Remote sensing image target detection method and device, storage medium and electronic equipment

Abstract

This application discloses a method, apparatus, storage medium, and electronic device for target detection in remote sensing images. The method includes: acquiring a first remote sensing image to be detected; extracting features from the first remote sensing image using a pre-set feature extraction network to obtain first feature images at multiple different levels; processing the first feature images at multiple different levels using a first feature processing unit of a pre-trained remote sensing target detection model to obtain corresponding first fused feature images; detecting the first fused feature images at multiple different levels using a first detection unit of the remote sensing target detection model to obtain detection results corresponding to the first fused feature images at multiple different levels, and determining a first multi-scale target in the first remote sensing image based on the detection results corresponding to different levels. This application solves the technical problem that remote sensing target detection technology cannot capture long-distance dependencies of cross-scale features, leading to aliasing effect verification.

Description

Technical Field

[0001] This application relates to the field of target detection technology, and more specifically, to a method, apparatus, storage medium, and electronic device for target detection in remote sensing images. Background Technology

[0002] With the development of computer technology and the continuous advancement of visual principles, target detection using computer image processing technology has been widely applied in numerous fields such as military reconnaissance, public security, intelligent transportation, and urban planning. Target detection technology using remote sensing images is a new technology that has emerged with the development of remote sensing technology. Its advantages, including long operating distance, wide coverage, and high efficiency, give it significant military and civilian value.

[0003] However, although high-resolution remote sensing images can provide detailed information about ground objects, they also have extensive and complex backgrounds (such as occlusion conditions), which leads to low detection accuracy of related remote sensing image target detection methods and limits the application scope of remote sensing images in target detection technology.

[0004] There is currently no effective solution to the above problems. Summary of the Invention

[0005] This application provides a remote sensing image target detection method, apparatus, storage medium, and electronic device to at least solve the technical problem that remote sensing target detection technology cannot capture long-distance dependencies of cross-scale features, leading to aliasing effect verification.

[0006] According to one aspect of the embodiments of this application, a remote sensing image target detection method is provided, comprising: acquiring a first remote sensing image to be detected; extracting features from the first remote sensing image using a preset feature extraction network to obtain first feature images of the first remote sensing image at multiple different levels; and performing feature processing on the first feature images at multiple different levels through a first feature processing unit of a pre-trained remote sensing target detection model to obtain first fused feature images at multiple different levels, wherein the first feature processing unit includes: a target feature pyramid network based on non-local perception and attention mechanism for multi-scale feature fusion processing of the first feature images at multiple different levels, and an input method for the target feature pyramid network. The system employs a multi-scale feature decoupling network to decouple the first shallow fused feature image, an adaptive multi-scale feature allocation network to adjust the multi-scale feature ratio of the first mid-layer fused feature image and the first deep fused feature image output by the target feature pyramid network, and a filtering invalid information strategy network to separate image information from the optimized fused feature image output by the multi-scale feature decoupling network and the adaptive multi-scale feature allocation network. The first detection unit of the remote sensing target detection model detects multiple first fused feature images at different levels, obtaining detection results corresponding to multiple first fused feature images at different levels, and determining the first multi-scale target of the first remote sensing image based on the detection results corresponding to different levels.

[0007] Optionally, a preset feature extraction network is used to extract features from the remote sensing image to obtain first feature images of the remote sensing image at multiple different levels, including: preprocessing the first remote sensing image; and using the feature extraction network to extract features from the processed first remote sensing image to obtain first feature images of the first remote sensing image at multiple different levels, wherein the first feature images include: a first shallow feature image, a first medium feature image, and a first deep feature image.

[0008] Optionally, the training process of the remote sensing target detection model includes: acquiring a sample image set, wherein the sample image set includes multiple second remote sensing images, and each second remote sensing image includes at least one second remote sensing target; using a feature extraction network to extract features from each first remote sensing image in the sample image set, obtaining second feature images of each first remote sensing image at multiple different levels, wherein the second feature images include: a second shallow feature image, a second middle feature image, and a second deep feature image; constructing an initial remote sensing target detection model, wherein the initial remote sensing target detection model includes: a second feature processing unit and a second detection unit; training the initial remote sensing target detection model based on the second feature images of each second remote sensing image in the sample image set at multiple different levels, to obtain the remote sensing target detection model.

[0009] Optionally, the construction process of the second feature processing unit includes: constructing a target feature pyramid network that integrates non-local perception and dual attention mechanisms, wherein the target feature pyramid module includes: a shallow feature network, a middle feature network, and a deep feature network; constructing an adaptive allocation multi-scale feature network that integrates a position-aware attention mechanism; constructing a multi-scale feature decoupling network using a gate function, and constructing a filtering invalid information strategy network using a membership function.

[0010] Optionally, the initial remote sensing target detection model is trained based on the second feature images of each second remote sensing image in the sample image set at multiple different levels to obtain the remote sensing target detection model. This includes: for each second remote sensing image, inputting the second feature images of the second remote sensing image at multiple different levels into the initial remote sensing target detection model to obtain the second multi-scale target output by the initial remote sensing target detection model; constructing a target loss function based on the second remote sensing target and the second multi-scale target in the sample image set, wherein the target loss function is used to reflect the difference between the second remote sensing target and the second multi-scale target; and adjusting the model parameters of the initial remote sensing target detection model based on the target loss function to obtain the remote sensing target detection model.

[0011] Optionally, the second feature images of the second remote sensing image at multiple different levels are input into the initial remote sensing target detection model to obtain the second multi-scale target output by the initial remote sensing target detection model. This includes: using a target feature pyramid network to perform multi-scale feature fusion processing on the second feature images at multiple different levels to obtain a second shallow fused feature image corresponding to the second shallow feature image, a second middle fused feature image corresponding to the second middle feature image, and a second deep fused feature image corresponding to the second deep feature image; inputting the second shallow fused feature image into a multi-scale feature decoupling network to obtain a third shallow fused feature image, and inputting the second middle fused feature image and the second deep fused feature image into an adaptively assigned multi-scale feature network to obtain a third middle fused feature image and a third deep fused feature image; and inputting the third shallow fused feature image into a multi-scale feature decoupling network to obtain a third middle fused feature image and a third deep fused feature image. The image, the third mid-level fusion feature image, and the third deep fusion feature image are respectively input into the corresponding filtering invalid information strategy network to obtain the first image information corresponding to the second shallow fusion feature image, the second image information corresponding to the second mid-level fusion feature image, and the third image information corresponding to the second deep fusion feature image. The first image information, the second image information, and the third image information include foreground information or background information. The first image information, the second image information, and the third image information are respectively input into the second detection unit at the corresponding level to obtain the shallow feature detection results, mid-level feature detection results, and deep feature detection results output by the second detection unit at different levels. Based on the shallow feature detection results, mid-level feature detection results, and deep feature detection results output by the second detection unit at different levels, the second multi-scale target of the second remote sensing image is determined.

[0012] Optionally, the feature extraction network includes at least one of the following: LeNet-5 network, AlexNet network, VGGNet network, GoogLeNet network, and ResNet network.

[0013] According to another aspect of the embodiments of this application, a remote sensing image target detection device is also provided, comprising: an acquisition module for acquiring a first remote sensing image to be detected; a feature extraction module for extracting features from the first remote sensing image using a preset feature extraction network to obtain first feature images of the first remote sensing image at multiple different levels; and a feature processing module for performing feature processing on the first feature images at multiple different levels through a first feature processing unit of a pre-trained remote sensing target detection model to obtain first fused feature images at multiple different levels, wherein the first feature processing unit includes: a target feature pyramid network based on non-local perception and attention mechanism for performing multi-scale feature fusion processing on the first feature images at multiple different levels; and a feature processing module for... The system comprises: a multi-scale feature decoupling network for decoupling the first shallow fused feature image output by the target feature pyramid network; an adaptive multi-scale feature allocation network for adjusting the proportion of multi-scale features in the first mid-layer fused feature image and the first deep fused feature image output by the target feature pyramid network; and a filtering invalid information strategy network for separating image information from the fused feature images output by the multi-scale feature decoupling network and the adaptive multi-scale feature allocation network. The target detection module is used to detect multiple first fused feature images at different levels through the first detection unit of the remote sensing target detection model, obtain detection results corresponding to multiple first fused feature images at different levels, and determine the first multi-scale target in the first remote sensing image based on the detection results corresponding to different levels.

[0014] According to another aspect of the embodiments of this application, a non-volatile storage medium is also provided, the non-volatile storage medium including a stored computer program, wherein the device where the non-volatile storage medium is located executes the above-described remote sensing image target detection method by running the computer program.

[0015] According to another aspect of the embodiments of this application, an electronic device is also provided, the electronic device including: a memory and a processor, wherein the memory stores a computer program, and the processor is configured to execute the above-described remote sensing image target detection method through the computer program.

[0016] In this embodiment, a first remote sensing image to be detected is acquired; features are extracted from the first remote sensing image using a preset feature extraction network to obtain first feature images of the first remote sensing image at multiple different levels; feature processing is performed on the first feature images at multiple different levels by a first feature processing unit of a pre-trained remote sensing target detection model to obtain first fused feature images at multiple different levels. The first feature processing unit includes: a target feature pyramid network based on non-local perception and attention mechanisms for multi-scale feature fusion processing of the first feature images at multiple different levels; and a first shallow fused feature image output by the target feature pyramid network. The system includes a multi-scale feature decoupling network for decoupling processing, an adaptive multi-scale feature allocation network for adjusting the multi-scale feature ratio of the first mid-level fusion feature image and the first deep fusion feature image output by the target feature pyramid network, and a filtering invalid information strategy network for separating image information from the optimized fusion feature image output by the multi-scale feature decoupling network and the adaptive multi-scale feature allocation network. The first detection unit of the remote sensing target detection model detects multiple first fusion feature images at different levels, obtains the detection results corresponding to multiple first fusion feature images at different levels, and determines the first multi-scale target of the first remote sensing image based on the detection results corresponding to different levels.

[0017] In the above technical solution, a target feature pyramid network that integrates nonlocal perception and attention mechanisms can capture salient features across scales, improve the representation capability of multi-scale features, and reduce aliasing effects. An adaptive multi-scale feature allocation network adaptively modifies the proportion of multi-scale features, enhancing the flexibility of feature fusion. A multi-scale feature decoupling network eliminates the influence of large-scale target features on small-scale target features, avoiding missed detection of small-scale object features in shallow fusion. An invalid information filtering strategy network filters invalid information from the processed feature image, suppressing background information and improving the accuracy of target detection. This solves the problem of aliasing effect verification techniques caused by the inability of remote sensing target detection technology to capture long-distance dependencies of cross-scale features. Attached Figure Description

[0018] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0019] Figure 1 This is a hardware structure block diagram of an optional computer terminal (or mobile device) for implementing a remote sensing image target detection method according to an embodiment of this application;

[0020] Figure 2 This is a schematic flowchart of an optional remote sensing image target detection method according to an embodiment of this application;

[0021] Figure 3 This is a schematic diagram of the structure of an optional remote sensing target detection model according to an embodiment of this application;

[0022] Figure 4 This is a schematic diagram of an optional nonlocal dual attention network according to an embodiment of this application;

[0023] Figure 5 This is a schematic diagram of a position-aware attention network structure for an optional network embedding a position-aware attention mechanism according to an embodiment of this application;

[0024] Figure 6 This is a schematic diagram of an optional multi-scale feature decoupling network structure according to an embodiment of this application;

[0025] Figure 7 This is a schematic diagram of an optional remote sensing image target detection result according to an embodiment of this application;

[0026] Figure 8 This is a schematic diagram of an optional remote sensing image target detection device according to an embodiment of this application. Detailed Implementation

[0027] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0028] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0029] Furthermore, all information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for display, data used for analysis, etc.) involved in this application are information and data authorized by the user or fully authorized by all parties. For example, this system has an interface with the relevant user or organization. Before obtaining relevant information, it needs to send an acquisition request to the aforementioned user or organization through the interface, and obtain the relevant information after receiving consent information from the aforementioned user or organization.

[0030] Example 1

[0031] Target detection in optical remote sensing images presents numerous challenges, including the large number of instances, large image size, and complex background textures. Furthermore, the significant untapped potential of optical remote sensing images in both civilian and military applications remains a major obstacle. In recent years, the increased resolution of satellite optical images has enabled the identification of even smaller targets within remote sensing images. While high-resolution remote sensing images provide detailed information about ground objects, the resulting increased complexity of the background further complicates target detection. This is particularly true for detecting targets of various types and scales at different times, as remote sensing targets are often distributed and their locations and distances are not fixed, posing a significant challenge to the field.

[0032] In recent years, researchers have made continuous efforts in deep convolutional neural networks. In multi-scale object recognition, most detection techniques typically use Feature Pyramid Networks (FPNs) as their neck structure to construct multi-scale features. However, FPNs have been found to have some well-known drawbacks, such as the easy loss of small-scale feature information during downsampling due to the limited pixel information of small objects, and aliasing effects in cross-scale fusion. Current research has made many attempts to alleviate these problems, and while these improved methods have shown effectiveness in improving performance, their structures still cannot fully resolve aliasing effects.

[0033] Due to semantic differences and feature misalignment, direct additive cross-scale fusion leads to inconsistent feature integration. Furthermore, the aliasing effect caused by cross-scale fusion becomes more severe as the backbone network deepens. Related research has found that attention mechanisms can be used to alleviate this problem. However, attention mechanisms fail to fully consider the advantages of channel and spatial features and ignore long-distance dependencies between cross-scale features.

[0034] Therefore, to solve this problem, this application provides an embodiment of a remote sensing image target detection method, which is described in detail below. It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowcharts, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0035] The methods and embodiments provided in this application can be executed on mobile terminals, computer terminals, or similar computing devices. Figure 1 A hardware block diagram of a computer terminal (or mobile device) for implementing a target detection method in remote sensing images is shown. Figure 1 As shown, the computer terminal 10 (or mobile device 10) may include one or more processors 102 (shown as 102a, 102b, ..., 102n in the figure) 102 (processor 102 may include, but is not limited to, a microprocessor MCU or a programmable logic device FPGA, etc.), a memory 104 for storing data, and a transmission device 106 for communication functions. In addition, it may also include: a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of a BUS bus), a network interface, a power supply, and / or a camera. Those skilled in the art will understand that... Figure 1 The structure shown is for illustrative purposes only and does not limit the structure of the aforementioned electronic device. For example, computer terminal 10 may also include... Figure 1 The more or fewer components shown, or having the same Figure 1 The different configurations shown.

[0036] It should be noted that the aforementioned one or more processors 102 and / or other data processing circuits are generally referred to herein as "data processing circuits". These data processing circuits may be embodied, in whole or in part, in software, hardware, firmware, or any other combination thereof. Furthermore, the data processing circuits may be a single, independent processing module, or may be integrated, in whole or in part, into any other element within the computer terminal 10 (or mobile device). As involved in the embodiments of this application, the data processing circuits serve as a processor control mechanism (e.g., selection of a variable resistor termination path connected to an interface).

[0037] The memory 104 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the remote sensing image target detection method in this embodiment. The processor 102 executes various functional applications and data processing by running the software programs and modules stored in the memory 104, thereby realizing the aforementioned remote sensing image target detection method. The memory 104 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include memory remotely located relative to the processor 102, and these remote memories can be connected to the computer terminal 10 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.

[0038] The transmission device 106 is used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the communication provider of the computer terminal 10. In one example, the transmission device 106 includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device 106 may be a Radio Frequency (RF) module, used for wireless communication with the Internet.

[0039] The display may be, for example, a touchscreen liquid crystal display (LCD) that allows the user to interact with the user interface of the computer terminal 10 (or mobile device).

[0040] Under the above operating environment, Figure 2 This is a schematic flowchart of an optional remote sensing image target detection method according to an embodiment of this application, as shown below. Figure 2 As shown, the method includes at least steps S202-S208, wherein:

[0041] Step S202: Obtain the first remote sensing image to be detected.

[0042] In the technical solution provided in step S202, the remote sensing image is an image of the Earth's surface acquired through remote sensing technology. Therefore, the remote sensing image may include remote sensing targets such as aircraft, ports, and ships. Due to its special structure and imaging method, the objects within the remote sensing image are characterized by wide distribution, large number, and diverse shapes and sizes. Therefore, it is difficult to accurately extract the objects within the remote sensing image using traditional image processing methods. Therefore, in this embodiment, target detection will be performed on the first remote sensing image to be detected through the following steps.

[0043] Step S204: Use a preset feature extraction network to extract features from the first remote sensing image to obtain the first feature image of the first remote sensing image at multiple different levels.

[0044] In the technical solution provided in step S204, the aforementioned feature extraction network can also be called the backbone network, which is used to extract multi-level feature maps at multiple scale resolutions from the first remote sensing image.

[0045] Step S206: The first feature processing unit in the pre-trained remote sensing target detection model performs feature processing on multiple first feature images at different levels to obtain multiple first fused feature images at different levels.

[0046] In the technical solution provided in step S206, the first feature processing unit includes: a target feature pyramid network based on nonlocal perception and attention mechanism, a multi-scale feature decoupling network, an adaptive multi-scale feature allocation network, and a strategy network for filtering invalid information.

[0047] Specifically, aliasing effects may occur when integrating features at multiple resolution scales without considering semantic differences. These aliasing effects negatively impact the localization and recognition of objects in remote sensing images. Therefore, this application proposes an improved target feature pyramid network (NDA-FPN) based on nonlocal perception and attention mechanisms to perform multi-scale feature fusion processing on multiple first feature images at different levels, resulting in multiple first fused feature images at different levels, including: a first shallow fused feature image, a first mid-level fused feature image, and a first deep fused feature image. Thus, the target feature pyramid network using nonlocal perception and attention mechanisms can capture long-distance dependencies across scales on feature maps and obtain salient features across scales.

[0048] Within a shallow fused feature image, detailed features of both large-scale and small-scale objects coexist. However, the detailed features of large-scale objects are often more prominent. Furthermore, the relevant object detector focuses more on prominent features, leading to the neglect of detailed features of small-sized objects within the shallow fused feature image, resulting in missed detections. Therefore, to avoid the above problems, this application proposes a multi-scale feature decoupling network to decouple the features of large-scale and small-scale objects within the first shallow fused feature image output by the target feature pyramid network, thereby improving the detection accuracy of small-scale objects.

[0049] Since objects in remote sensing images have different scales, and fixed-scale feature extraction methods cannot capture object information at all scales, in order to avoid the above problems, the adaptive multi-scale feature allocation network proposed in this application is used to adjust the multi-scale feature ratio of the first middle-layer fusion feature image and the first deep-layer fusion feature image output by the target feature pyramid network, so as to extract features at different scales, thereby better capturing the details and contextual information of the target object and improving the accuracy and robustness of object detection and recognition.

[0050] Because remote sensing images have complex backgrounds and dense distributions, a large amount of noise information may be introduced during feature extraction. Therefore, in order to improve the influence of salient information on relevant features and reduce the influence of irrelevant background factors, in this embodiment, an invalid information filtering strategy network is connected after the output of the multi-scale feature decoupling network and the adaptive allocation multi-scale feature network, respectively, to separate image information in the output first fused feature image and improve the accuracy of target detection.

[0051] Step S208: The first detection unit of the remote sensing target detection model detects multiple first fusion feature images at different levels to obtain the detection results corresponding to the first fusion feature images at different levels, and determines the first multi-scale target of the first remote sensing image based on the detection results corresponding to different levels.

[0052] In the technical solution provided in step S208, the first detection unit in the pre-trained remote sensing target detection model detects multiple first fusion feature images at different levels. The first detection unit performs target recognition and classification on the first fusion feature images at different levels to obtain the first multi-scale target of the first remote sensing image.

[0053] Based on the scheme defined in steps S202 to S208 above, it can be understood that in the embodiment, a first remote sensing image to be detected is acquired; features of the first remote sensing image are extracted using a preset feature extraction network to obtain first feature images of the first remote sensing image at multiple different levels; the first feature processing unit of a pre-trained remote sensing target detection model performs feature processing on the first feature images at multiple different levels to obtain first fused feature images at multiple different levels. The first feature processing unit includes: a target feature pyramid network based on non-local perception and attention mechanisms for multi-scale feature fusion processing of the first feature images at multiple different levels; and a target feature pyramid network... The system comprises a multi-scale feature decoupling network for decoupling the first shallow fused feature image output; an adaptive multi-scale feature allocation network for adjusting the multi-scale feature ratio of the first mid-layer fused feature image and the first deep fused feature image output by the target feature pyramid network; and a filtering invalid information strategy network for separating image information from the optimized fused feature image output by the multi-scale feature decoupling network and the adaptive multi-scale feature allocation network. The first detection unit of the remote sensing target detection model detects multiple first fused feature images at different levels, obtaining detection results corresponding to multiple first fused feature images at different levels. Based on the detection results corresponding to different levels, the first multi-scale target of the first remote sensing image is determined.

[0054] Therefore, the target feature pyramid network fusion mechanism combining nonlocal perception and attention in this application embodiment can capture salient features across scales, improve the representation capability of multi-scale features, and reduce aliasing effects. The adaptive multi-scale feature allocation network adaptively modifies the proportion of multi-scale features, enhancing the flexibility of feature fusion. The multi-scale feature decoupling network eliminates the influence of large-scale target features on small-scale target features, avoiding missed detection of small-scale object target features in shallow fusion features. The invalid information filtering strategy network filters invalid information from the processed feature image, suppressing background information and improving the accuracy of target detection. This solves the problem of remote sensing target detection technology's inability to capture long-distance dependencies of cross-scale features, leading to aliasing effect verification issues.

[0055] The implementation methods of the above steps will be further described in detail below.

[0056] As an optional implementation, in the technical solution provided in step S204 above, the method may include: preprocessing the first remote sensing image; using a feature extraction network to extract features from the processed first remote sensing image to obtain first feature images of the first remote sensing image at multiple different levels, wherein the first feature images include: a first shallow feature image, a first medium-level feature image, and a first deep feature image.

[0057] In this embodiment, since the features of remote sensing targets vary with different resolution scales, the acquired first remote sensing image to be detected can first be preprocessed. This preprocessing includes, but is not limited to, histogram equalization to improve contrast, filtering to remove noise, and image augmentation through flipping, rotating, and scaling to ensure the quality of the remote sensing image and enhance the features of the remote sensing target, thereby improving subsequent detection accuracy. Next, the processed first remote sensing image is input into a pre-defined feature extraction network for continuous local convolution and pooling operations to obtain feature maps at different levels of the first remote sensing image. Each convolution or pooling operation yields a feature map at a specific resolution scale, helping to obtain the spatial distribution information of the first remote sensing target at multiple resolution scales, thus improving the accuracy of multi-scale target detection.

[0058] Optionally, the feature extraction network includes at least one of the following: LeNet-5 network, AlexNet network, VGGNet network, GoogLeNet network, and ResNet network. For example, in this embodiment, VGGNet network can be selected as the feature extraction network, wherein the VGGNet network includes three layers: Conv1_1, Conv3_3, and Conv5_5, which serve as feature extraction layers to output multiple first feature maps of the first remote sensing image.

[0059] Optionally, the training process of the remote sensing target detection model includes the following steps S1-S4, wherein:

[0060] Step S1: Obtain a sample image set, wherein the sample image set includes multiple second remote sensing images, and each second remote sensing image includes at least one second remote sensing target;

[0061] Step S2: Use a feature extraction network to extract features from each of the first remote sensing images in the sample image set to obtain second feature images of each of the first remote sensing images at multiple different levels. The second feature images include: a second shallow feature image, a second middle feature image, and a second deep feature image.

[0062] Step S3: Construct an initial remote sensing target detection model, which includes: a second feature processing unit and a second detection unit;

[0063] Step S4: Train the initial remote sensing target detection model based on the second feature images of each second remote sensing image in the sample image set at multiple different levels to obtain the remote sensing target detection model.

[0064] In this embodiment, the feature extraction network can also be any one of LeNet-5, AlexNet, VGGNet, GoogLeNet, ResNet, etc.

[0065] Specifically, Figure 3 This is a schematic diagram of the structure of an optional initial remote sensing target detection model according to an embodiment of this application, as shown below. Figure 3 As shown, the initial remote sensing target detection model includes a second feature processing unit and a second detection unit. The construction process of the second detection unit includes the following steps S31-S33, wherein:

[0066] Step S31: Construct a target feature pyramid network that integrates nonlocal perception and dual attention mechanisms.

[0067] like Figure 3 As shown, the target feature pyramid network includes a shallow feature network, a mid-level feature network, and a deep feature network. Furthermore, the aforementioned non-local perception refers to capturing global and local feature information at different scales, while the attention mechanism includes two dual attention mechanisms: spatial attention and channel attention. By fusing these dual attention mechanisms, features at different levels can be adaptively added to achieve better feature fusion. In other words, the target feature pyramid in this embodiment is an improved feature pyramid that embeds non-local dual attention on top of an existing feature pyramid, due to the limitations of existing feature pyramids.

[0068] Figure 4 This is a schematic diagram of an optional nonlocal dual attention network according to an embodiment of this application, such as... Figure 4 As shown, the backbone network is transmitted through... Convolutional layers Convolutional layers and The multi-level feature maps output from the convolutional layers are concatenated and transformed into a vector format, which is then used as input to the non-local dual attention network, while the feature maps input to the non-local dual attention module... Furthermore, since the weights are distributed across multiple multi-resolution scale feature maps at different levels through weight sharing, the input feature map of the input non-local dual attention network can be written as the following expression:

[0069]

[0070] Where N represents the number of pixels in the spatial dimension (N=H) W), C represents the number of channels. Vectorized functions representing the spatial and channel dependencies of feature maps Represents a nonlocal feature map. These represent multi-scale features. Therefore, the three input functions of the dual attention mechanism can be defined as follows:

[0071]

[0072]

[0073]

[0074] in, This represents a learnable transformation of the input. These represent the query, key, and value, respectively. represents the embedding matrix of the input feature map in different feature layers (the size of the embedding matrix is ​​equal to the size of the convolution kernel in the feature layer), and X represents the original input feature map.

[0075] When calculating nonlocal attention weights, the sum of the matrices of K and Q can be used as the similarity matrix, where the expression for the similarity matrix is:

[0076]

[0077] in, Let S denote the transpose of Q. After obtaining the similarity matrix S, an attention matrix can be constructed using the softmax function. This attention matrix can represent the correlation between spatial multi-scale dependencies and channel dependencies. Therefore, the expression for the attention matrix can be written as:

[0078]

[0079] Next, the weights of all positions within the input feature map are weighted to obtain a weighted feature map. Then, the formaldehyde feature map is pooled to obtain the final output feature. The expression for the input feature is:

[0080]

[0081]

[0082] Among them, the above This represents the attention weight of position i to position j. This represents all the attention weights for that row. Therefore, the saliency features output by the non-local attention network... It can be represented as:

[0083]

[0084] in, Representation layer normalization, Indicates the eigenvector Reshaped into (i.e., the original shape), matrix This indicates that the weight matrix obtained through Conv1_1 convolution with a non-linear activation function ensures that the size of the feature map remains unchanged.

[0085] Therefore, by integrating the above Figure 4 The target feature pyramid network of the nonlocal dual attention network shown can extract features at different scales through multi-level convolution and pooling operations, thereby better capturing details and contextual information in the image and making the model more robust to scale changes of the input feature image. At the same time, it can automatically adjust the size of the receptive field to adapt to target objects at different scales, and can learn spatial multi-scale nonlocal dependencies and channel dependencies.

[0086] Step S32: Construct an adaptive multi-scale feature network that integrates a location-aware attention mechanism.

[0087] Specifically, since targets at different scales within remote sensing images depend on multi-scale features to varying degrees, it is necessary to adjust the proportion of multi-scale feature information using an adaptive multi-scale feature allocation network. Furthermore, most adaptive multi-scale methods omit the importance of location information; therefore, this application uses an adaptive multi-scale feature allocation network with an embedded location-aware attention mechanism to adjust the proportion of multi-scale features.

[0088] Figure 5 This is a schematic diagram of the structure of a position-aware attention network according to an optional embodiment of this application, which incorporates a position-aware attention mechanism. Figure 5 As shown.

[0089] First, salient features are obtained through a location-aware attention network, and the location attention weights can be expressed as:

[0090]

[0091] in, and sign Let L2 normalization function and time function represent the time function, respectively. Cross-layer features representing the target feature pyramid network This results in a cross-layer feature. Since semantic information of objects at different scales is represented at different scales in different network layers (e.g., higher layers have larger receptive fields), this embodiment leverages the advantages of cross-layer semantics to further improve the interpretability of multi-scale learning. Here, we use... Indicates the ratio of the next layer Therefore, the function of variable P is to identify spatial correlations between different levels as attention weights.

[0092] Next, the original feature map is enhanced using positional weights, and multi-scale information features are adaptively assigned according to the following formula, thereby selecting the most relevant features for feature maps at different levels:

[0093]

[0094]

[0095] Among them, the above express The values ​​in the i-th row and j-th column, where n and N represent the number of multi-scale features, respectively. These represent the spatial adaptive weights at three different levels. Indicates the output features, where S represents the number of layers in the backbone network output. The channels are first compressed using Conv1_1 convolution to reduce workload.

[0096] It should be noted that the sum of the attention weights of the three different levels of multi-scale features at the same location is equal to 1. Furthermore, as the location changes, the attention weights of the three different levels of multi-scale features also change, thereby achieving the purpose of adaptive adjustment of the proportion of multi-scale information.

[0097] Step S33: A multi-scale feature decoupling network is constructed using a gate function, and a membership function is used to construct a strategy network for filtering invalid information.

[0098] Specifically, Figure 6 This is a schematic diagram of an optional multi-scale feature decoupling network according to an embodiment of this application, such as... Figure 6 As shown, the first fused feature output by the large-scale object target in the shallow feature network and the second fused feature output by the small-scale object target in the shallow feature network are decoupled using a gate function. Therefore, the expression for the multi-scale feature decoupling network can be written as:

[0099]

[0100]

[0101] in, This represents the mid-level fusion feature output from the target feature pyramid, where S is a feature containing information about large-scale objects. and These represent the average pooling set and the max pooling set, respectively. Indicates hyperparameters, This represents the Hadamard product, and Gate represents the gate that restricts large-scale information. This represents the function for obtaining the spatial mask. The following function can be used to obtain a feature map composed of features of small-scale objects, and its expression is as follows:

[0102]

[0103] in, This represents the shallow fused features output by the target feature pyramid. This represents a shallow fusion feature that contains only features of small-scale objects and , This represents a function that obtains the characteristics of small-sized objects.

[0104] In addition, according to Figure 5 The schematic diagram of the invalid information filtering policy network shown can be used to express the saliency features of the input invalid information filtering policy network as follows:

[0105]

[0106]

[0107]

[0108] in, This represents a learnable weight tensor. Since the background of a scene can be considered as its boundary pixels, edge values ​​can be initialized to 1, and values ​​in other regions to 0. Therefore, in this embodiment, the weight tensor is... The edge value is initialized to 1, and the value in other areas is 0. This represents an average aggregation operation to obtain background information. Indicates hyperparameters, Indicates the input feature map, Indicates image background information, express and The difference, The membership function represents the input feature map. Each element belongs to the foreground member level. Furthermore, prominent salience features are achieved through... Therefore, the final feature map can be represented as:

[0109]

[0110] in, This indicates element-wise summation.

[0111] As an optional implementation, in the technical solution provided in step S4 above, the method may include the following steps S41-S43, wherein:

[0112] Step S41: For each second remote sensing image, input the second feature images of the second remote sensing image at multiple different levels into the initial remote sensing target detection model to obtain the second multi-scale target output by the initial remote sensing target detection model;

[0113] Step S42: Construct a target loss function based on the second remote sensing target and the second multi-scale target in the sample image set, wherein the target loss function is used to reflect the difference between the second remote sensing target and the second multi-scale target;

[0114] Step S43: Adjust the model parameters of the initial remote sensing target detection model based on the target loss function to obtain the remote sensing target detection model.

[0115] Further, for the scheme provided in step S41, the method may include: using a target feature pyramid network to perform multi-scale feature fusion processing on multiple second feature images at different levels, respectively obtaining a second shallow fused feature image corresponding to the second shallow feature image, a second middle fused feature image corresponding to the second middle feature image, and a second deep fused feature image corresponding to the second deep feature image; inputting the second shallow fused feature image into a multi-scale feature decoupling network to obtain a third shallow fused feature image, and inputting the second middle fused feature image and the second deep fused feature image into an adaptive multi-scale information feature network to obtain a third middle fused feature image and a third deep fused feature image; and combining the third shallow fused feature image, the third middle fused feature image, and the third deep fused feature image... The fused feature images are input into the corresponding filtering invalid information strategy network to obtain the first image information corresponding to the second shallow fused feature image, the second image information corresponding to the second middle fused feature image, and the third image information corresponding to the second deep fused feature image. The first, second, and third image information include foreground information or background information. The first, second, and third image information are input into the second detection unit at the corresponding level to obtain the shallow feature detection results, middle feature detection results, and deep feature detection results output by the second detection unit at different levels. Based on the shallow, middle, and deep feature detection results output by the second detection unit at different levels, the second multi-scale target of the second remote sensing image is determined.

[0116] Specifically, such as Figure 3As shown, firstly, the second shallow feature image, the second middle feature image, and the second deep feature image output from the backbone network are input into the target feature pyramid that fuses nonlocal perception and attention mechanisms, respectively, to obtain the second shallow fused feature image output from the shallow feature network, the second middle fused feature image output from the middle feature network, and the second deep fused feature image output from the deep feature network. Next, the second shallow fused feature image is input into a multi-scale feature decoupling network to obtain the third shallow fused feature image, and the second middle fused feature image and the second deep fused feature image are input into the corresponding attention mechanisms at the fusion locations. In the adaptive multi-scale information feature network, the corresponding third middle-layer fusion feature image and third deep-layer fusion feature image are obtained respectively. Then, the obtained third shallow-layer fusion feature image, third middle-layer fusion feature image and third deep-layer fusion feature image are respectively input into the filtering invalid information strategy network to determine whether the image information of each fusion feature image is a foreground image or a background image, and obtain their respective image information. The obtained image information is respectively input into the second detection unit (i.e., the detection head) to determine the detection results in the fusion feature images of different levels, and based on the detection results, the multi-scale targets of the remote sensing image input into the remote sensing target detection model are determined.

[0117] Furthermore, after obtaining the second multi-scale target output by the initial remote sensing target detection model, the model parameters of the initial remote sensing target detection model can be adjusted according to the difference between the second multi-scale target and the second remote sensing targets in each second remote sensing image in the sample image set, so as to obtain the trained remote sensing target detection model.

[0118] To determine the detection accuracy of the remote sensing target detection model provided in this application embodiment, the detection performance of the remote sensing target detection model in this application embodiment will be verified in NWPU VHR-10 and DIOR respectively. The DIOR dataset includes 23,463 images and 190,288 object instances, and contains 20 categories, including aircraft (AL), airport (AT), baseball field (BF), basketball court (BC), bridge (B), chimney (C), dam (D), highway service area (ESA), highway toll station (ETS), golf course (GC), ground runway (GTF), port (HB), overpass (O), ship (S), stadium (SD), storage tank (ST), tennis court (TC), railway station (TS), vehicle (V), and windmill (WM). Therefore, the dataset can be divided into training set (5,862 images), validation set (5,863 images), and test set (11,738 images). The NWPU VHR-10 dataset contains a total of 800 images, of which 650 contain objects and the remaining 150 do not. It includes 10 object categories: airplane (AP), baseball field (BD), basketball court (BC), bridge (BG), ground runway (GTF), port (HB), ship (SP), storage tank (ST), tennis court (TC), and vehicle (VC). Since this dataset contains 800 images, with 650 containing objects and 150 not, it can be divided into an 80% training set and a 20% test set.

[0119] Table 1 below compares the detection results of the remote sensing target detection model (CSGNDA-Net) and related remote sensing target detection methods using the embodiments of this application on the DIOR dataset and the NWPU VHR-10 dataset. The detection performance is analyzed using two parameters: mAP (mean Average Precision) and FPS (Frames Per Second).

[0120] Table 1

[0121]

[0122] As can be seen from Table 1 above, the remote sensing target detection model of this application can ensure an mAP of up to 71.6% and an inference speed of 12 frames / second, confirming the superiority of CSGNDA-Net. Figure 7 This is a schematic diagram of an optional remote sensing image target detection result according to an embodiment of this application. It can be seen that the CSGNDA-Net used in this embodiment of the application can perform complete and accurate detection of aircraft and ships.

[0123] Example 2

[0124] Based on Embodiment 1 of this application, an embodiment of a remote sensing image target detection device is also provided, which executes the remote sensing image target detection method described in the above embodiment when it is in operation. Figure 8 This is a schematic diagram of the structure of an optional remote sensing image target detection device according to an embodiment of this application, such as... Figure 8 As shown, the remote sensing image target detection device includes at least an acquisition module 81, a feature extraction module 82, a feature processing module 83, and a target detection module 84, wherein:

[0125] Acquisition module 81 is used to acquire the first remote sensing image to be detected;

[0126] Feature extraction module 82 is used to extract features from the first remote sensing image using a preset feature extraction network to obtain the first feature image of the first remote sensing image at multiple different levels;

[0127] The feature processing module 83 is used to perform feature processing on multiple first feature images at different levels through the first feature processing unit of the pre-trained remote sensing target detection model to obtain multiple first fused feature images at different levels. The first feature processing unit includes: a target feature pyramid network based on non-local perception and attention mechanism for multi-scale feature fusion processing of multiple first feature images at different levels; a multi-scale feature decoupling network for decoupling the first shallow fused feature image output by the target feature pyramid network; an adaptive multi-scale information feature network for adjusting the ratio of multi-scale features between the first middle-layer fused feature image and the first deep fused feature image output by the target feature pyramid network; and an invalid information filtering strategy network for separating image information from the fused feature image output by the multi-scale feature decoupling network and the adaptive multi-scale information feature network.

[0128] The target detection module 84 is used to detect multiple first fusion feature images at different levels through the first detection unit of the remote sensing target detection model, obtain the detection results corresponding to the first fusion feature images at different levels, and determine the first multi-scale target of the first remote sensing image based on the detection results corresponding to different levels.

[0129] It should be noted that each module in the above-mentioned remote sensing image target detection device can be a program module (for example, a set of program instructions to implement a certain function) or a hardware module. For the latter, it can be manifested in the following forms, but is not limited to them: each of the above modules is manifested as a processor, or the functions of each of the above modules are implemented by a processor.

[0130] Example 3

[0131] According to an embodiment of this application, a non-volatile storage medium is also provided, which stores a program, wherein when the program runs, it controls the device where the non-volatile storage medium is located to execute the remote sensing image target detection method in Embodiment 1.

[0132] Optionally, the device containing the non-volatile storage medium executes the following steps by running the program: acquiring a first remote sensing image to be detected; extracting features from the first remote sensing image using a preset feature extraction network to obtain first feature images of the first remote sensing image at multiple different levels; performing feature processing on the first feature images at multiple different levels through a first feature processing unit of a pre-trained remote sensing target detection model to obtain first fused feature images at multiple different levels, wherein the first feature processing unit includes: a target feature pyramid network based on non-local perception and attention mechanism for multi-scale feature fusion processing of the first feature images at multiple different levels, and processing the output of the target feature pyramid network... A multi-scale feature decoupling network for decoupling the first shallow fused feature image; an adaptive multi-scale information feature network for adjusting the multi-scale feature ratio of the first mid-layer fused feature image and the first deep fused feature image output by the target feature pyramid network; and a filtering invalid information strategy network for separating image information from the optimized fused feature image output by the multi-scale feature decoupling network and the adaptive allocation multi-scale feature network. The first detection unit of the remote sensing target detection model detects the first fused feature images at multiple different levels to obtain detection results corresponding to the first fused feature images at multiple different levels, and determines the first multi-scale target of the first remote sensing image based on the detection results corresponding to different levels.

[0133] According to an embodiment of this application, a processor is also provided for running a program, wherein the program executes the remote sensing image target detection method in Embodiment 1 during runtime.

[0134] Optionally, the program executes the following steps during runtime: acquiring a first remote sensing image to be detected; extracting features from the first remote sensing image using a preset feature extraction network to obtain first feature images of the first remote sensing image at multiple different levels; and performing feature processing on the first feature images at multiple different levels through a first feature processing unit of a pre-trained remote sensing target detection model to obtain first fused feature images at multiple different levels. The first feature processing unit includes: a target feature pyramid network based on non-local perception and attention mechanisms for multi-scale feature fusion processing of the first feature images at multiple different levels; and a first shallow fused feature image output by the target feature pyramid network. The system comprises a multi-scale feature decoupling network for decoupling feature images, an adaptive multi-scale information feature network for adjusting the multi-scale feature ratio of the first mid-level fused feature image and the first deep fused feature image output by the target feature pyramid network, and a filtering invalid information strategy network for separating image information from the optimized fused feature image output by the multi-scale feature decoupling network and the adaptive allocation multi-scale feature network. The first detection unit of the remote sensing target detection model detects multiple first fused feature images at different levels, obtaining detection results corresponding to multiple first fused feature images at different levels, and determining the first multi-scale target of the first remote sensing image based on the detection results corresponding to different levels.

[0135] According to an embodiment of this application, an electronic device is also provided, which includes one or more processors; a memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors are configured to run the programs, wherein the programs are configured to execute the remote sensing image target detection method in Embodiment 1 above.

[0136] Optionally, the processor is configured to execute the following steps via a computer program: acquiring a first remote sensing image to be detected; extracting features from the first remote sensing image using a preset feature extraction network to obtain first feature images of the first remote sensing image at multiple different levels; performing feature processing on the first feature images at multiple different levels through a first feature processing unit of a pre-trained remote sensing target detection model to obtain first fused feature images at multiple different levels, wherein the first feature processing unit includes: a target feature pyramid network based on non-local perception and attention mechanism for multi-scale feature fusion processing of the first feature images at multiple different levels, and a first shallow depth sensor output by the target feature pyramid network. The system comprises a multi-scale feature decoupling network for decoupling layer fusion feature images, an adaptive multi-scale information feature network for adjusting the multi-scale feature ratio of the first mid-layer fusion feature image and the first deep fusion feature image output by the target feature pyramid network, and a filtering invalid information strategy network for separating image information from the optimized fusion feature image output by the multi-scale feature decoupling network and the adaptive allocation multi-scale feature network. The first detection unit of the remote sensing target detection model detects multiple first fusion feature images at different levels, obtaining detection results corresponding to multiple first fusion feature images at different levels, and determining the first multi-scale target of the first remote sensing image based on the detection results corresponding to different levels.

[0137] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0138] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0139] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some interfaces; indirect couplings or communication connections between units or modules may be electrical or other forms.

[0140] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0141] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0142] If the integrated unit is implemented as 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 this application, in essence, or the part that contributes to related technologies, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard drive, magnetic disk, or optical disk.

[0143] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A remote sensing image target detection method, characterized in that, include: Acquire the first remote sensing image to be detected; The first remote sensing image is subjected to feature extraction using a preset feature extraction network to obtain first feature images of the first remote sensing image at multiple different levels; The first feature processing unit of the pre-trained remote sensing target detection model performs feature processing on multiple first feature images at different levels to obtain multiple first fused feature images at different levels. The first feature processing unit includes: a target feature pyramid network based on non-local perception and attention mechanism for multi-scale feature fusion processing of multiple first feature images at different levels; a multi-scale feature decoupling network for decoupling the first shallow fused feature image output by the target feature pyramid network; an adaptive multi-scale information feature network for adjusting the multi-scale feature ratio of the first mid-level fused feature image and the first deep fused feature image output by the target feature pyramid network; and an invalid information filtering strategy network for separating image information from the optimized fused feature image output by the multi-scale feature decoupling network and the adaptive multi-scale information feature network. The first detection unit of the remote sensing target detection model detects multiple first fused feature images at different levels to obtain detection results corresponding to the first fused feature images at multiple different levels, and determines the first multi-scale target of the first remote sensing image based on the detection results corresponding to different levels. The multi-scale feature decoupling network is used to decouple the first shallow-layer fused feature image output by the target feature pyramid network using a gate function, and the output expression of the multi-scale feature decoupling network is: ； In the formula, This represents the first shallow fused feature image output by the target feature pyramid network. This represents the first shallow fused feature image output by the target feature pyramid network, containing only small-scale objects. C represents the number of channels, W represents the width, and H represents the height. A function representing the features used to obtain information about small objects. Where S represents a feature containing information about large-scale objects, and , This represents the first middle-layer fused feature image output from the target feature pyramid. and These represent the average pooling set and the max pooling set, respectively. Indicates hyperparameters, This represents the Hadamard product, and Gate represents the gate that restricts large-scale information. This represents a function to obtain the spatial mask; The adaptive multi-scale information feature network is used to determine the positional attention weights at different locations within the first mid-level fused feature image and the first deep fused feature image output by the target feature pyramid network, and the expression for the positional attention weights is: ； In the formula, This represents the L2 normalization function. Sign Represents the time-time function. This represents the k-th layer original feature image output by the target feature pyramid network. The feature image obtained after preprocessing, and The location attention weights are used to adaptively assign multi-scale information features to the first mid-layer fusion feature image and the first deep fusion feature image output by the target feature pyramid network.

2. The method of claim 1, wherein, The first remote sensing image is subjected to feature extraction using a preset feature extraction network to obtain first feature images of the first remote sensing image at multiple different levels, including: The first remote sensing image is preprocessed; The feature extraction network is used to extract features from the processed first remote sensing image to obtain first feature images of the first remote sensing image at multiple different levels, wherein the first feature images include: a first shallow feature image, a first medium feature image, and a first deep feature image.

3. The method of claim 1, wherein, The training process of the remote sensing target detection model includes: Obtain a sample image set, wherein the sample image set includes multiple second remote sensing images, and each second remote sensing image includes at least one second remote sensing target; The feature extraction network is used to extract features from each of the second remote sensing images in the sample image set to obtain second feature images of each second remote sensing image at multiple different levels. The second feature images include: a second shallow feature image, a second middle feature image, and a second deep feature image. An initial remote sensing target detection model is constructed, wherein the initial remote sensing target detection model includes: a second feature processing unit and a second detection unit; The initial remote sensing target detection model is trained based on the second feature images at multiple different levels of each of the second remote sensing images in the sample image set to obtain the remote sensing target detection model.

4. The method of claim 3, wherein, The construction process of the second feature processing unit includes: Construct the target feature pyramid network that integrates nonlocal perception and dual attention mechanisms, wherein the target feature pyramid network includes: a shallow feature network, a medium feature network, and a deep feature network; Construct the adaptive multi-scale information feature network that integrates the position-aware attention mechanism; The multi-scale feature decoupling network is constructed using a gate function, and the invalid information filtering strategy network is constructed using a membership function.

5. The method of claim 3, wherein, The initial remote sensing target detection model is trained based on second feature images at multiple different levels for each of the second remote sensing images in the sample image set to obtain the remote sensing target detection model, including: For each of the second remote sensing images, the second feature images of the second remote sensing image at multiple different levels are input into the initial remote sensing target detection model to obtain the second multi-scale target output by the initial remote sensing target detection model; A target loss function is constructed based on the second remote sensing target and the second multi-scale target in the sample image set, wherein the target loss function is used to reflect the difference between the second remote sensing target and the second multi-scale target; The model parameters of the initial remote sensing target detection model are adjusted based on the target loss function to obtain the remote sensing target detection model.

6. The method of claim 3, wherein, The second feature images at multiple different levels of the second remote sensing image are input into the initial remote sensing target detection model to obtain the second multi-scale target output by the initial remote sensing target detection model, including: The target feature pyramid network is used to perform multi-scale feature fusion processing on the second feature images at multiple different levels to obtain the second shallow fused feature image corresponding to the second shallow feature image, the second middle fused feature image corresponding to the second middle feature image, and the second deep fused feature image corresponding to the second deep feature image. The second shallow fusion feature image is input into the multi-scale feature decoupling network to obtain the third shallow fusion feature image. The second middle fusion feature image and the second deep fusion feature image are input into the adaptive multi-scale information feature network to obtain the third middle fusion feature image and the third deep fusion feature image. The third shallow fusion feature image, the third middle fusion feature image, and the third deep fusion feature image are respectively input into the corresponding filtering invalid information policy network to obtain the first image information corresponding to the second shallow fusion feature image, the second image information corresponding to the second middle fusion feature image, and the third image information corresponding to the second deep fusion feature image. The first image information, the second image information, and the third image information include foreground information or background information. The first image information, the second image information, and the third image information are respectively input to the second detection unit at the corresponding level to obtain shallow feature detection results, middle feature detection results, and deep feature detection results output by the second detection unit at different levels; The second multi-scale target in the second remote sensing image is determined based on the shallow feature detection results, the middle feature detection results, and the deep feature detection results output by the second detection unit at different levels.

7. The method according to claim 1, characterized in that, The feature extraction network includes at least one of the following: LeNet-5 network, AlexNet network, VGGNet network, GoogLeNet network, and ResNet network.

8. A remote sensing image target detection apparatus, characterized by comprising: include: The acquisition module is used to acquire the first remote sensing image to be detected; The feature extraction module is used to extract features from the first remote sensing image using a preset feature extraction network to obtain the first feature image of the first remote sensing image at multiple different levels. The feature processing module is used to perform feature processing on multiple first feature images at different levels through the first feature processing unit of the pre-trained remote sensing target detection model to obtain multiple first fused feature images at different levels. The first feature processing unit includes: a target feature pyramid network based on non-local perception and attention mechanism for multi-scale feature fusion processing of multiple first feature images at different levels; a multi-scale feature decoupling network for decoupling the first shallow fused feature image output by the target feature pyramid network; an adaptive multi-scale information feature network for adjusting the ratio of multi-scale features in the first middle-layer fused feature image and the first deep fused feature image output by the target feature pyramid network; and an invalid information filtering strategy network for separating image information in the fused feature image output by the multi-scale feature decoupling network and the adaptive multi-scale information feature network. The target detection module is used to detect multiple first fused feature images at different levels through the first detection unit of the remote sensing target detection model, obtain detection results corresponding to multiple first fused feature images at different levels, and determine the first multi-scale target of the first remote sensing image based on the detection results corresponding to different levels. The multi-scale feature decoupling network is used to decouple the first shallow-layer fused feature image output by the target feature pyramid network using a gate function, and the output expression of the multi-scale feature decoupling network is: ； In the formula, This represents the first shallow fused feature image output by the target feature pyramid network. This represents the first shallow fused feature image output by the target feature pyramid network, containing only small-scale objects. C represents the number of channels, W represents the width, and H represents the height. A function representing the features used to obtain information about small-sized objects. Where S represents a feature containing information about large-scale objects, and , This represents the first middle-layer fused feature image output from the target feature pyramid. and These represent the average pooling set and the max pooling set, respectively. Indicates hyperparameters, This represents the Hadamard product, and Gate represents the gate that restricts large-scale information. This represents a function to obtain the spatial mask; The adaptive multi-scale information feature network is used to determine the positional attention weights at different locations within the first mid-level fused feature image and the first deep fused feature image output by the target feature pyramid network, and the expression for the positional attention weights is: ； In the formula, This represents the L2 normalization function. Sign Represents the time-time function. This represents the k-th layer original feature image output by the target feature pyramid network. The feature image obtained after preprocessing, and The location attention weights are used to adaptively assign multi-scale information features to the first mid-layer fusion feature image and the first deep fusion feature image output by the target feature pyramid network.

9. A non-volatile storage medium, characterized by The non-volatile storage medium stores a computer program, wherein the device containing the non-volatile storage medium executes the remote sensing image target detection method according to any one of claims 1 to 7 by running the computer program.

10. An electronic device, characterized in that, include: A memory and a processor, the processor being configured to run a program stored in the memory, wherein the program, when running, executes the remote sensing image target detection method according to any one of claims 1 to 7.

Images