Remote sensing image scene classification method and device, equipment and storage medium

Abstract

The application discloses a remote sensing image scene classification method, device and equipment and a storage medium. The application obtains a remote sensing image to be classified, and then performs scene classification on the remote sensing image to be classified through a preset remote sensing image scene classification model to obtain a category corresponding to the remote sensing image to be classified. The preset remote sensing image scene classification model is obtained through dense connection convolution network training. The application performs scene classification on the remote sensing image to be classified through the preset remote sensing image scene classification model, and the preset remote sensing image scene classification model is obtained through dense connection convolution network training. Compared with the existing remote sensing scene classification through a deep learning network, the application can perform scene classification through the preset remote sensing scene classification model, can accurately classify remote sensing image scenes, and can effectively reduce the calculation amount and parameter amount of the preset remote sensing scene classification model backbone network, thereby improving the operation efficiency of the network.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, and particularly relates to a remote sensing image scene classification method and device, equipment and a storage medium. BACKGROUND

[0002] With the continuous improvement of remote sensing technology, the spatial and temporal resolution, revisit rate and accuracy of satellites are steadily improved, and the functions of remote sensing satellites are becoming more powerful. The traditional method is to study remote sensing scene classification through a deep learning network, but for the problems of large intra-class difference and high inter-class similarity of remote sensing images, the above method is not accurate enough for extracting key information of remote sensing images. At the same time, with the increase of network layers, the network calculation amount increases, and the network becomes more difficult to train. Therefore, how to accurately classify remote sensing image scenes while improving the efficiency of scene classification has become a problem to be solved.

[0003] The above content is only used to assist in understanding the technical solutions of the present application and does not represent the acknowledgement of the above content as prior art. SUMMARY

[0004] The main purpose of the present application is to provide a remote sensing image scene classification method, device, equipment and storage medium, which aims to solve the technical problem of how to accurately classify remote sensing image scenes while improving the efficiency of scene classification.

[0005] To achieve the above purpose, the present application provides a remote sensing image scene classification method, which comprises the following steps:

[0006] acquiring a remote sensing image to be classified;

[0007] classifying the remote sensing image to be classified through a preset remote sensing image scene classification model to obtain the category corresponding to the remote sensing image to be classified, wherein the preset remote sensing image scene classification model is obtained based on a densely connected convolutional network.

[0008] Optionally, before the step of acquiring the remote sensing image to be classified, the method further comprises:

[0009] training an initial remote sensing image scene classification model to obtain a preset remote sensing image scene classification model, wherein the initial remote sensing image scene classification model comprises a first weighted pooling layer, a first bottleneck layer, a second weighted pooling layer, a Ghost bottleneck layer, a second bottleneck layer, a third bottleneck layer and a third weighted pooling layer.

[0010] Optionally, the step of training the initial remote sensing image scene classification model to obtain the preset remote sensing image scene classification model specifically comprises:

[0011] acquiring a remote sensing image training set;

[0012] inputting the remote sensing image training set into the first weighted pooling layer for weighted processing to obtain a first weighted feature map;

[0013] inputting the first weighted feature map into the first bottleneck layer for convolution to obtain a first convolutional feature map;

[0014] inputting the first convolutional feature map into the second weighted pooling layer for weighted processing to obtain a second weighted feature map;

[0015] inputting the second weighted feature map into the Ghost bottleneck layer for convolution to obtain a second convolutional feature map;

[0016] inputting the second convolutional feature map into the second bottleneck layer, the third bottleneck layer and the third weighted pooling layer in sequence, and obtaining a current iteration number corresponding to the initial remote sensing image scene classification model;

[0017] when the current iteration number is greater than a preset iteration number, obtaining a preset remote sensing image scene classification model.

[0018] Optionally, the step of inputting the remote sensing image training set into the first weighted pooling layer for weighted processing to obtain a first weighted feature map specifically comprises:

[0019] inputting the remote sensing image training set into the first weighted pooling layer to obtain a weight corresponding to each remote sensing image in the remote sensing image training set;

[0020] accumulating the weight to obtain a first weighted feature map.

[0021] Optionally, the step of inputting the remote sensing image training set into the first weighted pooling layer to obtain a weight corresponding to each remote sensing image in the remote sensing image training set specifically comprises:

[0022] inputting the remote sensing image training set into the first weighted pooling layer and determining a weight corresponding to each remote sensing image in the remote sensing image training set through a first preset formula, the first preset formula being:

[0023]

[0024] wherein, a i is an activation of an i-th element of a remote sensing image, w i is a weight of the activation of the i-th element of the remote sensing image, R is a pooling neighborhood size of the weighted pooling layer, |R|=k 2 , and k is a height or width of a pooling region of the weighted pooling layer.

[0025] Optionally, the step of inputting the second weighted feature map into the Ghost bottleneck layer for convolution to obtain a second convolutional feature map specifically comprises:

[0026] inputting the second weighted feature map into the Ghost bottleneck layer to generate a first feature map according to a preset number of feature maps in the second weighted feature map;

[0027] performing linear operation on the first feature map to generate a second feature map;

[0028] performing convolution according to the first feature map and the second feature map to obtain a second convolutional feature map.

[0029] Optionally, the Ghost bottleneck layer comprises a first Ghost module and a second Ghost module.

[0030] The step of inputting the second weighted feature map into the Ghost bottleneck layer to generate a first feature map according to a preset number of feature maps in the second weighted feature map specifically comprises:

[0031] inputting the second weighted feature map into the first Ghost module to increase the number of channels corresponding to each feature map in the second weighted feature map as an expansion layer;

[0032] reducing the increased number of channels through the second Ghost module to make the reduced number of channels consistent with the initial number of channels in the second weighted feature map, and generating a first feature map according to a preset number of feature maps in the second weighted feature map.

[0033] In addition, to achieve the above object, the present application further provides a remote sensing image scene classification device, which comprises:

[0034] a remote sensing image acquisition module, configured to acquire a remote sensing image to be classified;

[0035] a remote sensing image scene classification module, configured to perform scene classification on the remote sensing image to be classified through a preset remote sensing image scene classification model to obtain a category corresponding to the remote sensing image to be classified, wherein the preset remote sensing image scene classification model is obtained based on a densely connected convolutional network.

[0036] In addition, to achieve the above object, the present application further provides a remote sensing image scene classification device, which comprises a memory, a processor and a remote sensing image scene classification program stored in the memory and executable on the processor, wherein the remote sensing image scene classification program is configured to implement the steps of the remote sensing image scene classification method as described above.

[0037] In addition, to achieve the above object, the application further provides a storage medium, wherein the storage medium stores a remote sensing image scene classification program, and the remote sensing image scene classification program is executed by a processor to implement the steps of the remote sensing image scene classification method.

[0038] The application obtains a remote sensing image to be classified, and then classifies the remote sensing image to be classified by using a preset remote sensing image scene classification model to obtain a category corresponding to the remote sensing image to be classified, wherein the preset remote sensing image scene classification model is obtained based on a dense connection convolution network. The application classifies the remote sensing image to be classified by using the preset remote sensing image scene classification model, and the preset remote sensing image scene classification model is obtained based on the dense connection convolution network. Compared with the existing remote sensing scene classification by using a deep learning network, the application can classify the remote sensing image by using the preset remote sensing scene classification model, can accurately classify the remote sensing image scene, and can effectively reduce the calculation amount and the parameter amount of the preset remote sensing scene classification model backbone network, thereby improving the operation efficiency of the network. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 FIG. 1 is a structural schematic diagram of a remote sensing image scene classification device of a hardware running environment related to an embodiment scheme of the application;

[0040] Figure 2 FIG. 2 is a flowchart of a remote sensing image scene classification method according to a first embodiment of the application;

[0041] Figure 3 FIG. 3 is a flowchart of a remote sensing image scene classification method according to a second embodiment of the application;

[0042] Figure 4 FIG. 4 is a down-sampling process diagram of a weighted pooling layer of the application;

[0043] Figure 5 FIG. 5 is an implementation process diagram of a Ghost bottleneck layer of the application;

[0044] Figure 6 FIG. 6 is a structural block diagram of a remote sensing image scene classification device according to a first embodiment of the application.

[0045] The implementation, functional features and advantages of the application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0046] It should be understood that the specific embodiments described herein are intended to explain the application, but not to limit the application.

[0047] Reference Figure 1 , Figure 1A remote sensing image scene classification device structure schematic diagram of a hardware running environment involved in an embodiment of the present application is shown.

[0048] As shown in Figure 1 , the remote sensing image scene classification device can include a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to realize the connection and communication between the components. The user interface 1003 can include a display, an input unit such as a keyboard, and can also include a standard wired interface, a wireless interface. The network interface 1004 can optionally include a standard wired interface, a wireless interface (such as a wireless fidelity (Wi-Fi) interface). The memory 1005 can be a high-speed random access memory (RAM), and can also be a stable non-volatile memory (NVM), such as a disk memory. The memory 1005 can also be a storage device independent of the aforementioned processor 1001.

[0049] Those skilled in the art can understand that Figure 1 the structure shown in the above description does not constitute a limitation on the remote sensing image scene classification device, and can include more or fewer components than the diagram, or combine certain components, or different component arrangements.

[0050] As shown in Figure 1 , the memory 1005 as a storage medium can include an operating system, a network communication module, a user interface module, and a remote sensing image scene classification program.

[0051] In the remote sensing image scene classification device shown in Figure 1 , the network interface 1004 is mainly used for data communication with a network server; the user interface 1003 is mainly used for data interaction with a user; the processor 1001 and the memory 1005 in the remote sensing image scene classification device of the present application can be arranged in the remote sensing image scene classification device, and the remote sensing image scene classification device calls the remote sensing image scene classification program stored in the memory 1005 through the processor 1001, and executes the remote sensing image scene classification method provided by the embodiment of the present application.

[0052] Based on the above remote sensing image scene classification device, an embodiment of the present application provides a remote sensing image scene classification method, which is described with reference to Figure 2 , Figure 2A flowchart of a first embodiment of the remote sensing image scene classification method of the present application is shown in FIG. 1.

[0053] In this embodiment, the remote sensing image scene classification method comprises the following steps:

[0054] Step S10: obtaining a remote sensing image to be classified;

[0055] It should be noted that the execution subject of this embodiment can be a computing service device with data processing, network communication and program running functions, such as a mobile phone, a tablet computer, a personal computer, etc., or an electronic device or a remote sensing image scene classification device capable of realizing the above functions. In the following, the remote sensing image scene classification device is taken as an example to describe this embodiment and the following embodiments.

[0056] It can be understood that the remote sensing image to be classified refers to a remote sensing image that needs to be classified.

[0057] Step S20: classifying the remote sensing image to be classified by a preset remote sensing image scene classification model to obtain a category corresponding to the remote sensing image to be classified, wherein the preset remote sensing image scene classification model is obtained based on a densely connected convolutional network.

[0058] It should be understood that the preset remote sensing image scene classification model is a model that can be used to classify the scene of a remote sensing image. The preset remote sensing image scene classification model in this embodiment is obtained based on a densely connected convolutional network, i.e., a Densenet121 network.

[0059] In a specific implementation, this embodiment can input the remote sensing image to be classified into the preset remote sensing scene classification model to classify the scene and obtain the corresponding category.

[0060] This embodiment obtains a remote sensing image to be classified, and then classifies the scene of the remote sensing image to be classified by a preset remote sensing image scene classification model to obtain a category corresponding to the remote sensing image to be classified, wherein the preset remote sensing image scene classification model is obtained based on a densely connected convolutional network. Compared with the existing remote sensing scene classification by a deep learning network, this embodiment can classify the scene by the preset remote sensing scene classification model, can accurately classify the scene of a remote sensing image, and can effectively reduce the calculation amount and parameter amount of the backbone network of the preset remote sensing scene classification model, thereby improving the operation efficiency of the network.

[0061] Reference Figure 3 , Figure 3A flowchart of a second embodiment of the remote sensing image scene classification method of the present application is shown.

[0062] Based on the first embodiment, in the present embodiment, before the step S10, further comprising:

[0063] Step S00: training an initial remote sensing image scene classification model to obtain a preset remote sensing image scene classification model, the initial remote sensing image scene classification model comprising: a first weighted pooling layer, a first bottleneck layer, a second weighted pooling layer, a Ghost bottleneck layer, a second bottleneck layer, a third bottleneck layer, and a third weighted pooling layer.

[0064] It should be understood that the first weighted pooling layer and the second weighted pooling layer have the same structure. The pooling layer is the key to reduce the size of the activation map in the convolutional neural network. The pooling layer reduces the network computing demand while also realizing spatial invariance and increasing the receptive field of subsequent convolution. Commonly used pooling layers mainly include two types: maximum pooling and average pooling. The maximum pooling only selects the highest single activation in the region, while the average pooling reduces the influence of all activations in the region. The first weighted pooling layer and the second weighted pooling layer (SoftPool) are between the two, because all activations in the region contribute to the final output, and higher activations are more dominant than lower activations. This balances the influence of average pool and maximum pool, while taking advantage of the beneficial properties of both. SoftPool is based on the Softmax weighting method to preserve the basic properties of the input. By weighting Softmax, the pooling region is also less susceptible to local kernel activation vanishing, solving a common problem of average pooling; unlike maximum pooling, SoftPool is differentiable, getting the gradient of each input during backpropagation, improving neural connectivity during training.

[0065] It should be noted that the first weighted pooling layer and the second weighted pooling layer in the present embodiment can be weighted, and the first bottleneck layer, the Ghost bottleneck layer, the second bottleneck layer, and the third bottleneck layer can be convolutional.

[0066] In a specific implementation, the embodiment applies the Ghost bottleneck layer to the dense connection structure with reference to the idea of the Densenet121 network, and improves the Densenet121 network by using the SoftPool. First, the SoftPool is used to replace the average pooling and the maximum pooling in the Densenet121; second, the Ghost bottleneck layer is used to replace the Dense Block2 bottleneck layer in the original model, and the first bottleneck layer, the second bottleneck layer and the third bottleneck layer are the DenseBlock2 bottleneck layer in the original model, so that the initial remote sensing image scene classification model in the embodiment is obtained. Therefore, the initial remote sensing image scene classification model in the embodiment includes: a first weighted pooling layer, a first bottleneck layer, a second weighted pooling layer, a Ghost bottleneck layer, a second bottleneck layer, a third bottleneck layer and a third weighted pooling layer.

[0067] Further, in order to accurately obtain the preset remote sensing image scene classification model, in the embodiment, the step S00 includes:

[0068] Step S001: obtaining a remote sensing image training set;

[0069] It should be understood that the verification experiment in the embodiment can be performed on the UCMerced_Land Use (UC Merced) public remote sensing image dataset, and the model robustness test can be performed on four remote sensing image datasets, including the Aerial Image Dataset (AID), the WHU-RS19 Date Set (WHU-RS19), the RSSCN7 Date Set (RSSCN7) and the SIRI-WHU Date Set (SIRI-WHU).

[0070] Step S002: inputting the remote sensing image training set into the first weighted pooling layer for weighted processing to obtain a first weighted feature map;

[0071] Further, in order to accurately obtain the first weighted feature map, in the embodiment, the step S002 includes: inputting the remote sensing image training set into the first weighted pooling layer to obtain the weight corresponding to each remote sensing image in the remote sensing image training set; and accumulating the weight to obtain a first weighted remote sensing image training set.

[0072] It can be understood that, with reference to Figure 4 , Figure 4 is a down-sampling process diagram of the weighted pooling layer of the present application. As Figure 4As shown, taking any region of a remote sensing image in the remote sensing image training set as an example, the numbers on the graph represent the elements of the remote sensing image, that is, 2, 1, 2, 3, 5, 3…, and 1e-3 represents the weight, the weight corresponding to the element "3" in the third row and the third column is "1e-3", the weight corresponding to the element "4" in the third row and the fourth column is "2e-3", and so on. The present embodiment will not be described in detail.

[0073] Further, in order to accurately obtain the weight, in the present embodiment, the step of inputting the remote sensing image training set into the first weighted pooling layer to obtain the weight corresponding to each remote sensing image in the remote sensing image training set specifically comprises:

[0074] The remote sensing image training set is input into the first weighted pooling layer, and the weight corresponding to each remote sensing image in the remote sensing image training set is determined by a first preset formula, the first preset formula being:

[0075]

[0076] In the formula, a i is the activation of the i-th element of the remote sensing image, w i is the weight of the activation of the i-th element of the remote sensing image, R is the pooling neighborhood size of the weighted pooling layer, |R|=k 2 , and k is the height or width of the pooling region of the weighted pooling layer.

[0077] It can be understood that the weight corresponding to each element in the remote sensing image can be obtained by the first preset formula.

[0078] It should be understood that the output value of the pooling field can also be obtained by a second preset formula in the present embodiment, the second preset formula being:

[0079]

[0080] In the formula, is the output value of the pooling neighborhood R.

[0081] Step S003: inputting the first weighted feature map into the first bottleneck layer for convolution to obtain a first convolutional feature map;

[0082] It should be understood that the first bottleneck layer in the present embodiment is the Dense Block2 bottleneck layer in the Densenet121 network, and the specific convolutional processing process can refer to the prior art, and the present embodiment will not be described in detail.

[0083] Step S004: inputting the first convolutional feature map into the second weighted pooling layer for weighted processing to obtain a second weighted feature map;

[0084] It can be understood that the weighting processing process of step S004 in the embodiment is the same as that in step S002, and the embodiment will not be described in detail.

[0085] Step S005: inputting the second weighted feature map into the Ghost bottleneck layer for convolution to obtain a second convolutional feature map;

[0086] Further, in the embodiment, the step S005 includes: inputting the second weighted feature map into the Ghost bottleneck layer to generate a first feature map according to a preset number of feature maps in the second weighted feature map; performing linear operation on the first feature map to generate a second feature map; and performing convolution according to the first feature map and the second feature map to obtain a second convolutional feature map.

[0087] It can be understood that referring to Figure 5 , Figure 5 is the implementation process diagram of the Ghost bottleneck layer of the application. The Ghost bottleneck layer is divided into two steps. In the first step, a part of feature maps is generated using a small number of convolution kernels. In the second step, another part of feature maps is generated by performing linear operation on the feature maps generated in the first step. Finally, the two groups of feature maps are spliced together to obtain complete feature maps.

[0088] In specific implementation, the operation of generating n feature maps by the ordinary convolution layer can be represented as: Y=X*f+b, wherein X∈R c×h×w represents the input, and h and w represent the height and width of the input feature; Y∈R h′×w′×n represents the output feature map of n channels, and h' and w' represent the height and width of the output feature respectively; f∈R c×k×k×n represents the convolution kernel of c×n size k×k; and b represents the bias term.

[0089] And the Ghost convolution only needs to perform a small number of conventional convolution operations, and generates feature maps by linear operation to reduce feature redundancy, so as to achieve the goal of reducing the amount of calculation and the amount of parameters. The Ghost convolution operation can be represented as: Y'=X*f', wherein Y'∈R h′×w′×m represents m(m≤n) feature maps generated by linear convolution, and the bias term b is omitted here; f'∈R c×k×k×m represents the convolution kernel of c×m size k×k.

[0090] Under the condition of outputting the same feature map size and the same number of channels, the FLOPs ratio of the conventional convolution and the Ghost bottleneck layer is: wherein the size of k×k is approximately equal to the size of d×d, and s<<c, so formula (7) can be simplified as: According to the above formula, it can be inferred that using the Ghost bottleneck layer for convolution operation can reduce the amount of calculation in the convolution process to a certain extent.

[0091] Further, in the embodiment, the step of inputting the second weighted feature map into the Ghost bottleneck layer to generate a first feature map according to a preset number of feature maps in the second weighted feature map specifically includes: inputting the second weighted feature map into the first Ghost module to increase the number of channels corresponding to each feature map in the second weighted feature map as an expansion layer; reducing the increased number of channels through the second Ghost module to make the reduced number of channels consistent with the initial number of channels in the second weighted feature map, and generating a first feature map according to a preset number of feature maps in the second weighted feature map.

[0092] It should be understood that the Ghost bottleneck layer in the embodiment includes a first Ghost module and a second Ghost module; the first Ghost module is used as an expansion layer to increase the number of channels, and the second Ghost module is used to reduce the number of channels to match the number of channels of the input connection.

[0093] Step S006: sequentially inputting the second convolutional feature map into the second bottleneck layer, the third bottleneck layer and the third weighted pooling layer, and obtaining a current iteration number corresponding to the initial remote sensing image scene classification model;

[0094] It should be understood that the processing process of the second bottleneck layer, the third bottleneck layer and the third weighted pooling layer can refer to the above steps, and the embodiment will not be described in detail.

[0095] Step S007: obtaining a preset remote sensing image scene classification model when the current iteration number is greater than a preset iteration number.

[0096] It should be understood that when the iteration number of the initial remote sensing image scene classification model, i.e. the current iteration number, is greater than the preset iteration number, it means that the initial remote sensing image scene classification model is trained, i.e. the preset remote sensing image scene classification model can be obtained. The specific preset iteration number can be set to 200 times, and can also be set to other numerical values, which are not limited in the embodiment.

[0097] The embodiment trains an initial remote sensing image scene classification model to obtain a preset remote sensing image scene classification model. The initial remote sensing image scene classification model includes a first weighted pooling layer, a first bottleneck layer, a second weighted pooling layer, a Ghost bottleneck layer, a second bottleneck layer, a third bottleneck layer, and a third weighted pooling layer. The embodiment uses a Densenet121 network as a backbone network, and replaces part of dense connection blocks in the backbone network with the Ghost bottleneck layer, thereby effectively reducing the calculation amount and parameter amount of the backbone network, and improving the operation efficiency of the network.

[0098] Referring to Figure 6 , Figure 6 FIG. 1 is a structural block diagram of a remote sensing image scene classification device according to an embodiment of the present application.

[0099] As shown in Figure 6 , the remote sensing image scene classification device according to the embodiment of the present application includes:

[0100] A remote sensing image acquisition module 10 is configured to acquire a remote sensing image to be classified.

[0101] A remote sensing image scene classification module 20 is configured to perform scene classification on the remote sensing image to be classified by using a preset remote sensing image scene classification model, and obtain a category corresponding to the remote sensing image to be classified. The preset remote sensing image scene classification model is obtained based on a dense connection convolutional network.

[0102] The embodiment acquires a remote sensing image to be classified, and then performs scene classification on the remote sensing image to be classified by using a preset remote sensing image scene classification model, and obtains a category corresponding to the remote sensing image to be classified. The preset remote sensing image scene classification model is obtained based on a dense connection convolutional network. Compared with the existing remote sensing scene classification by using a deep learning network, the above-mentioned method of the embodiment can perform scene classification by using a preset remote sensing scene classification model, can accurately classify remote sensing image scenes, and can effectively reduce the calculation amount and parameter amount of a backbone network of the preset remote sensing scene classification model, thereby improving the operation efficiency of the network.

[0103] It should be noted that the above-described workflow is merely illustrative and does not limit the scope of protection of the present application. In actual applications, a person skilled in the art can select part or all of the above-described workflow to achieve the purpose of the embodiment, and the selection is not limited herein.

[0104] In addition, technical details not described in detail in the embodiment can be referred to the remote sensing image scene classification method provided by any embodiment of the present application, and will not be described herein.

[0105] Based on the first embodiment of the remote sensing image scene classification device, the second embodiment of the remote sensing image scene classification device is proposed.

[0106] In the embodiment, the remote sensing image scene classification device further comprises a model training module 00, which is configured to train an initial remote sensing image scene classification model to obtain a preset remote sensing image scene classification model. The initial remote sensing image scene classification model comprises a first weighted pooling layer, a first bottleneck layer, a second weighted pooling layer, a Ghost bottleneck layer, a second bottleneck layer, a third bottleneck layer, and a third weighted pooling layer.

[0107] Further, the model training module 00 is further configured to obtain a remote sensing image training set, input the remote sensing image training set into the first weighted pooling layer for weighted processing to obtain a first weighted feature map, input the first weighted feature map into the first bottleneck layer for convolution to obtain a first convolution feature map, input the first convolution feature map into the second weighted pooling layer for weighted processing to obtain a second weighted feature map, input the second weighted feature map into the Ghost bottleneck layer for convolution to obtain a second convolution feature map, input the second convolution feature map into the second bottleneck layer, the third bottleneck layer, and the third weighted pooling layer in sequence, and obtain a current iteration number corresponding to the initial remote sensing image scene classification model. When the current iteration number is greater than a preset iteration number, a preset remote sensing image scene classification model is obtained.

[0108] Further, the model training module 00 is further configured to input the remote sensing image training set into the first weighted pooling layer to obtain a weight corresponding to each remote sensing image in the remote sensing image training set, and accumulate the weight to obtain a first weighted feature map.

[0109] Further, the model training module 00 is further configured to input the remote sensing image training set into the first weighted pooling layer, and determine the weight corresponding to each remote sensing image in the remote sensing image training set by a first preset formula, the first preset formula being:

[0110]

[0111] wherein, a i is the activation of the i-th element of the remote sensing image, w i is the weight of the activation of the i-th element of the remote sensing image, R is the pooling neighborhood size of the weighted pooling layer, |R|=k 2 , k is the height or width of the pooling region of the weighted pooling layer.

[0112] Further, the model training module 00 is further configured to input the second weighted feature map into the Ghost bottleneck layer, to generate a first feature map according to a preset number of remote sensing images in the second weighted feature map, perform linear operation on the first feature map to generate a second feature map, and perform convolution according to the first feature map and the second feature map to obtain a second convolutional feature map.

[0113] Further, the Ghost bottleneck layer comprises a first Ghost module and a second Ghost module, and the model training module 00 is further configured to input the second weighted feature map into the first Ghost module to increase the number of channels corresponding to each feature map in the second weighted feature map as an expansion layer, reduce the increased number of channels through the second Ghost module, so that the reduced number of channels is consistent with the initial number of channels in the second weighted feature map, and generate a first feature map according to a preset number of feature maps in the second weighted feature map.

[0114] Other embodiments or specific implementations of the remote sensing image scene classification device of the present application can refer to the above-mentioned method embodiments, which will not be described here.

[0115] In addition, the embodiments of the present application also propose a storage medium, wherein the storage medium stores a remote sensing image scene classification program, and the remote sensing image scene classification program is executed by a processor to realize the steps of the remote sensing image scene classification method as described above.

[0116] It should be noted that in this paper, the term "includes", "contains" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or system including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or includes elements inherent to such process, method, article or system. Without more limitations, the element defined by the statement "includes a" does not exclude the presence of another identical element in the process, method, article or system including the element.

[0117] The above-mentioned serial numbers of the embodiments of the present application are only for description, and do not represent the advantages and disadvantages of the embodiments.

[0118] Through the above description of the embodiments, those skilled in the art can clearly understand that the above-mentioned example method can be realized by means of software and a necessary general hardware platform, and of course, it can also be realized by hardware, but in many cases, the former is a better embodiment. Based on such understanding, the technical solutions of the present application can be embodied in the form of a software product, which is stored in a storage medium (such as a read-only memory / random access memory, a magnetic disk, or an optical disk) and includes a number of instructions for causing an end device (which can be a mobile phone, a computer, a server, an air conditioner, or a network device) to execute the method described in each embodiment of the present application.

[0119] The above is only the preferred embodiment of the present application, and does not limit the patent scope of the present application, and any equivalent structure or equivalent flow transformation made by using the content of the specification and drawings of the present application, or directly or indirectly applied to other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A method for classifying scenes in remote sensing images, characterized in that, The remote sensing image scene classification method includes the following steps: The initial remote sensing image scene classification model is trained to obtain a preset remote sensing image scene classification model. The initial remote sensing image scene classification model includes: a first weighted pooling layer, a first bottleneck layer, a second weighted pooling layer, a Ghost bottleneck layer, a second bottleneck layer, a third bottleneck layer, and a third weighted pooling layer. Acquire the remote sensing image to be classified; The remote sensing image to be classified is classified by a preset remote sensing image scene classification model to obtain the category corresponding to the remote sensing image to be classified. The preset remote sensing image scene classification model is trained based on a densely connected convolutional network, and the densely connected convolutional network is a Densenet121 network.

2. The remote sensing image scene classification method as described in claim 1, characterized in that, The step of training the initial remote sensing image scene classification model to obtain the preset remote sensing image scene classification model specifically includes: Obtain a training set of remote sensing images; The remote sensing image training set is input into the first weighted pooling layer for weighted processing to obtain the first weighted feature map; The first weighted feature map is input into the first bottleneck layer for convolution to obtain the first convolutional feature map. The feature map after the first convolution is input into the second weighted pooling layer for weighted processing to obtain the second weighted feature map; The second weighted feature map is input into the Ghost bottleneck layer for convolution to obtain the second convolutional feature map. The feature map after the second convolution is sequentially input into the second bottleneck layer, the third bottleneck layer, and the third weighted pooling layer, and the current iteration number corresponding to the initial remote sensing image scene classification model is obtained; When the current iteration number is greater than the preset iteration number, a preset remote sensing image scene classification model is obtained.

3. The remote sensing image scene classification method as described in claim 2, characterized in that, The step of inputting the remote sensing image training set into the first weighted pooling layer for weighted processing to obtain the first weighted feature map specifically includes: The remote sensing image training set is input into the first weighted pooling layer to obtain the weights corresponding to each remote sensing image in the remote sensing image training set. The weights are accumulated to obtain the first weighted feature map.

4. The remote sensing image scene classification method as described in claim 3, characterized in that, The step of inputting the remote sensing image training set into the first weighted pooling layer to obtain the weights corresponding to each remote sensing image in the remote sensing image training set specifically includes: The remote sensing image training set is input into the first weighted pooling layer, and the weights corresponding to each remote sensing image in the training set are determined by a first preset formula, which is: In the formula, For the activation of the i-th element in the remote sensing image, Let R be the weight activated by the i-th element in the remote sensing image, and let R be the pooling neighborhood size of the weighted pooling layer. , The height or width of the pooling region in the weighted pooling layer.

5. The remote sensing image scene classification method as described in claim 2, characterized in that, The step of inputting the second weighted feature map into the Ghost bottleneck layer for convolution to obtain the second convolutional feature map specifically includes: The second weighted feature map is input into the Ghost bottleneck layer to generate a first feature map based on a preset number of feature maps in the second weighted feature map. Perform linear operations on the first feature map to generate a second feature map; Convolution is performed on the first feature map and the second feature map to obtain the second convolutional feature map.

6. The remote sensing image scene classification method as described in claim 5, characterized in that, The Ghost bottleneck layer includes: a first Ghost module and a second Ghost module; The step of inputting the second weighted feature map into the Ghost bottleneck layer to generate a first feature map based on a preset number of feature maps in the second weighted feature map specifically includes: The second weighted feature map is input into the first Ghost module to serve as an extension layer to increase the number of channels corresponding to each feature map in the second weighted feature map; The second Ghost module reduces the number of channels after the increase so that the reduced number of channels is consistent with the initial number of channels in the second weighted feature map, and generates a first feature map based on a preset number of feature maps in the second weighted feature map.

7. A remote sensing image scene classification device, characterized in that, The remote sensing image scene classification device includes: The model training module is used to train the initial remote sensing image scene classification model to obtain a preset remote sensing image scene classification model. The initial remote sensing image scene classification model includes: a first weighted pooling layer, a first bottleneck layer, a second weighted pooling layer, a Ghost bottleneck layer, a second bottleneck layer, a third bottleneck layer, and a third weighted pooling layer. The remote sensing image acquisition module is used to acquire remote sensing images to be classified. The remote sensing image scene classification module is used to classify the remote sensing image to be classified by means of a preset remote sensing image scene classification model, and to obtain the category corresponding to the remote sensing image to be classified. The preset remote sensing image scene classification model is trained based on a densely connected convolutional network, and the densely connected convolutional network is a Densenet121 network.

8. A remote sensing image scene classification device, characterized in that, The device includes: a memory, a processor, and a remote sensing image scene classification program stored in the memory and executable on the processor, the remote sensing image scene classification program being configured to implement the steps of the remote sensing image scene classification method as described in any one of claims 1 to 6.

9. A storage medium, characterized in that, The storage medium stores a remote sensing image scene classification program, which, when executed by a processor, implements the steps of the remote sensing image scene classification method as described in any one of claims 1 to 6.

Images