Lightweight remote sensing cloud detection method for edge deployment and related device
By introducing lightweight residual module, spatial pyramid pooling module and structural reparameterization module in the YOLOv3 framework, a preset remote sensing cloud detection model is generated, which solves the accuracy and complexity of background complex cloud detection on edge devices, and realizes the realization of the realization of the following methods. Efficient cloud detection effect.
Patent Information
- Application Number
- CN202411174469.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-26
- Publication Date
- 2024-11-08
AI Technical Summary
The prior art is difficult to accurately detect remote sensing clouds with complex backgrounds, and the existing algorithms are highly complex and difficult to apply on edge devices.
The YOLOv3 framework is adopted, and the preset remote sensing cloud detection model is generated by combining the lightweight residual module, the spatial pyramid pooling module, the structural reparameterization module and the spanning connection structure. The model is optimized and trained through the Adam optimizer and the learning rate linear attenuation strategy. Implement cloud detection.
It reduces model parameters and calculation complexity, improves computing efficiency and detection accuracy, and can accurately detect remote sensing clouds with complex backgrounds on edge devices.
Smart Images

Figure CN118918493A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of target detection technology, and in particular to a lightweight remote sensing cloud detection method and related devices for edge deployment. Background Art
[0002] Remote sensing technology is a means of obtaining information about the earth's surface through long-distance sensing technologies such as satellites, aircraft, and drones. With the advancement of remote sensing technology and satellite technology, remote sensing images have become crucial in studying information about the earth's surface, especially the prominent position of satellite remote sensing images in this field. According to data from the International Satellite Cloud Climate Program, the global average annual cloud coverage is about 68%. Among them, the cloud cover on land is about 55%, while the cloud cover on the ocean is about 72%, highlighting the large proportion of cloud pixels in remote sensing images. Therefore, cloud detection has become a crucial research task in satellite remote sensing image processing. Although observing cloud cover is of great significance to the study of the earth's water cycle and climate change, when using satellite images for ground object analysis, clouds may become an interference signal and constitute an obstacle to the accurate acquisition of surface information. Therefore, accurately detecting cloud-covered areas and removing clouds are crucial to obtaining reliable surface information.
[0003] In cloud detection, when clouds and other surface objects with similar reflective properties coexist in remote sensing images, the underlying background of the clouds becomes complex, and it may be difficult to distinguish them. In addition, the existing detection algorithms are relatively complex, which limits the application of the methods to edge devices and makes it difficult to produce practical significance. Summary of the invention
[0004] The present application provides a lightweight remote sensing cloud detection method and related devices for edge deployment, which are used to solve the technical problems that the existing technology is difficult to accurately detect remote sensing clouds with complex backgrounds, and the algorithm is complex and difficult to be applied to edge devices.
[0005] In view of this, the first aspect of the present application provides a lightweight remote sensing cloud detection method for edge deployment, including:
[0006] Acquire cloud images through remote sensing technology to obtain target cloud images;
[0007] Based on the YOLOv3 framework, a preset remote sensing cloud detection model is generated according to the lightweight residual module, spatial pyramid pooling module, structural reparameterization module and spanning connection structure;
[0008] The target cloud image is input into the preset remote sensing cloud detection model to perform cloud detection and obtain a detection mark cloud.
[0009] Preferably, the method of generating a preset remote sensing cloud detection model based on the YOLOv3 framework according to a lightweight residual module, a spatial pyramid pooling module, a structural reparameterization module and a spanning connection structure includes:
[0010] In the YOLOv3 framework, a lightweight residual module, a spatial pyramid pooling module, a structural reparameterization module, and a spanning connection structure are configured to build an initial remote sensing cloud detection model.
[0011] The Adam optimizer and the learning rate linear decay strategy are used to perform detection optimization training on the initial remote sensing cloud detection model according to the preset remote sensing cloud training data set to obtain the preset remote sensing cloud detection model.
[0012] Preferably, the step of inputting the target cloud image into the preset remote sensing cloud detection model to perform cloud detection to obtain a detection marker cloud comprises:
[0013] Performing initial convolution processing on the target cloud image through the input layer of the preset remote sensing cloud detection model to obtain an initial convolution feature map;
[0014] Using the lightweight residual module to perform feature extraction analysis on the initial convolution feature map to obtain a residual feature map, the lightweight residual module includes a depth-separable convolution and an SE attention unit;
[0015] Extracting feature information of different scales from the residual feature map through the spatial pyramid pooling module to generate a multi-scale feature map;
[0016] Performing convolution fusion processing on the multi-scale feature map according to the structure reparameterization module to generate a fused feature map;
[0017] Based on the spanning connection structure, multiple upsampling-based detection analyses are performed on the fused feature map to output a detection tag cloud.
[0018] Preferably, extracting feature information of different scales from the residual feature map by the spatial pyramid pooling module to generate a multi-scale feature map includes:
[0019] Extracting feature information of different scales from the residual feature map through a 1×1 convolution layer and a plurality of preset hole convolution layers with different hole rates in the spatial pyramid pooling module to obtain sub-feature maps of multiple scales;
[0020] After the multiple scale sub-feature maps are spliced and fused, a 1×1 convolutional layer is used to perform channel compression adjustment, and a multi-scale feature map is output.
[0021] The second aspect of the present application provides a lightweight remote sensing cloud detection device for edge deployment, including:
[0022] An image acquisition unit, used to acquire cloud images through remote sensing technology to obtain target cloud images;
[0023] A model generation unit is used to generate a preset remote sensing cloud detection model based on the YOLOv3 framework, according to a lightweight residual module, a spatial pyramid pooling module, a structural reparameterization module, and a spanning connection structure;
[0024] The cloud detection unit is used to input the target cloud image into the preset remote sensing cloud detection model to perform cloud detection and obtain a detection mark cloud.
[0025] Preferably, the model generating unit is specifically used for:
[0026] In the YOLOv3 framework, a lightweight residual module, a spatial pyramid pooling module, a structural reparameterization module, and a spanning connection structure are configured to build an initial remote sensing cloud detection model.
[0027] The Adam optimizer and the learning rate linear decay strategy are used to perform detection optimization training on the initial remote sensing cloud detection model according to the preset remote sensing cloud training data set to obtain the preset remote sensing cloud detection model.
[0028] Preferably, the cloud detection unit comprises:
[0029] An initial convolution subunit, used to perform initial convolution processing on the target cloud image through the input layer of the preset remote sensing cloud detection model to obtain an initial convolution feature map;
[0030] A residual analysis subunit, configured to perform feature extraction analysis on the initial convolution feature map using the lightweight residual module to obtain a residual feature map, wherein the lightweight residual module includes a depthwise separable convolution and an SE attention unit;
[0031] A multi-scale extraction subunit, used for extracting feature information of different scales from the residual feature map through the spatial pyramid pooling module to generate a multi-scale feature map;
[0032] A convolution fusion subunit, used for performing convolution fusion processing on the multi-scale feature map according to the structure re-parameterization module to generate a fused feature map;
[0033] The sampling detection subunit is used to perform multiple upsampling-based detection analyses on the fused feature map based on the spanning connection structure, and output a detection tag cloud.
[0034] Preferably, the multi-scale extraction subunit is specifically used for:
[0035] Extracting feature information of different scales from the residual feature map through a 1×1 convolution layer and a plurality of preset hole convolution layers with different hole rates in the spatial pyramid pooling module to obtain sub-feature maps of multiple scales;
[0036] After the multiple scale sub-feature maps are spliced and fused, a 1×1 convolutional layer is used to perform channel compression adjustment, and a multi-scale feature map is output.
[0037] The third aspect of the present application also provides a lightweight remote sensing cloud detection device for edge deployment, the device comprising a processor and a memory;
[0038] The memory is used to store program code and transmit the program code to the processor;
[0039] The processor is used to execute the lightweight remote sensing cloud detection method for edge deployment described in the first aspect according to the instructions in the program code.
[0040] The fourth aspect of the present application also provides a computer-readable storage medium, which is used to store program code, and the program code is used to execute the lightweight remote sensing cloud detection method for edge deployment described in the first aspect.
[0041] It can be seen from the above technical solutions that the embodiments of the present application have the following advantages:
[0042] In this application, a lightweight remote sensing cloud detection method for edge deployment is provided, including: acquiring cloud images through remote sensing technology to obtain target cloud images; based on the YOLOv3 framework, generating a preset remote sensing cloud detection model according to a lightweight residual module, a spatial pyramid pooling module, a structural reparameterization module and a spanning connection structure; inputting the target cloud image into the preset remote sensing cloud detection model for cloud detection to obtain a detection marker cloud.
[0043] The lightweight remote sensing cloud detection method for edge deployment provided by this application adds a lightweight residual module to the preset remote sensing cloud detection model, which can reduce model parameters and computational complexity and improve computing efficiency, and the structural reparameterization module is also a configuration for improving network computing speed; in addition, the introduction of the spatial pyramid pooling module can extract features of different scales, including spanning connection structures, and is also to enrich multi-scale feature information, thereby ensuring the accuracy of the detection results. This process can ensure detection accuracy while ensuring computing efficiency, and can adapt to more application scenarios, including edge devices. Therefore, the present application can solve the technical problems that the prior art is difficult to accurately detect remote sensing clouds with complex backgrounds, and the algorithm is complex and difficult to be applied to edge devices. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] Figure 1A schematic diagram of a lightweight remote sensing cloud detection method for edge deployment provided in an embodiment of the present application;
[0045] Figure 2 A schematic diagram of the structure of a lightweight remote sensing cloud detection device for edge deployment provided in an embodiment of the present application;
[0046] Figure 3 A schematic diagram of the overall structure of a preset remote sensing cloud detection model provided in an embodiment of the present application;
[0047] Figure 4 A schematic diagram of a network structure of a lightweight residual module provided in an embodiment of the present application;
[0048] Figure 5 A schematic diagram of the network structure of the spatial pyramid pooling module provided in an embodiment of the present application. DETAILED DESCRIPTION
[0049] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.
[0050] For easier understanding, see Figure 1 , an embodiment of a lightweight remote sensing cloud detection method for edge deployment provided by the present application includes:
[0051] Step 101: Acquire cloud images through remote sensing technology to obtain target cloud images.
[0052] The target cloud image may be an image acquired in real time, including clouds in the sky and some other objects, that is, the target cloud image may be a remote sensing cloud image with a complex background. This embodiment separates clouds from other ground objects by improving the algorithm and detects the target cloud.
[0053] Step 102: Based on the YOLOv3 framework, a preset remote sensing cloud detection model is generated according to a lightweight residual module, a spatial pyramid pooling module, a structural reparameterization module, and a spanning connection structure.
[0054] Further, step 102 includes:
[0055] In the YOLOv3 framework, a lightweight residual module, a spatial pyramid pooling module, a structural reparameterization module, and a spanning connection structure are configured to build an initial remote sensing cloud detection model.
[0056] The Adam optimizer and the learning rate linear decay strategy are used to perform detection optimization training on the initial remote sensing cloud detection model according to the preset remote sensing cloud training dataset to obtain the preset remote sensing cloud detection model.
[0057] See also Figure 3 , the preset remote sensing cloud detection model in this embodiment is based on the YOLOv3 framework. In this case, a lightweight residual module is added, which can significantly reduce model parameters and computational complexity while maintaining a high accuracy rate, thereby achieving faster reasoning. The introduction of the spatial pyramid pooling module can effectively extract feature information of different scales in the image, so that the network can learn features of various scales well; and the use of the UNet spanning link structure can further enrich the multi-scale feature information, thereby improving the accuracy of the model. The structural reparameterization module can increase the reasoning speed of the network. Only by improving the detection accuracy and detection speed of the model at the same time can the algorithm be deployed on the edge device to realize specific scenario applications. The improvement of the network model in this embodiment is mainly based on these two aspects.
[0058] It is understandable that the constructed initial remote sensing cloud detection model needs to be optimized and trained before it can be used for cloud detection tasks in real-time scenarios. In order to improve the performance of the model, this embodiment uses the Adam optimizer and the learning rate linear decay strategy to train the model, so that after the model detects the target cloud, it can mark it as white and the rest as black to achieve target detection.
[0059] Step 103: Input the target cloud image into a preset remote sensing cloud detection model to perform cloud detection and obtain a detection mark cloud.
[0060] Furthermore, step 103 includes:
[0061] Perform initial convolution processing on the target cloud image through the input layer of the preset remote sensing cloud detection model to obtain an initial convolution feature map;
[0062] A lightweight residual module is used to perform feature extraction and analysis on the initial convolution feature map to obtain a residual feature map. The lightweight residual module includes a depth-separable convolution and a SE attention unit.
[0063] The spatial pyramid pooling module is used to extract feature information of different scales in the residual feature map and generate a multi-scale feature map.
[0064] Perform convolution fusion processing on the multi-scale feature maps according to the structure reparameterization module to generate a fused feature map;
[0065] Based on the spanning connection structure, multiple upsampling-based detection analyses are performed on the fused feature map to output a detection tag cloud.
[0066] Furthermore, feature information of different scales is extracted from the residual feature map through the spatial pyramid pooling module to generate a multi-scale feature map, including:
[0067] Through the 1×1 convolution layer in the spatial pyramid pooling module and multiple preset dilated convolution layers with different dilation rates, feature information of different scales is extracted from the residual feature map to obtain sub-feature maps of multiple scales.
[0068] After the sub-feature maps of multiple scales are spliced and fused, a 1×1 convolutional layer is used to perform channel compression adjustment and output a multi-scale feature map.
[0069] It should be noted that after the target cloud image is input into the preset remote sensing cloud detection model, preliminary convolution processing can be performed to adjust the image size, etc.; in addition to inputting the network layer of the lightweight residual module, the obtained initial convolution feature map also needs to be used as input information across the connection and transmitted to the subsequent upsampling and splicing network layers to avoid network learning forgetting.
[0070] See also Figure 4 , the lightweight residual module of this embodiment first uses 1×1 grouped convolution to adjust the channels of the input feature map. It can divide the input feature map into multiple groups, perform independent convolution operations on each group, and then merge the outputs of all groups to obtain the overall output. Then, use 3×3 depth-separable convolution to extract the input features; and then use 1×1 grouped convolution kernel to reduce the dimension of the features to reduce the computational complexity of the model. Then, the result of the 1×1 grouped convolution is added to the input of the residual connection and input into the SE attention unit to strengthen the interaction between channels, thereby improving the generalization ability and accuracy of the model.
[0071] It is understandable that compared with the traditional convolution operation, the grouped convolution used in this embodiment can reduce the amount of calculation and parameters, reduce the complexity of the model, thereby accelerating the calculation and reducing the risk of overfitting. In addition, since grouped convolution may cause some features to be isolated in different groups, which cannot be effectively interactively fused, thereby limiting the network's expressive power, this embodiment will also rearrange the output results of the SE attention unit through channels, which can strengthen the information exchange between different channels, increase the interaction and information transfer between feature channels, and thus improve the efficiency of feature extraction.
[0072] See also Figure 5, the spatial pyramid pooling module of this embodiment divides the input feature map into channels through a multi-scale feature extraction unit, and divides the input feature map into 4 equal parts according to the channel dimension, and the dimension of each equal part is 1 / 4 of the number of channels of the input feature map; then three hole convolutions with hole rates of 1, 3, and 5 and a 1×1 convolution layer are used to extract feature information of different scales from the feature data after channel division, thereby generating receptive fields of different scales and obtaining contextual information of different scales. The extracted feature information is input into the feature fusion layer for splicing, and the channel is adjusted after passing through the 1×1 convolution layer to reduce the number of channels of the feature map and reduce the computational complexity. It can be understood that the specific number of equal parts to be divided needs to be determined according to the number of scale selection or the type of scale feature extraction layer, and it is not unique; and the hole rate can also be adjusted as needed, which is only used as an example and not limited here.
[0073] The structural reparameterization module in this embodiment needs to perform multiple convolution and padding operations. First, the 1×1 convolution needs to be converted into a 3×3 convolution to keep the convolution kernel size consistent. Padding is used here for zero padding. Secondly, the identity mapping is converted into a 3×3 convolution, that is, the separate 1×1 convolution of each channel in the depth-separable convolution is converted into a 3×3 convolution, and padding is used for zero padding. Next, the convolution and batch normalization layers are fused, that is, the two operators Conv2d and BN are fused into RepConv, thereby reducing the amount of computation and memory usage. The fusion of operators is performed after the network training is completed.
[0074] Specifically, let the input feature map be x, the convolution kernel size can be expressed as w, then the expression of the convolution layer can be expressed as:
[0075]
[0076] Among them, b is the bias of the convolution layer, and the subsequent input value batch normalization layer is specifically expressed as:
[0077]
[0078] in, is the mean of the BN layer data, is the variance of the BN layer data, Represents a very small constant, used to prevent the occurrence of 0; , They represent the scale factor and translation factor of the BN layer, which are used to adjust the data to the standard normal distribution. Substituting the operation of the convolution layer into the BN layer, we can get the following formula:
[0079]
[0080] After simplification, we can get:
[0081]
[0082] As can be seen from the formula, the first part on the right side of the formula and the second part after the plus sign are used as the convolution kernel and bias of the convolution layer respectively, so as to achieve the fusion of the convolution layer and the BN layer and obtain the fusion feature map. It can be replaced with a form with fewer parameters while maintaining higher performance.
[0083] The cross-connection structure of this embodiment is mainly based on the lightweight residual module to extract the initial features. In the upsampling splicing operation process after the input network, this is to achieve a similar effect of residual connection from a macro level to prevent the model from learning and forgetting. It can be understood that after training with the Adam optimizer and the linear attenuation strategy of the learning rate, the target cloud detected by the preset remote sensing cloud detection model will be marked as white, and the other parts will be marked as black, so after model detection, the detection mark cloud can be obtained.
[0084] The lightweight remote sensing cloud detection method for edge deployment provided in the embodiment of the present application adds a lightweight residual module to the preset remote sensing cloud detection model, which can reduce model parameters and computational complexity and improve computing efficiency, and the structural reparameterization module is also a configuration for improving network computing speed; in addition, the introduction of the spatial pyramid pooling module can extract features of different scales, including spanning connection structures, and is also to enrich multi-scale feature information, thereby ensuring the accuracy of the detection results. This process can ensure detection accuracy while ensuring computing efficiency, and can adapt to more application scenarios, including edge devices. Therefore, the embodiment of the present application can solve the technical problems that the prior art is difficult to accurately detect remote sensing clouds with complex backgrounds, and the algorithm is complex and difficult to be applied to edge devices.
[0085] For easier understanding, see Figure 2 , the present application provides an embodiment of a lightweight remote sensing cloud detection device for edge deployment, including:
[0086] The image acquisition unit 201 is used to acquire cloud images through remote sensing technology to obtain target cloud images;
[0087] The model generation unit 202 is used to generate a preset remote sensing cloud detection model based on the YOLOv3 framework, according to the lightweight residual module, the spatial pyramid pooling module, the structure re-parameterization module and the cross-connection structure;
[0088] The cloud detection unit 203 is used to input the target cloud image into a preset remote sensing cloud detection model to perform cloud detection and obtain a detection mark cloud.
[0089] Furthermore, the model generation unit 202 is specifically configured to:
[0090] In the YOLOv3 framework, a lightweight residual module, a spatial pyramid pooling module, a structural reparameterization module, and a spanning connection structure are configured to build an initial remote sensing cloud detection model.
[0091] The Adam optimizer and the learning rate linear decay strategy are used to perform detection optimization training on the initial remote sensing cloud detection model according to the preset remote sensing cloud training dataset to obtain the preset remote sensing cloud detection model.
[0092] Further, the cloud detection unit 203 includes:
[0093] The initial convolution subunit 2031 is used to perform initial convolution processing on the target cloud image through the input layer of the preset remote sensing cloud detection model to obtain an initial convolution feature map;
[0094] A residual analysis subunit 2032 is used to perform feature extraction analysis on the initial convolution feature map using a lightweight residual module to obtain a residual feature map, wherein the lightweight residual module includes a depthwise separable convolution and an SE attention unit;
[0095] A multi-scale extraction subunit 2033 is used to extract feature information of different scales from the residual feature map through a spatial pyramid pooling module to generate a multi-scale feature map;
[0096] The convolution fusion subunit 2034 is used to perform convolution fusion processing on the multi-scale feature map according to the structure re-parameterization module to generate a fused feature map;
[0097] The sampling detection subunit 2035 is used to perform multiple upsampling-based detection analyses on the fused feature map based on the spanning connection structure, and output a detection tag cloud.
[0098] Furthermore, the multi-scale extraction subunit 2033 is specifically used for:
[0099] Through the 1×1 convolution layer in the spatial pyramid pooling module and multiple preset dilated convolution layers with different dilation rates, feature information of different scales is extracted from the residual feature map to obtain sub-feature maps of multiple scales.
[0100] After the sub-feature maps of multiple scales are spliced and fused, a 1×1 convolutional layer is used to perform channel compression adjustment and output a multi-scale feature map.
[0101] The present application also provides a lightweight remote sensing cloud detection device for edge deployment, the device including a processor and a memory;
[0102] The memory is used to store the program code and transmit the program code to the processor;
[0103] The processor is used to execute the lightweight remote sensing cloud detection method for edge deployment in the above method embodiment according to the instructions in the program code.
[0104] The present application also provides a computer-readable storage medium, which is used to store program code, and the program code is used to execute the lightweight remote sensing cloud detection method for edge deployment in the above method embodiment.
[0105] In the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.
[0106] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.
[0107] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.
[0108] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including a number of instructions for executing all or part of the steps of the method described in each embodiment of the present application through a computer device (which can be a personal computer, server, or network device, etc.). The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (full name in English: Read-Only Memory, English abbreviation: ROM), random access memory (full name in English: Random Access Memory, English abbreviation: RAM), disk or optical disk and other media that can store program codes.
[0109] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.
Claims
1. Lightweight remote sensing cloud detection method for edge deployment, It is characterized in that include: Acquire cloud images through remote sensing technology to obtain target cloud images; Based on the YOLOv3 framework, a preset remote sensing cloud detection model is generated according to the lightweight residual module, spatial pyramid pooling module, structural reparameterization module and spanning connection structure; The target cloud image is input into the preset remote sensing cloud detection model to perform cloud detection and obtain a detection mark cloud.
2. According to claim 1, the lightweight remote sensing cloud detection method for edge deployment, It is characterized in that The preset remote sensing cloud detection model is generated based on the YOLOv3 framework according to the lightweight residual module, the spatial pyramid pooling module, the structure reparameterization module and the cross-connection structure, including: In the YOLOv3 framework, a lightweight residual module, a spatial pyramid pooling module, a structural reparameterization module, and a spanning connection structure are configured to build an initial remote sensing cloud detection model. The Adam optimizer and the learning rate linear decay strategy are used to perform detection optimization training on the initial remote sensing cloud detection model according to the preset remote sensing cloud training data set to obtain the preset remote sensing cloud detection model.
3. According to claim 1, the lightweight remote sensing cloud detection method for edge deployment, It is characterized in that The step of inputting the target cloud image into the preset remote sensing cloud detection model to perform cloud detection and obtain a detection mark cloud comprises: Performing initial convolution processing on the target cloud image through the input layer of the preset remote sensing cloud detection model to obtain an initial convolution feature map; Using the lightweight residual module to perform feature extraction analysis on the initial convolution feature map to obtain a residual feature map, the lightweight residual module includes a depth-separable convolution and an SE attention unit; Extracting feature information of different scales from the residual feature map through the spatial pyramid pooling module to generate a multi-scale feature map; Performing convolution fusion processing on the multi-scale feature map according to the structure reparameterization module to generate a fused feature map; Based on the spanning connection structure, multiple upsampling-based detection analyses are performed on the fused feature map to output a detection tag cloud.
4. According to claim 3, the lightweight remote sensing cloud detection method for edge deployment, It is characterized in that The extracting feature information of different scales from the residual feature map by the spatial pyramid pooling module to generate a multi-scale feature map includes: Extracting feature information of different scales from the residual feature map through a 1×1 convolution layer and a plurality of preset hole convolution layers with different hole rates in the spatial pyramid pooling module to obtain sub-feature maps of multiple scales; After the multiple scale sub-feature maps are spliced and fused, a 1×1 convolutional layer is used to perform channel compression adjustment, and a multi-scale feature map is output.
5. Lightweight remote sensing cloud detection device for edge deployment, It is characterized in that include: An image acquisition unit, used to acquire cloud images through remote sensing technology to obtain target cloud images; A model generation unit is used to generate a preset remote sensing cloud detection model based on the YOLOv3 framework, according to a lightweight residual module, a spatial pyramid pooling module, a structural reparameterization module, and a spanning connection structure; The cloud detection unit is used to input the target cloud image into the preset remote sensing cloud detection model to perform cloud detection and obtain a detection mark cloud.
6. The lightweight remote sensing cloud detection device for edge deployment according to claim 5, It is characterized in that The model generation unit is specifically used for: In the YOLOv3 framework, a lightweight residual module, a spatial pyramid pooling module, a structural reparameterization module, and a spanning connection structure are configured to build an initial remote sensing cloud detection model. The Adam optimizer and the learning rate linear decay strategy are used to perform detection optimization training on the initial remote sensing cloud detection model according to the preset remote sensing cloud training data set to obtain the preset remote sensing cloud detection model.
7. The lightweight remote sensing cloud detection device for edge deployment according to claim 5, It is characterized in that The cloud detection unit comprises: An initial convolution subunit, used to perform initial convolution processing on the target cloud image through the input layer of the preset remote sensing cloud detection model to obtain an initial convolution feature map; A residual analysis subunit, configured to perform feature extraction analysis on the initial convolution feature map using the lightweight residual module to obtain a residual feature map, wherein the lightweight residual module includes a depthwise separable convolution and an SE attention unit; A multi-scale extraction subunit, used for extracting feature information of different scales from the residual feature map through the spatial pyramid pooling module to generate a multi-scale feature map; A convolution fusion subunit, used for performing convolution fusion processing on the multi-scale feature map according to the structure re-parameterization module to generate a fused feature map; The sampling detection subunit is used to perform multiple upsampling-based detection analyses on the fused feature map based on the spanning connection structure, and output a detection tag cloud.
8. The lightweight remote sensing cloud detection device for edge deployment according to claim 7, It is characterized in that The multi-scale extraction subunit is specifically used for: Extracting feature information of different scales from the residual feature map through a 1×1 convolution layer and a plurality of preset hole convolution layers with different hole rates in the spatial pyramid pooling module to obtain sub-feature maps of multiple scales; After the multiple scale sub-feature maps are spliced and fused, a 1×1 convolutional layer is used to perform channel compression adjustment, and a multi-scale feature map is output.
9. Lightweight remote sensing cloud detection equipment for edge deployment, It is characterized in that The device comprises a processor and a memory; The memory is used to store program code and transmit the program code to the processor; The processor is used to execute the lightweight remote sensing cloud detection method for edge deployment according to any one of claims 1-4 according to the instructions in the program code.
10. A computer-readable storage medium, It is characterized in that The computer-readable storage medium is used to store program code, and the program code is used to execute the lightweight remote sensing cloud detection method for edge deployment as described in any one of claims 1-4.
Citation Information
Patent Citations
Remote sensing image segmentation method and device, storage medium and server
CN109801293A
Remote sensing image cloud and cloud shadow detection method based on double-attention neural network
CN114220001A
Embedded lightweight remote sensing target detection system
CN114494872A
Optical remote sensing image cloud and mist segmentation method based on spectrum guidance and depth attention
CN115713537A
Remote sensing image aircraft detection method based on multiple adaptive mechanisms
CN117853944A
