categories.

[0006] (2) Technical solution To achieve the above object, the present invention provides a remote sensing identification method for building roof types using a residual capsule network, the method comprising the following steps: S1, obtain sample roof remote sensing images through the camera fixed on the bottom of the drone to obtain a roof sample library; the number of roof remote sensing images in the roof sample library is M , respectively recorded as the first image to the M image.

[0007] S2, obtain the first image to the MThe actual number of eaves lines, actual number of ridge lines, actual number of waist lines and actual angles between eaves lines, ridge lines and waist lines of the roof corresponding to the image are obtained. M Data group; the first image to the M Image, first data set to M The data group is input into a pre-constructed initial residual capsule neural network, and the parameters of the initial residual capsule neural network are trained to obtain a first residual capsule neural network; the input of the first residual capsule neural network is the remote sensing image of the roof to be measured, and the output is the number of eaves lines, the number of ridge lines, the number of waist lines of the roof to be measured, and the angles between the eaves lines, the ridge lines, and the waist lines.

[0008] S3, obtains remote sensing images of the roof to be measured through a camera fixed on the bottom of the drone.

[0009] S4, inputting the remote sensing image of the roof to be measured into the first residual capsule neural network, and outputting the number of eaves lines, the number of ridge lines, the number of waist lines, and the angles between the eaves lines, the ridge lines, and the waist lines of the roof to be measured; matching the number of eaves lines, the number of ridge lines, the number of waist lines, and the angles between the eaves lines, the ridge lines, and the waist lines of the roof to be measured with a pre-obtained roof type library to obtain the roof type to be measured.

[0010] Furthermore, the initial residual capsule neural network includes a standard convolutional layer, a capsule layer, and a feedback iterative layer.

[0011] Furthermore, the standard convolutional layer uses a convolutional neural network algorithm to extract the first image to the M The local images of the eaves, ridges and waistlines in the image; the convolutional neural network algorithm is used to extract the local images from the first image to the M The parameters used in the process of local images of eaves, ridges, and waistlines in the image are recorded as standard convolution layer parameters; the standard convolution layer parameters include convolution kernel, convolution kernel width, number of input channels, number of output channels, and bias top; the number of standard convolution layer parameters is obtained H , relabel the standard convolution layer parameters as the first standard convolution layer parameters to the H Standard convolutional layer parameters.

[0012] Furthermore, the capsule layer is based on the first image to the M The local images of the eaves, ridges and waistlines in the image are extracted using a dynamic routing algorithm from the first image to the second image. M The number of eaves lines, ridge lines, waist lines and the angles between the eaves lines, ridge lines and waist lines in the image are obtained from the first recognition data set to the second recognition data set. M Identify the data group; MThe local images of the eaves, ridges and waistlines in the image are extracted using a dynamic routing algorithm from the first image to the second image. M The number of eaves lines, ridge lines, waist lines and the angles between the eaves lines, ridge lines and waist lines in the image are obtained from the first recognition data set to the second recognition data set. M The parameters used in the process of identifying the data group are recorded as capsule layer parameters; the capsule layer parameters include the number of input capsules, the number of output capsules, the output capsule dimension, the input capsule dimension, and the coupling coefficient; the number of capsule layer parameters is obtained L , relabel the capsule layer parameters as the first capsule layer parameters to the L Capsule layer parameters.

[0013] Furthermore, the first identification data group to the M The identification data set is represented as: ; in, Indicates the Identify data groups; Indicates the the number of eave lines in the image; Indicates the The number of roof lines in the image; Indicates the The number of waist lines in the image; is a data set, indicating the The set of angles between the eaves, ridges, and waistlines in the image; The value range is 1 to M Integer variable between .

[0014] Furthermore, the feedback iterative layer uses a particle swarm optimization algorithm to optimize and calculate the standard convolution layer parameters and the capsule layer parameters to obtain the optimal standard convolution layer parameters and the optimal capsule layer parameters; the optimal standard convolution layer parameters and the optimal capsule layer parameters are respectively used as the standard convolution layer parameters and the capsule layer parameters of the first residual capsule neural network.

[0015] Furthermore, the particle swarm optimization algorithm is used to optimize and calculate the standard convolution layer parameters and the capsule layer parameters to obtain the optimal standard convolution layer parameters and the optimal capsule layer parameters. The method is as follows: Construct an optimization objective function; the optimization objective function is expressed as: ; in, represents the optimization objective function; represents the standard convolutional layer parameters; Represents the capsule layer parameters; Indicates the The actual number of eaves lines on the roof to which the image corresponds; Indicates the The actual number of ridge lines on the roof that the image corresponds to; Indicates the The actual number of waistlines on the roof to which the image corresponds; Indicates the The actual angles between the eaves line, ridge line and waist line of the roof corresponding to the image are Identify the first The sum of the absolute values ​​of the angle differences between the eaves line, ridge line, and waist line in the image, in radians.

[0016] Construct optimization constraints.

[0017] With the optimization constraints as constraints and the goal of minimizing the value of the optimization objective function, the particle swarm optimization algorithm is used to calculate the optimal values ​​of the standard convolutional layer parameters and the optimal values ​​of the capsule layer parameters, which are recorded as the optimal standard convolutional layer parameters and the optimal capsule layer parameters respectively.

[0018] Furthermore, the optimization constraints are: ; in, Indicates the t Standard convolutional layer parameters; Indicates the preset t The lower limit of standard convolutional layer parameters; Indicates the preset t Upper limit of standard convolutional layer parameters; Indicates the c Capsule layer parameters; Indicates the preset c The lower limit of capsule layer parameters; Indicates the preset c Upper limit of capsule layer parameters; t The value range is 1 to H Integer between ; c The value range is 1 to L Integer between .

[0019] Furthermore, the method of matching the number of eaves lines, the number of ridge lines, the number of waist lines, and the angles between the eaves lines, the ridge lines, and the waist lines of the roof to be tested with the pre-obtained roof type library to obtain the type of the roof to be tested is as follows: The roof type library includes preset roof types and the preset number of eaves lines, the preset number of ridge lines, the preset number of waist lines corresponding to the preset roof types, and the preset angles between the eaves lines, the ridge lines, and the waist lines; the preset roof types include the first type to theG Type; Type 1 to Type G The preset number of eaves lines, the preset number of ridge lines, the preset number of waist lines, and the preset angles between the eaves lines, ridge lines, and waist lines corresponding to the type are recorded as the first type characteristic parameters to the G Type characteristic parameters.

[0020] Calculate the number of eaves lines, ridge lines, waist lines, and the angles between the eaves lines, ridge lines, and waist lines of the roof to be tested, and the first type of characteristic parameters to the G weighted Euclidean distance of type characteristic parameters; the roof type corresponding to the minimum value of the weighted Euclidean distance is recorded as the roof type to be tested; in the calculation process of the weighted Euclidean distance, a preset first weight is assigned to the number of ridge lines, a preset second weight is assigned to the number of eaves lines, a preset third weight is assigned to the number of waist lines, and a preset fourth weight is assigned to the angles between the eaves lines, the ridge lines, and the waist lines; the first weight is greater than the second weight, the second weight is greater than the third weight, the third weight is greater than the fourth weight, and the sum of the first weight, the second weight, the third weight, and the fourth weight is 1.

[0021] (3) Beneficial effects Compared with the prior art, the present invention has the following beneficial effects: By inputting the remote sensing image of the roof to be tested into the first residual capsule neural network, the output is the number of eaves lines, ridge lines, and waist lines of the roof to be tested, as well as the angles between each of the eaves lines, ridge lines, and waist lines, and then matching the roof type to be tested. This method uses the angles between each of the eaves lines, ridge lines, and waist lines as recognition image features, allowing for more accurate identification of complex roof types. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 This is a flowchart of a method for remote sensing identification of building roof types using a residual capsule network according to Example 1 of the present invention. DETAILED DESCRIPTION

[0023] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0024] Before giving examples, it is necessary to explain the application scenarios of the present invention. This embodiment is used to accurately identify the roof types of buildings, and the complex spatial positional relationship between the eaves line, ridge line, and waist line is considered during the identification process.

[0025] Example 1: Figure 1 As shown, this embodiment provides a remote sensing identification method for building roof types using a residual capsule network, the method comprising the following steps: S1, obtain sample roof remote sensing images through the camera fixed on the bottom of the drone to obtain a roof sample library; the number of roof remote sensing images in the roof sample library is M , respectively recorded as the first image to the M image.

[0026] S2, obtain the first image to the M The actual number of eaves lines, actual number of ridge lines, actual number of waist lines and actual angles between eaves lines, ridge lines and waist lines of the roof corresponding to the image are obtained. M Data group; the first image to the M Image, first data set to M The data group is input into a pre-constructed initial residual capsule neural network, and the parameters of the initial residual capsule neural network are trained to obtain a first residual capsule neural network; the input of the first residual capsule neural network is the remote sensing image of the roof to be measured, and the output is the number of eaves lines, the number of ridge lines, the number of waist lines of the roof to be measured, and the angles between the eaves lines, the ridge lines, and the waist lines.

[0027] S3, obtains remote sensing images of the roof to be measured through a camera fixed on the bottom of the drone.

[0028] S4, inputting the remote sensing image of the roof to be measured into the first residual capsule neural network, and outputting the number of eaves lines, the number of ridge lines, the number of waist lines, and the angles between the eaves lines, the ridge lines, and the waist lines of the roof to be measured; matching the number of eaves lines, the number of ridge lines, the number of waist lines, and the angles between the eaves lines, the ridge lines, and the waist lines of the roof to be measured with a pre-obtained roof type library to obtain the roof type to be measured.

[0029] For example, data preparation is first performed. Specifically, a camera fixed on the bottom of the drone is used to take vertical aerial photos of the roofs of buildings in the target area to obtain high-resolution rooftop remote sensing images. These collected rooftop images are sorted to form a rooftop sample library for training the recognition model. The sample library contains a total of M The roof images are named from the first image to the second image according to the collection order or specific rules. M The core purpose of this step is to provide sufficient and representative raw image data for the subsequent construction and training of recognition models.

[0030] Then the model is constructed and trained. For each image in the roof sample library (i.e. the first image to theM Images) require manual interpretation or other high-precision measurement methods to accurately obtain the key geometric parameters of the corresponding roof structure. These key parameters include: the actual number of eaves lines, the actual number of ridge lines, the actual number of waist lines, and the actual angles between the eaves lines, ridge lines, and waist lines. Each image and its corresponding complete structural parameters, namely the quantity information and angle information, together constitute a data unit, which are respectively referred to as the first data group to the first data group. M Then, the first image to the M Image and corresponding data sets from the first to the M The data sets are collectively input into a pre-designed network framework, namely the initial residual capsule neural network. The parameters of the initial residual capsule neural network are pre-set. Model training is then performed, allowing the network to learn the complex mapping relationship between the input roof image and its actual structural parameters (the number of eaves lines, the number of ridge lines, the number of waist lines, and the angles between them). By optimizing the internal network parameters, a fully trained and stable network model is ultimately obtained, referred to as the first residual capsule neural network. This first residual capsule neural network has the following functions: when inputting a new, unknown roof image (i.e., a remote sensing image of the roof to be measured), it can automatically analyze and output the number of eaves lines, the number of ridge lines, the number of waist lines, and the angles between each of the eaves lines, the ridge lines, and the waist lines.

[0031] Next, data collection for the actual object to be identified is performed. When the roof type of a specific building needs to be identified, the camera fixed to the bottom of the drone is used to take aerial photos of the target building again to obtain a high-resolution image of its roof.

[0032] Finally, the acquired remote sensing image of the roof to be tested is fed into the trained first residual capsule neural network (RCN). Based on the learned features and patterns, the network performs computational analysis and outputs the number of eaves lines, ridge lines, and waist lines of the roof to be tested, as well as the angles between each of these lines. Once these specific structural parameters are obtained, they are then compared with a pre-built library of roof types. This library contains standard structural parameter ranges (including the standard number of eaves lines, ridge lines, waist lines, and the standard angle ranges between each) for various known standard roof types (including flat roofs, gable roofs, hip roofs, and mansard roofs). Using a specific matching algorithm, the output parameters are compared with the standard parameters of each type in the library to identify the known roof type that best matches the structural characteristics of the roof to be tested. Ultimately, the specific type of the roof to be tested is determined and obtained.

[0033] Furthermore, the initial residual capsule neural network includes a standard convolutional layer, a capsule layer, and a feedback iterative layer.

[0034] For example, the standard convolutional layer is used to identify the first image to the M The partial images of the eaves, ridges, and waistlines in the image cannot be identified, but the angles between the eaves, ridges, and waistlines, that is, the spatial position relationship, cannot be identified. The capsule layer is used to identify the first image to the M The number of eaves, ridges, and waistlines in the image, as well as the angles between each of them, are calculated. The feedback iteration layer is used to iteratively optimize the parameters of the standard convolutional and capsule layers to achieve better recognition results.

[0035] Furthermore, the standard convolutional layer uses a convolutional neural network algorithm to extract the first image to the M The local images of the eaves, ridges and waistlines in the image; the convolutional neural network algorithm is used to extract the local images from the first image to the M The parameters used in the process of local images of eaves, ridges, and waistlines in the image are recorded as standard convolution layer parameters; the standard convolution layer parameters include convolution kernel, convolution kernel width, number of input channels, number of output channels, and bias top; the number of standard convolution layer parameters is obtained H , relabel the standard convolution layer parameters as the first standard convolution layer parameters to the H Standard convolutional layer parameters.

[0036] Furthermore, the capsule layer is based on the first image to the M The local images of the eaves, ridges and waistlines in the image are extracted using a dynamic routing algorithm from the first image to the second image. M The number of eaves lines, ridge lines, waist lines and the angles between the eaves lines, ridge lines and waist lines in the image are obtained from the first recognition data set to the second recognition data set. M Identify the data group; M The local images of the eaves, ridges and waistlines in the image are extracted using a dynamic routing algorithm from the first image to the second image. M The number of eaves lines, ridge lines, waist lines and the angles between the eaves lines, ridge lines and waist lines in the image are obtained from the first recognition data set to the second recognition data set. M The parameters used in the process of identifying the data group are recorded as capsule layer parameters; the capsule layer parameters include the number of input capsules, the number of output capsules, the output capsule dimension, the input capsule dimension, and the coupling coefficient; the number of capsule layer parameters is obtained L , relabel the capsule layer parameters as the first capsule layer parameters to the L Capsule layer parameters.

[0037] Furthermore, the first identification data group to the M The identification data set is represented as: ; in, Indicates the Identify data groups; Indicates the the number of eave lines in the image; Indicates the The number of roof lines in the image; Indicates the The number of waist lines in the image; is a data set, indicating the The set of angles between the eaves, ridges, and waistlines in the image; The value range is 1 to M Integer variable between .

[0038] Furthermore, the feedback iterative layer uses a particle swarm optimization algorithm to optimize and calculate the standard convolution layer parameters and the capsule layer parameters to obtain the optimal standard convolution layer parameters and the optimal capsule layer parameters; the optimal standard convolution layer parameters and the optimal capsule layer parameters are respectively used as the standard convolution layer parameters and the capsule layer parameters of the first residual capsule neural network.

[0039] Furthermore, the particle swarm optimization algorithm is used to optimize and calculate the standard convolution layer parameters and the capsule layer parameters to obtain the optimal standard convolution layer parameters and the optimal capsule layer parameters. The method is as follows: Construct an optimization objective function; the optimization objective function is expressed as: ; in, represents the optimization objective function; represents the standard convolutional layer parameters; Represents the capsule layer parameters; Indicates the The actual number of eaves lines on the roof to which the image corresponds; Indicates the The actual number of ridge lines on the roof that the image corresponds to; Indicates the The actual number of waistlines on the roof to which the image corresponds; Indicates the The actual angles between the eaves line, ridge line and waist line of the roof corresponding to the image are Identify the first The sum of the absolute values ​​of the angle differences between the eaves line, ridge line, and waist line in the image, in radians.

[0040] Construct optimization constraints.

[0041] With the optimization constraints as constraints and the goal of minimizing the value of the optimization objective function, the particle swarm optimization algorithm is used to calculate the optimal values ​​of the standard convolutional layer parameters and the optimal values ​​of the capsule layer parameters, which are recorded as the optimal standard convolutional layer parameters and the optimal capsule layer parameters respectively.

[0042] For example, Further explanation is given. The number of eaves lines and the number of ridge lines corresponding to the first image is 4. Renumber the 4 eaves lines as actual eaves line 1, actual eaves line 2, actual eaves line 3, and actual eaves line 4. Record one ridge line as actual ridge line 1. The actual angle between actual eaves line 1 and actual ridge line 1 is 0 radians. The actual angle between actual eaves line 2 and actual ridge line 1 is The actual angle between the actual eaves line No. 3 and the actual ridge line No. 1 is 0 radians. The actual angle between the actual eaves line No. 4 and the actual ridge line No. 1 is radians. Since in the first recognition data set, the number of eaves lines in the first image is 4 and the number of ridge lines in the first image is 1. Renumber the 4 eaves lines as eaves line 1, eaves line 2, eaves line 3, and eaves line 4. Record one ridge line as ridge line 1. The actual angle between eaves line 1 and ridge line 1 is 0.143 radians. The actual angle between eaves line 2 and ridge line 1 is 1.52 radians. The actual angle between actual eaves line 3 and actual ridge line 1 is 0.156 radians. The actual angle between actual eaves line 4 and actual ridge line 1 is 1.53 radians. Therefore, the calculation results are =0.391 radians. The number of eaves lines and the number of ridge lines corresponding to the second image is 4. Renumber the 4 eaves lines as actual eaves line 1, actual eaves line 2, actual eaves line 3, and actual eaves line 4. Record one ridge line as actual ridge line 1. The actual angle between actual eaves line 1 and actual ridge line 1 is 0 radians. The actual angle between actual eaves line 2 and actual ridge line 1 is The actual angle between the actual eaves line No. 3 and the actual ridge line No. 1 is 0 radians. The actual angle between the actual eaves line No. 4 and the actual ridge line No. 1 is radians. Since in the first recognition data set, the number of eaves lines in the second image is 3 and the number of ridge lines in the first image is 1. Renumber the 3 eaves lines as eaves line No. 1, eaves line No. 2, and eaves line No. 3. Record one ridge line as ridge line No. 1. The actual angle between eaves line No. 1 and ridge line No. 1 is 0.133 radians. The actual angle between eaves line No. 2 and ridge line No. 1 is 1.48 radians. The actual angle between actual eaves line No. 3 and actual ridge line No. 1 is 0.126 radians. Since there is no eave No. 4 in the second image, in order to make the calculation result more robust, a virtual eave No. 4 is constructed, and the angle between the virtually constructed eaves line No. 4 and ridge line No. 1 is artificially determined to be radians. Therefore, the calculation is The actual angle between the eaves line, ridge line and waist line ranges from 0 to The angle between the eaves line, ridge line and waist line ranges from 0 to radian.

[0043] Furthermore, the optimization constraints are: ; in, Indicates the t Standard convolutional layer parameters; Indicates the preset t The lower limit of standard convolutional layer parameters; Indicates the preset t Upper limit of standard convolutional layer parameters; Indicates the c Capsule layer parameters; Indicates the preset c The lower limit of capsule layer parameters; Indicates the preset c Upper limit of capsule layer parameters; t The value range is 1 to H Integer between ; c The value range is 1 to L Integer between .

[0044] Furthermore, the method of matching the number of eaves lines, the number of ridge lines, the number of waist lines, and the angles between the eaves lines, the ridge lines, and the waist lines of the roof to be tested with the pre-obtained roof type library to obtain the type of the roof to be tested is as follows: The roof type library includes preset roof types and the preset number of eaves lines, the preset number of ridge lines, the preset number of waist lines corresponding to the preset roof types, and the preset angles between the eaves lines, the ridge lines, and the waist lines; the preset roof types include the first type to the G Type; Type 1 to TypeG The preset number of eaves lines, the preset number of ridge lines, the preset number of waist lines, and the preset angles between the eaves lines, ridge lines, and waist lines corresponding to the type are recorded as the first type characteristic parameters to the G Type characteristic parameters.

[0045] Calculate the number of eaves lines, ridge lines, waist lines, and the angles between the eaves lines, ridge lines, and waist lines of the roof to be tested, and the first type of characteristic parameters to the G weighted Euclidean distance of type characteristic parameters; the roof type corresponding to the minimum value of the weighted Euclidean distance is recorded as the roof type to be tested; in the calculation process of the weighted Euclidean distance, a preset first weight is assigned to the number of ridge lines, a preset second weight is assigned to the number of eaves lines, a preset third weight is assigned to the number of waist lines, and a preset fourth weight is assigned to the angles between the eaves lines, the ridge lines, and the waist lines; the first weight is greater than the second weight, the second weight is greater than the third weight, the third weight is greater than the fourth weight, and the sum of the first weight, the second weight, the third weight, and the fourth weight is 1.

[0046] For example, in the process of roof type identification, the number of ridge lines explains the roof type to a greater extent than the number of eaves lines, the number of eaves lines explains the roof type to a greater extent than the number of waist lines, and the number of waist lines explains the roof type to a greater extent than the angles between the eaves lines, ridge lines, and waist lines.

[0047] Finally, it should be noted that although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments, or make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A remote sensing recognition method for building roof types using residual capsule networks, characterized in that: The method comprises the following steps: S1, obtain sample roof remote sensing images through the camera fixed on the bottom of the drone to obtain a roof sample library; the number of roof remote sensing images in the roof sample library is M , respectively recorded as the first image to the M image; S2, obtain the first image to the M The actual number of eaves lines, actual number of ridge lines, actual number of waist lines and actual angles between eaves lines, ridge lines and waist lines of the roof corresponding to the image are obtained. M Data group; the first image to the M Image, first data set to M The data set is input into a pre-constructed initial residual capsule neural network, and the parameters of the initial residual capsule neural network are trained to obtain a first residual capsule neural network; the input of the first residual capsule neural network is a remote sensing image of the roof to be measured, and the output is the number of eaves lines, the number of ridge lines, the number of waist lines of the roof to be measured, and the angles between the eaves lines, the ridge lines, and the waist lines; S3, obtains remote sensing images of the roof to be measured through a camera fixed on the bottom of the drone; S4, inputting the remote sensing image of the roof to be measured into the first residual capsule neural network, and outputting the number of eaves lines, the number of ridge lines, the number of waist lines, and the angles between the eaves lines, the ridge lines, and the waist lines of the roof to be measured; matching the number of eaves lines, the number of ridge lines, the number of waist lines, and the angles between the eaves lines, the ridge lines, and the waist lines of the roof to be measured with a pre-obtained roof type library to obtain the roof type to be measured.

2. The method for remote sensing identification of building roof types using a residual capsule network according to claim 1, wherein: The initial residual capsule neural network includes a standard convolution layer, a capsule layer, and a feedback iteration layer.

3. The method for remote sensing identification of building roof types using a residual capsule network according to claim 2, wherein: The standard convolutional layer uses a convolutional neural network algorithm to extract the first image to the M The local images of the eaves, ridges and waistlines in the image; the convolutional neural network algorithm is used to extract the local images from the first image to the M The parameters used in the process of local images of the eaves, ridges, and waistlines in the image are recorded as standard convolution layer parameters; the standard convolution layer parameters include convolution kernel, convolution kernel width, number of input channels, number of output channels, and bias top; Get the number of standard convolutional layer parameters H , relabel the standard convolution layer parameters as the first standard convolution layer parameters to the H Standard convolutional layer parameters.

4. The method for remote sensing identification of building roof types using a residual capsule network according to claim 3, wherein: The capsule layer is based on the first image to the M The local images of the eaves, ridges and waistlines in the image are extracted using a dynamic routing algorithm from the first image to the second image. M The number of eaves lines, ridge lines, waist lines and the angles between the eaves lines, ridge lines and waist lines in the image are obtained from the first recognition data set to the second recognition data set. M Identify the data group; M The local images of the eaves, ridges and waistlines in the image are extracted using a dynamic routing algorithm from the first image to the second image. M The number of eaves lines, ridge lines, waist lines and the angles between the eaves lines, ridge lines and waist lines in the image are obtained from the first recognition data set to the second recognition data set. M The parameters used in the process of identifying the data set are recorded as capsule layer parameters; the capsule layer parameters include the number of input capsules, the number of output capsules, the output capsule dimension, the input capsule dimension, and the coupling coefficient; Get the number of capsule layer parameters L , relabel the capsule layer parameters as the first capsule layer parameters to the L Capsule layer parameters.

5. The method for remote sensing identification of building roof types using a residual capsule network according to claim 4, wherein: The first to the second identification data sets M The identification data set is represented as: ； in, Indicates the Identify data groups; Indicates the the number of eave lines in the image; Indicates the The number of roof lines in the image; Indicates the The number of waist lines in the image; is a data set, indicating the The set of angles between the eaves, ridges, and waistlines in the image; The value range is 1 to M Integer variable between .

6. The method for remote sensing identification of building roof types using a residual capsule network according to claim 5, wherein: The feedback iterative layer uses a particle swarm optimization algorithm to optimize and calculate the standard convolution layer parameters and the capsule layer parameters to obtain the optimal standard convolution layer parameters and the optimal capsule layer parameters; the optimal standard convolution layer parameters and the optimal capsule layer parameters are respectively used as the standard convolution layer parameters and the capsule layer parameters of the first residual capsule neural network.

7. The method for remote sensing identification of building roof types using a residual capsule network according to claim 6, wherein: The method of optimizing and calculating the standard convolution layer parameters and the capsule layer parameters by using the particle swarm optimization algorithm to obtain the optimal standard convolution layer parameters and the optimal capsule layer parameters is as follows: Construct an optimization objective function; the optimization objective function is expressed as: ； in, represents the optimization objective function; represents the standard convolutional layer parameters; Represents the capsule layer parameters; Indicates the The actual number of eaves lines on the roof to which the image corresponds; Indicates the The actual number of ridge lines on the roof that the image corresponds to; Indicates the The actual number of waistlines on the roof to which the image corresponds; Indicates the The actual angles between the eaves line, ridge line and waist line of the roof corresponding to the image are Identify the first The sum of the absolute values ​​of the angle differences between the eaves line, ridge line, and waist line in the image, in radians; Construct optimization constraints; With the optimization constraints as constraints and the goal of minimizing the value of the optimization objective function, the particle swarm optimization algorithm is used to calculate the optimal values ​​of the standard convolutional layer parameters and the optimal values ​​of the capsule layer parameters, which are recorded as the optimal standard convolutional layer parameters and the optimal capsule layer parameters respectively.

8. The method for remote sensing identification of building roof types using a residual capsule network according to claim 7, wherein: The optimization constraints are: ； in, Indicates the t Standard convolutional layer parameters; Indicates the preset t The lower limit of standard convolutional layer parameters; Indicates the preset t Upper limit of standard convolutional layer parameters; Indicates the c Capsule layer parameters; Indicates the preset c The lower limit of capsule layer parameters; Indicates the preset c Upper limit of capsule layer parameters; t The value range is 1 to H Integer between ; c The value range is 1 to L Integer between .

9. The method for remote sensing identification of building roof types using a residual capsule network according to claim 8, wherein: The method of matching the number of eaves lines, ridge lines, waist lines, and the angles between the eaves lines, ridge lines, and waist lines of the roof to be tested with the pre-obtained roof type library to obtain the type of the roof to be tested is as follows: The roof type library includes preset roof types and the preset number of eaves lines, the preset number of ridge lines, the preset number of waist lines corresponding to the preset roof types, and the preset angles between the eaves lines, the ridge lines, and the waist lines; the preset roof types include the first type to the G Type; Type 1 to Type G The preset number of eaves lines, the preset number of ridge lines, the preset number of waist lines, and the preset angles between the eaves lines, ridge lines, and waist lines corresponding to the type are recorded as the first type characteristic parameters to the G Type characteristic parameters; Calculate the number of eaves lines, ridge lines, waist lines, and the angles between the eaves lines, ridge lines, and waist lines of the roof to be tested, and the first type of characteristic parameters to the G weighted Euclidean distance of type characteristic parameters; the roof type corresponding to the minimum value of the weighted Euclidean distance is recorded as the roof type to be tested; in the calculation process of the weighted Euclidean distance, a preset first weight is assigned to the number of ridge lines, a preset second weight is assigned to the number of eaves lines, a preset third weight is assigned to the number of waist lines, and a preset fourth weight is assigned to the angles between the eaves lines, the ridge lines, and the waist lines; the first weight is greater than the second weight, the second weight is greater than the third weight, the third weight is greater than the fourth weight, and the sum of the first weight, the second weight, the third weight, and the fourth weight is 1.

Images