Wavelet transform-based remote sensing image target detection method and device

Abstract

The application provides a target detection method and device for a remote sensing image based on wavelet transform, comprising: acquiring a remote sensing image; performing multi-scale feature adaptive fusion processing on the remote sensing image to obtain an aggregated feature map; performing discrete wavelet transform on the aggregated feature map to obtain a low-frequency subband and at least one high-frequency subband, the low-frequency subband being used for indicating smooth background and structural information of the remote sensing image, and the at least one high-frequency subband being used for indicating edge features and texture features of the remote sensing image; performing inverse wavelet transform on the low-frequency subband and the at least one high-frequency subband to obtain a reconstructed feature map; performing residual connection on the reconstructed feature map and the aggregated feature map to obtain an enhanced feature map; and locating and identifying a target in the remote sensing image according to the enhanced feature map. In this way, the target detection efficiency in the remote sensing image can be improved.

Description

Technical Field

[0001] This application relates to the field of remote sensing image processing technology, and in particular to a target detection method and apparatus for remote sensing images based on wavelet transform. Background Technology

[0002] Remote sensing target detection is a fundamental task in the field of remote sensing image analysis, widely used in military reconnaissance, environmental monitoring, disaster assessment, and urban planning. With the rapid development of Earth observation technology, the amount of satellite and aerial imagery acquired has increased exponentially, making manual annotation and interpretation increasingly impractical. This trend has spurred an urgent need for automated detection systems. These systems require the ability to accurately locate and classify various targets in complex remote sensing images. In recent years, deep learning-based methods have become the mainstream approach for remote sensing target detection due to their superior feature representation capabilities and good scalability. Representative methods include two-stage detectors (such as Faster R-CNN) and one-stage detectors (such as the YOLO series and RetinaNet), which have achieved excellent results in large-scale and mesoscale target detection. However, in the task of detecting small remote sensing targets, existing methods still face significant challenges, mainly in the following aspects:

[0003] Currently, small targets in remote sensing images typically occupy only a few pixels, and their recognition relies heavily on subtle edge and texture information. However, because deep neural networks usually extract features hierarchically through continuous downsampling and convolution operations, these high-frequency details gradually weaken, leading to blurred target boundaries and reduced discriminability, ultimately affecting the detection accuracy of small-scale targets. Summary of the Invention

[0004] This application provides a target detection method and apparatus for remote sensing images based on wavelet transform, which can improve the accuracy of target detection in remote sensing images.

[0005] In a first aspect, embodiments of this application provide a target detection method for remote sensing images based on wavelet transform, the method comprising:

[0006] Acquire remote sensing images;

[0007] Multi-scale feature adaptive fusion processing is performed on remote sensing images to obtain aggregated feature maps;

[0008] Discrete wavelet transform is performed on the aggregated feature map to obtain a low-frequency sub-band and at least one high-frequency sub-band. The low-frequency sub-band is used to indicate the smooth background and structural information of the remote sensing image, and at least one high-frequency sub-band is used to indicate the edge features and texture features of the remote sensing image.

[0009] Perform inverse wavelet transform on the low-frequency sub-band and at least one high-frequency sub-band to obtain the reconstructed feature map;

[0010] Residual connections are performed between the reconstructed feature map and the aggregated feature map to obtain the enhanced feature map;

[0011] Based on the enhanced feature map, targets in remote sensing images are located and identified.

[0012] According to the above scheme, by fusing multi-scale features of remote sensing images, the complementary effects of the precise localization capability of shallow features and the semantic expression capability of deep features can be leveraged, improving the integrity and robustness of the overall feature representation, especially demonstrating superior detection performance in complex scenes with densely distributed small targets or large scale variations. Next, the aggregated feature map is divided into sub-bands of different frequencies. Then, based on these sub-bands (e.g., low-frequency sub-bands and at least one high-frequency sub-band), the aggregated feature map is processed by inverse wavelet transform to obtain a feature-enhanced reconstructed feature map, which preserves and highlights weak edge, corner, and texture features. Compared to traditional direct convolutional downsampling feature extraction methods, the scheme in this application significantly alleviates the performance degradation problem caused by feature blurring of small targets, improves the representation capability of fine-grained details, and enhances the discriminability of small targets. By performing residual connections between the reconstructed feature map and the enhanced feature map, an enhanced feature map is obtained. Since the features in both the reconstructed and aggregated feature maps have been enhanced, targets can be identified more accurately from the enhanced feature map, improving the accuracy of target detection.

[0013] In one possible implementation, the remote sensing image undergoes multi-scale feature adaptive fusion processing to obtain an aggregated feature map, including:

[0014] Multi-scale feature extraction is performed on remote sensing images to obtain features at multiple scales;

[0015] Adaptive weighted fusion of features at multiple scales is performed to obtain an aggregated feature map.

[0016] In one possible implementation, multiple scale features are adaptively weighted and fused to obtain an aggregated feature map, including:

[0017] The aggregated feature map is obtained by adaptively weighting and summing the features at multiple scales at the channel level.

[0018] In one possible implementation, the multiple scale features include scale features for indicating the spatial structure of the remote sensing image and scale features for indicating the semantic context of the remote sensing image.

[0019] In one possible implementation, an inverse wavelet transform is performed on the low-frequency subband and at least one high-frequency subband to obtain a reconstructed feature map, including:

[0020] Based on the low-frequency sub-band and at least one high-frequency sub-band, generate the respective weights of the low-frequency sub-band and at least one high-frequency sub-band;

[0021] Based on the respective weights of the low-frequency sub-band and the at least one high-frequency sub-band, the low-frequency sub-band and the at least one high-frequency sub-band are weighted and enhanced respectively to obtain the enhanced sub-band;

[0022] The enhanced subband is subjected to inverse wavelet transform to obtain the reconstructed feature map.

[0023] Thus, when calculating attention weights, not only global statistical characteristics are considered, but also statistical information of local regions is dynamically integrated. The enhancement strategy is adaptively adjusted according to the feature differences of different regions, which effectively makes up for the problem that traditional global modeling attention mechanisms are prone to losing local key features in highly heterogeneous remote sensing images. It can more accurately focus on fine-grained variation regions, such as the boundaries of occluded objects and the outlines of small targets against complex texture backgrounds, further improving the utilization rate of high-frequency details.

[0024] In one possible implementation, at least one high-frequency sub-band is used to indicate edge and texture features of the remotely sensed image in at least one of the following directions:

[0025] Horizontal direction, vertical direction, diagonal direction.

[0026] Secondly, embodiments of this application provide a target detection device for remote sensing images based on wavelet transform, the device comprising:

[0027] The acquisition module is used to acquire remote sensing images;

[0028] The fusion processing module is used to perform multi-scale feature fusion processing on remote sensing images to obtain aggregated feature maps;

[0029] The first transformation module is used to perform discrete wavelet transform on the aggregated feature map to obtain a low-frequency sub-band and at least one high-frequency sub-band. The low-frequency sub-band is used to indicate the smooth background and structural information of the remote sensing image, and at least one high-frequency sub-band is used to indicate the edge features and texture features of the remote sensing image.

[0030] The second transformation module is used to perform inverse wavelet transform on the aggregated feature map based on the low-frequency sub-band and at least one high-frequency sub-band to obtain the reconstructed feature map.

[0031] The connection module is used to perform residual connections between the reconstructed feature map and the aggregated feature map to obtain the enhanced feature map;

[0032] The recognition module is used to identify targets in remote sensing images based on the enhanced feature map.

[0033] In one possible implementation, the fusion processing module is used for:

[0034] Multi-scale feature extraction is performed on remote sensing images to obtain features at multiple scales;

[0035] Adaptive weighted fusion of features at multiple scales is performed to obtain an aggregated feature map.

[0036] In one possible implementation, the fusion processing module is used to perform channel-level adaptive weighted summation of features at multiple scales to obtain an aggregated feature map.

[0037] In one possible implementation, the multiple scale features include scale features for indicating the spatial structure of the remote sensing image and scale features for indicating the semantic context of the remote sensing image.

[0038] In one possible implementation, the second transformation module is used for:

[0039] Based on the low-frequency sub-band and the at least one high-frequency sub-band, generate the respective weights of the low-frequency sub-band and the at least one high-frequency sub-band;

[0040] Based on the respective weights of the low-frequency sub-band and the at least one high-frequency sub-band, the low-frequency sub-band and the at least one high-frequency sub-band are weighted and enhanced respectively to obtain the enhanced sub-band;

[0041] The enhanced subband is subjected to inverse wavelet transform to obtain the reconstructed feature map.

[0042] In one possible implementation, at least one high-frequency sub-band is used to indicate edge and texture features of the remotely sensed image in at least one of the following directions:

[0043] Horizontal direction, vertical direction, diagonal direction.

[0044] Thirdly, embodiments of this application provide a computing device, including a processor, a memory, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the methods provided in the first aspect or any possible implementation of the first aspect described above.

[0045] Fourthly, embodiments of this application provide a computer storage medium storing instructions that, when executed on a computer, cause the computer to perform the methods provided in the first aspect or any possible implementation thereof. Attached Figure Description

[0046] Figure 1 The illustration shows a flowchart of a target detection method for remote sensing images based on wavelet transform, provided in an embodiment of this application.

[0047] Figure 2A schematic diagram of the architecture of a target detection model provided in an embodiment of this application is shown;

[0048] Figure 3 This illustration shows a schematic diagram of a feature pyramid fusion process provided in an embodiment of this application;

[0049] Figure 4 A schematic diagram of the structure of a target detection device for remote sensing images based on wavelet transform provided in an embodiment of this application is shown.

[0050] Figure 5 A schematic diagram of the structure of a computing device provided in an embodiment of this application is shown. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions in the embodiments of this application will be described below with reference to the accompanying drawings.

[0052] In the description of the embodiments of this application, the words "exemplary," "for example," or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "exemplary," "for example," or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the words "exemplary," "for example," or "for instance" is intended to present the relevant concepts in a specific manner.

[0053] In the description of the embodiments in this application, the term "and / or" is merely a description of the association relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, B existing alone, and A and B existing simultaneously. Furthermore, unless otherwise stated, the term "multiple" means two or more. For example, multiple systems refer to two or more systems, and multiple screen terminals refer to two or more screen terminals.

[0054] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. The terms "comprising," "including," "having," and their variations all mean "including but not limited to," unless otherwise specifically emphasized.

[0055] Remote sensing target detection is a fundamental task in the field of remote sensing image analysis, widely used in military reconnaissance, environmental monitoring, disaster assessment, and urban planning. With the rapid development of Earth observation technology, the amount of satellite and aerial imagery acquired has increased exponentially, making manual annotation and interpretation increasingly impractical. This trend has spurred a growing demand for automated detection systems. These systems are required to accurately locate and classify various targets within complex remote sensing images. In recent years, deep learning-based methods have become the mainstream approach for remote sensing target detection due to their superior feature representation capabilities and good scalability. Representative methods include two-stage detectors (such as Faster R-CNN) and one-stage detectors (such as the YOLO series and RetinaNet), which have achieved excellent results in large-scale and mesoscale target detection.

[0056] However, small targets in remote sensing images typically occupy only a few pixels, and their recognition relies heavily on subtle edge and texture information. But because deep neural networks usually extract features hierarchically through continuous downsampling and convolution operations, these high-frequency details gradually weaken, leading to blurred target boundaries and reduced discriminability, ultimately affecting the detection accuracy of small-scale targets.

[0057] Based on this, this application provides a target detection method and apparatus for remote sensing images based on wavelet transform. In this scheme, by fusing multi-scale features of the remote sensing image, the complementary effects of the precise localization capability of shallow features and the semantic expression capability of deep features are leveraged, improving the integrity and robustness of the overall feature representation. This is particularly evident in complex scenes with densely distributed small targets or large scale variations, demonstrating superior detection performance. Next, the aggregated feature map is divided into sub-bands of different frequencies. Then, after processing the aggregated feature map with inverse wavelet transform based on different frequency sub-bands, such as low-frequency sub-bands and at least one high-frequency sub-band, a reconstructed feature map with enhanced features is obtained, enabling the preservation and highlighting of weak edge, corner, and texture features. Compared to traditional direct convolutional downsampling feature extraction methods, the scheme in this application significantly alleviates the performance degradation problem caused by feature blurring of small targets, improves the representation capability of fine-grained details, and enhances the discriminability of small targets. An enhanced feature map is obtained by performing residual connections between the reconstructed feature map and the enhanced feature map. Since the features in the reconstructed feature map and the aggregated feature map have been enhanced, the target can be identified more accurately from the enhanced feature map, thus improving the accuracy of target detection.

[0058] Figure 1This application provides a target detection method for remote sensing images based on wavelet transform. This wavelet transform-based target detection method for remote sensing images can be applied to devices or platforms with computing and storage capabilities, such as computing devices (e.g., servers) and computing device clusters (e.g., cloud computing platforms). Figure 1 As shown, the target detection method for remote sensing images based on wavelet transform provided in this application includes the following steps S101 to S106.

[0059] S101, acquire remote sensing images.

[0060] Satellites can generate remote sensing images, which can then be transmitted to devices or platforms with computing and storage capabilities. Remote sensing images contain a wealth of information, such as buildings, vehicles, and people—in other words, targets within the images. In fields like military reconnaissance, environmental monitoring, disaster assessment, and urban planning, information from remote sensing images is crucial for achieving desired objectives. Therefore, target detection within remote sensing images is necessary.

[0061] S102, perform multi-scale feature adaptive fusion processing on the remote sensing image to obtain an aggregated feature map.

[0062] After acquiring the remote sensing image, multi-scale feature extraction is performed to obtain multiple scale features. These multiple scale features include scale features that indicate the spatial structure of the remote sensing image and scale features that indicate the semantic context of the remote sensing image.

[0063] For example, such as Figure 2 As shown, feature maps (F3, F4, F5) of different scales are extracted from the backbone network such as ResNet50, which correspond to different semantic levels. F3 contains fine spatial structure information in the remote sensing image, while F4 and F5 contain rich semantic context in the remote sensing image.

[0064] To fully integrate information from different levels, features at multiple scales can be weighted and fused to obtain an aggregated feature map. For example, ... Figure 2 As shown, features at each scale are weighted and fused using channel-level learnable parameters (α3, α4, α5) to form a unified aggregated feature map F. agg This achieves adaptive weighted fusion of features at different scales. Specifically, the aggregated feature map F... agg It satisfies the following formula (1)

[0065] F agg =Conv 1×1 (Concat(α3F3,α4F4,α5F5)) (1)

[0066] Thus, by designing an adaptive weighted fusion mechanism, this embodiment of the application no longer uses simple concatenation or equal weighting when aggregating feature maps F3, F4, and F5 at different scales. Instead, it dynamically adjusts the fusion ratio according to the importance of the features. This mechanism fully leverages the complementary effects of the precise localization capability of shallow features and the semantic expression capability of deep features, improving the integrity and robustness of the overall feature representation. In particular, it exhibits superior detection performance in complex scenarios with densely distributed small targets or large scale variations.

[0067] S103, Perform discrete wavelet transform on the aggregated feature map to obtain a low-frequency sub-band and at least one high-frequency sub-band. The low-frequency sub-band is used to indicate the smooth background and structural information of the remote sensing image, and at least one high-frequency sub-band is used to indicate the edge features and texture features of the remote sensing image.

[0068] By applying Discrete Wavelet Transform (DWT) to the aggregated feature map, the frequency components of the feature map can be decomposed to obtain a low-frequency sub-band and at least one high-frequency sub-band. For example, the aggregated feature map can be subjected to DWT using methods such as Haar DWT or Daubechies DWT. Thus, by introducing DWT, this invention can divide the feature map into different frequency sub-bands, especially extracting and processing the high-frequency sub-band separately, thereby preserving and highlighting weak edge, corner, and texture features.

[0069] For example, such as Figure 2 As shown, for the aggregated feature map F agg Applying the first-order Haar discrete wavelet transform (DWT), it is decomposed into four subbands F, as shown in formula (2) below. Among them, the low-frequency subband F LL It can capture smooth background and structural information; the three high-frequency subbands are high-frequency subband F. lh High-frequency subband F hl High-frequency subband F hh Among them, the high-frequency subband F lh High-frequency subband F hl High-frequency subband F hh It can encode edge and texture features in different directions, F lh For horizontal, F hl For vertical, F hh It is a diagonal line.

[0070]

[0071] Where h represents high frequency and l represents low frequency. Low-frequency sub-band F ll Both the row and column directions are low frequency, and the high frequency subband F lhThe row direction is low frequency, the column direction is high frequency, and the high frequency sub-band F hl The row direction is high frequency, the column direction is low frequency, and the high frequency sub-band F hh Both row and column directions are high frequency.

[0072] S104, perform inverse wavelet transform on the low-frequency subband and at least one high-frequency subband to obtain the reconstructed feature map.

[0073] The features in the reconstructed feature map can be enhanced after performing inverse wavelet transform on the low-frequency subband and at least one high-frequency subband.

[0074] In some embodiments, S104 specifically includes the following process: First, weights are generated for the low-frequency sub-band and the at least one high-frequency sub-band, respectively. Next, weighted enhancement is performed on the low-frequency sub-band and the at least one high-frequency sub-band according to their respective weights, resulting in enhanced sub-bands. Then, inverse wavelet transform is performed on the enhanced sub-bands to obtain a reconstructed feature map.

[0075] Specifically, such as Figure 2 As shown, to enhance the representation of key details in the high-frequency subband, an attention mechanism using the Local-Global Adaptive (LGA) probability density function (PDF) is applied to each subband in the low-frequency subband and at least one high-frequency subband to obtain an enhanced subband. The enhanced subband can highlight key response regions and suppress redundant noise. The specific implementation process is as follows:

[0076] Calculate the global mean / variance, and the global mean μ of the subband. global The global variance of the subband satisfies the following formula (3). The following formula (4) must be satisfied:

[0077]

[0078] Where i, j = 0, 1, 2, 3, ... X i,j This represents the pixel value at coordinates (i, j) in the sub-band. H represents the height of the aggregated feature map, and W represents the width of the aggregated feature map.

[0079] The local mean / variance is calculated using 3x3 average pooling, with the subband's local mean μ being... local The local variance of the subband satisfies the following formula (5). The following formula (6) must be satisfied:

[0080] μ local =AvgPool 3×3(F) (5)

[0081]

[0082] Where F represents a subband.

[0083] The adaptive weights β, learned through experimentation, satisfy the following formula (7). This is combined with local variance. and global variance Statistical analysis, generating mixed variance Mixed variance The following formula (8) must be satisfied:

[0084] β = Sigmoid(DWConv) 3×3 (F)) (7)

[0085]

[0086] Based on mixture variance Calculate the PDF of the Gaussian distribution (normal distribution). The PDF of the Gaussian distribution satisfies the following formula (9):

[0087]

[0088] Where, μ local It is a local mean. For mixed variance, λ represents the learnable PDF width control parameter.

[0089] Based on the attention weight A of the Gaussian generated subband map (F). The attention weights of subband F satisfy the following formula (10):

[0090]

[0091] Where ε represents the numerical stability factor in the attention function, and γ represents the scaling factor in the attention function used to adjust the PDF intensity.

[0092] Based on the attention weight A of each subband F map (F), determining the weighted features of subband F. Wherein, the weighted features... Satisfy the following formula (11):

[0093]

[0094] To further enhance the key response regions while maintaining the stability of the original features, we employ a residual method to fuse weighted features, resulting in enhanced subbands. Here, k is an empirical coefficient, 0 ≤ k ≤ 1; for example, k can be 0.1. This is used to prevent excessive perturbation of the original features.

[0095] Among them, the fusion feature of subbands (i.e., the enhanced subband F) ★ ) satisfies the following formula (12):

[0096]

[0097] Finally, an inverse wavelet transform is performed on the enhanced subband to obtain the reconstructed feature map. The reconstructed feature map satisfies the following formula (13):

[0098]

[0099] Thus, this embodiment of the application considers both global statistical characteristics and dynamically integrates statistical information from local regions, adaptively adjusting the enhancement strategy according to the feature differences of different regions. This approach effectively compensates for the problem that traditional global modeling attention mechanisms are prone to losing key local features in highly heterogeneous remote sensing images, and can more accurately focus on regions with fine-grained variations, such as the boundaries of occluded objects and the contours of small targets against complex texture backgrounds, further improving the utilization rate of high-frequency details.

[0100] S105, perform residual connection on the reconstructed feature map and the aggregated feature map to obtain the enhanced feature map.

[0101] like Figure 2 As shown, the enhanced feature map is restored using inverse wavelet transform (IDWT) and then processed by 1×1 convolution (Conv... 1×1 This achieves residual connections with the original aggregated features, ensuring stable fusion of the enhanced features and the original features and avoiding information shift. The enhanced feature map satisfies the following formula (14):

[0102] F enhanced =F Z +Conv 1×1 (F agg (14)

[0103] S106, Based on the enhanced feature map, locate and identify targets in the remote sensing image.

[0104] Finally, by locating and identifying the enhanced feature map, targets in the remote sensing image are obtained. These targets include objects such as houses, vehicles, people, and plants. Based on the enhanced image, the pixel location of the target within the remote sensing image and the type of target can be more accurately determined.

[0105] In some embodiments, such as Figure 3 As shown, the enhanced feature map F enhanced It can serve as a new enhancement branch, fused with features at multiple scales through feature pyramids, to jointly complete subsequent target detection tasks and achieve accurate identification and localization of small targets.

[0106] According to the embodiments of this application, by fusing multi-scale features of remote sensing images, the complementary effects of the precise localization capability of shallow features and the semantic expression capability of deep features can be leveraged, improving the integrity and robustness of the overall feature representation, especially demonstrating superior detection performance in complex scenes with densely distributed small targets or large scale variations. Next, the aggregated feature map is divided into sub-bands of different frequencies. Then, after processing the aggregated feature map with inverse wavelet transform based on different frequency sub-bands, such as low-frequency sub-bands and at least one high-frequency sub-band, a reconstructed feature map with enhanced features is obtained, which can preserve and highlight weak edge, corner, and texture features. Compared with traditional direct convolutional downsampling feature extraction methods, the scheme of this application can significantly alleviate the problem of decreased detection performance for small targets due to feature blurring, improve the representation capability of fine-grained details, and improve the discriminability of small targets. An enhanced feature map is obtained by performing residual connections between the reconstructed feature map and the enhanced feature map. Since the features in the reconstructed feature map and the aggregated feature map have been enhanced, the target can be identified more accurately from the enhanced feature map, thus improving the accuracy of target detection.

[0107] Based on the same concept as the method embodiments of this application, this application also provides a target detection device for remote sensing images based on wavelet transform. The target detection device for remote sensing images based on wavelet transform includes several modules, each module executing a step in the target detection method for remote sensing images based on wavelet transform provided in this application. The division of modules is not limited here. Those skilled in the art will understand that in practical applications, the steps in the target detection method for remote sensing images based on wavelet transform provided in this application can be assigned to different modules as needed, that is, the internal structure of the device can be divided into different modules to complete all or part of the functions described above. The modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more modules can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the modules in the above device can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0108] Figure 4 This application provides a target detection device for remote sensing images based on wavelet transform. For example... Figure 4 As shown in the figure, the target detection device for remote sensing images based on wavelet transform provided in this application embodiment includes:

[0109] Acquisition module 401 is used to acquire remote sensing images;

[0110] The fusion processing module 402 is used to perform multi-scale feature fusion processing on remote sensing images to obtain an aggregated feature map;

[0111] The first transformation module 403 is used to perform discrete wavelet transform on the aggregated feature map to obtain a low-frequency sub-band and at least one high-frequency sub-band. The low-frequency sub-band is used to indicate the smooth background and structural information of the remote sensing image, and the at least one high-frequency sub-band is used to indicate the edge features and texture features of the remote sensing image.

[0112] The second transformation module 404 is used to perform inverse wavelet transform on the aggregated feature map based on the low-frequency sub-band and at least one high-frequency sub-band to obtain the reconstructed feature map.

[0113] The connection module 405 is used to perform residual connection on the reconstructed feature map and the aggregated feature map to obtain the enhanced feature map;

[0114] The recognition module 406 is used to identify targets in remote sensing images based on the enhanced feature map.

[0115] In one possible implementation, the fusion processing module is used for:

[0116] Multi-scale feature extraction is performed on remote sensing images to obtain features at multiple scales;

[0117] Adaptive weighted fusion of features at multiple scales is performed to obtain an aggregated feature map.

[0118] In one possible implementation, the fusion processing module is used to perform channel-level adaptive weighted summation of features at multiple scales to obtain an aggregated feature map.

[0119] In one possible implementation, the multiple scale features include scale features for indicating the spatial structure of the remote sensing image and scale features for indicating the semantic context of the remote sensing image.

[0120] In one possible implementation, the second transformation module is used for:

[0121] Based on the low-frequency sub-band and the at least one high-frequency sub-band, generate the respective weights of the low-frequency sub-band and the at least one high-frequency sub-band;

[0122] Based on the respective weights of the low-frequency sub-band and the at least one high-frequency sub-band, the low-frequency sub-band and the at least one high-frequency sub-band are weighted and enhanced respectively to obtain the enhanced sub-band;

[0123] The enhanced subband is subjected to inverse wavelet transform to obtain the reconstructed feature map.

[0124] In one possible implementation, at least one high-frequency sub-band is used to indicate edge and texture features of the remotely sensed image in at least one of the following directions:

[0125] Horizontal direction, vertical direction, diagonal direction.

[0126] According to the embodiments of this application, by fusing multi-scale features of remote sensing images, the complementary effects of the precise localization capability of shallow features and the semantic expression capability of deep features can be leveraged, improving the integrity and robustness of the overall feature representation, especially demonstrating superior detection performance in complex scenes with densely distributed small targets or large scale variations. Next, the aggregated feature map is divided into sub-bands of different frequencies. Then, after processing the sub-bands of different frequencies, such as low-frequency sub-bands and at least one high-frequency sub-band, inverse wavelet transform is performed to obtain a reconstructed feature map with enhanced features, which can preserve and highlight weak edge, corner, and texture features. Compared with the traditional direct convolutional downsampling feature extraction method, the scheme of this application can significantly alleviate the problem of decreased detection performance of small targets due to feature blurring, improve the representation capability of fine-grained details, and improve the discriminability of small targets. By performing residual connection between the reconstructed feature map and the enhanced feature map, an enhanced feature map is obtained. Since the features in the reconstructed feature map and the aggregated feature map have been enhanced, targets can be located and identified more accurately from the enhanced feature map, improving the accuracy of target detection.

[0127] The following describes a computing device provided by an embodiment of this application.

[0128] Figure 5 This is a schematic diagram of the structure of a computing device provided in an embodiment of this application. Figure 5 As shown, the computing device includes one or more processors 501, a memory 502, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The components are interconnected via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processors can process instructions executed within the computing device, including instructions stored in or on memory to display graphical information of a GUI on an external input / output device (such as a display computing device coupled to the interface). In other embodiments, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple computing devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system). Figure 4 Take a processor 501 as an example.

[0129] The memory 502 is the non-transitory computer-readable storage medium provided in this application. The memory stores instructions executable by at least one processor to cause the at least one processor to perform the method provided in this application. The non-transitory computer-readable storage medium of this application stores computer instructions for causing a computer to perform the method provided in this application.

[0130] The memory 502, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as the program instructions / modules corresponding to the methods in the embodiments of this application. The processor 501 executes various functional applications and data processing of the server by running the non-transitory software programs, instructions, and modules stored in the memory 502, thereby implementing the methods in the above-described method embodiments.

[0131] Memory 502 may include a program storage area and a data storage area, wherein the program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the computing device, etc. Furthermore, memory 502 may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory 502 may optionally include memory remotely located relative to processor 501, and these remote memories may be connected to the computing device via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0132] The computing device may further include an input device 503 and an output device 504. The processor 501, memory 502, input device 503, and output device 504 can be connected via a bus or other means. Figure 4 Taking the example of a connection between China and Israel via a bus.

[0133] Input device 503 can receive input numerical or character information, and generate key signal inputs related to user settings and function control of the computing device, such as touch screens, keypads, mice, trackpads, touchpads, joysticks, one or more mouse buttons, trackballs, joysticks, etc. Output device 504 may include display computing devices, auxiliary lighting devices (e.g., LEDs), and haptic feedback devices (e.g., vibration motors). The display computing device may include, but is not limited to, liquid crystal displays (LCDs), light-emitting diode (LED) displays, and plasma displays. In some embodiments, the display computing device may be a touch screen.

[0134] Furthermore, in conjunction with the above embodiments, this application embodiment can be implemented using a computer storage medium. This computer storage medium stores computer program instructions; when these computer program instructions are executed by a processor, they implement any of the methods described in the above embodiments.

[0135] The functional blocks shown in the above-described structural diagram can be implemented as hardware, software, firmware, or a combination thereof. When implemented in hardware, they can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this application are programs or code segments used to perform the required tasks. Programs or code segments can be stored on a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried on a carrier wave. "Machine-readable medium" can include any medium capable of storing or transmitting information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, ROM, flash memory, erasable ROM (EROM), floppy disks, CD-ROMs, optical disks, hard disks, fiber optic media, radio frequency (RF) links, etc. Code segments can be downloaded via computer networks such as the Internet, intranets, etc.

[0136] It should also be noted that the exemplary embodiments mentioned in this application describe methods or systems based on a series of steps or apparatus. However, this application is not limited to the order of the above steps; that is, the steps can be performed in the order mentioned in the embodiments, or in a different order, or several steps can be performed simultaneously.

[0137] The aspects of this application have been described above with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It should be understood that each block in the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that these instructions, executable via the processor of the computer or other programmable data processing apparatus, enable the implementation of the functions / actions specified in one or more blocks of the flowchart illustrations and / or block diagrams. Such a processor can be, but is not limited to, a general-purpose processor, a special-purpose processor, a special application processor, or a field-programmable logic circuit. It is also understood that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can also be implemented by dedicated hardware performing the specified functions or actions, or can be implemented by a combination of dedicated hardware and computer instructions.

[0138] The above description is merely a specific implementation of this application. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, modules, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. It should be understood that the protection scope of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the protection scope of this application.

Claims

1. A target detection method for remote sensing images based on wavelet transform, characterized in that, The method includes: Acquire remote sensing images; The remote sensing image is subjected to multi-scale feature adaptive fusion processing to obtain an aggregated feature map; The aggregated feature map is subjected to discrete wavelet transform to obtain a low-frequency sub-band and at least one high-frequency sub-band. The low-frequency sub-band is used to indicate the smooth background and structural information of the remote sensing image, and the at least one high-frequency sub-band is used to indicate the edge features and texture features of the remote sensing image. Performing inverse wavelet transform on the low-frequency sub-band and at least one high-frequency sub-band to obtain a reconstructed feature map includes: generating weights for the low-frequency sub-band and the at least one high-frequency sub-band respectively; wherein, for each sub-band, attention weights are calculated using a local-global adaptive probability density function, the local-global adaptive probability density function being a Gaussian distribution probability density PDF function constructed based on the mixture variance generated from the local mean, local variance, global mean, and global variance of each sub-band; wherein, the calculation of the local mean and local variance of each sub-band is implemented using 3x3 average pooling, and the local mean of the sub-band satisfies... The local variance of the subband satisfies The formula for the mixture variance is: ,in, In the formula, Characterizing mixed variance, Characterize learnable adaptive weights. Characterizing local variance, Characterizing global variance, Characterizing subbands, Characterize the activation function, The characterization uses a 3x3 convolution kernel for convolution operations; the Gaussian probability density function (PDF) satisfies... In the formula, It is a local mean, where λ represents a learnable PDF width control parameter. For the mixed standard deviation, the attention weights of subband F satisfy... In the formula, γ represents the numerical stability factor in the attention function, and γ represents the scaling factor used to adjust the PDF intensity in the attention function. Based on the respective weights of the low-frequency sub-band and the at least one high-frequency sub-band, the low-frequency sub-band and the at least one high-frequency sub-band are weighted and enhanced respectively to obtain the enhanced sub-band; the enhanced sub-band is then subjected to inverse wavelet transform to obtain the reconstructed feature map; Residual connections are performed on the reconstructed feature map and the aggregated feature map to obtain an enhanced feature map; Based on the enhanced feature map, targets in the remote sensing image are located and identified.

2. The method according to claim 1, characterized in that, The process of performing multi-scale feature adaptive fusion processing on the remote sensing image to obtain an aggregated feature map includes: Multi-scale feature extraction is performed on the remote sensing image to obtain multiple scale features; The aggregated feature map is obtained by adaptively weighting and fusing the multiple scale features.

3. The method according to claim 2, characterized in that, The adaptive weighted fusion of the multiple scale features to obtain the aggregated feature map includes: The aggregated feature map is obtained by performing channel-level adaptive weighted summation on the multiple scale features.

4. The method according to claim 2 or 3, characterized in that, The plurality of scale features include scale features for indicating the spatial structure of the remote sensing image and scale features for indicating the semantic context of the remote sensing image.

5. The method according to claim 4, characterized in that, The at least one high-frequency sub-band is used to indicate the edge and texture features of the remote sensing image in at least one of the following directions: Horizontal direction, vertical direction, diagonal direction.

6. A target detection device for remote sensing images based on wavelet transform, characterized in that, The device includes: The acquisition module is used to acquire remote sensing images; The fusion processing module is used to perform multi-scale feature adaptive fusion processing on the remote sensing image to obtain an aggregated feature map; The first transformation module is used to perform discrete wavelet transform on the aggregated feature map to obtain a low-frequency sub-band and at least one high-frequency sub-band. The low-frequency sub-band is used to indicate the smooth background and structural information of the remote sensing image, and the at least one high-frequency sub-band is used to indicate the edge features and texture features of the remote sensing image. The second transformation module is used to perform inverse wavelet transform on the low-frequency sub-band and at least one high-frequency sub-band to obtain a reconstructed feature map; including: generating weights for the low-frequency sub-band and the at least one high-frequency sub-band respectively; wherein, for each sub-band, attention weights are calculated using a local-global adaptive probability density function, the local-global adaptive probability density function being a Gaussian distribution probability density PDF function constructed based on the mixture variance generated from the local mean, local variance, global mean, and global variance of each sub-band; wherein, the calculation of the local mean and local variance of each sub-band is implemented using 3x3 average pooling, and the local mean of the sub-band satisfies The local variance of the subband satisfies The formula for the mixture variance is: ,in, In the formula, Characterizing mixed variance, Characterize learnable adaptive weights. Characterizing local variance, Characterizing global variance, Characterizing subbands, Characterize the activation function, The characterization uses a 3x3 convolution kernel for convolution operations; the Gaussian probability density function (PDF) satisfies... In the formula, It is a local mean, where λ represents a learnable PDF width control parameter. For the mixed standard deviation, the attention weights of subband F satisfy... In the formula, γ represents the numerical stabilization factor in the attention function, and γ represents the scaling factor in the attention function used to adjust the PDF intensity. Based on the respective weights of the low-frequency sub-band and the at least one high-frequency sub-band, the low-frequency sub-band and the at least one high-frequency sub-band are weighted and enhanced to obtain an enhanced sub-band. An inverse wavelet transform is performed on the enhanced sub-band to obtain the reconstructed feature map. A connection module is used to perform residual connections on the reconstructed feature map and the aggregated feature map to obtain an enhanced feature map; The identification module is used to locate and identify targets in the remote sensing image based on the enhanced feature map.

7. The apparatus according to claim 6, characterized in that, The fusion processing module is used for: Multi-scale feature extraction is performed on the remote sensing image to obtain multiple scale features; The aggregated feature map is obtained by adaptively weighting and fusing the multiple scale features.

8. The apparatus according to claim 7, characterized in that, The fusion processing module is used to perform channel-level adaptive weighted summation on the multiple scale features to obtain the aggregated feature map.

9. The apparatus according to claim 7 or 8, characterized in that, The plurality of scale features include scale features for indicating the spatial structure of the remote sensing image and scale features for indicating the semantic context of the remote sensing image.

Images