Method and system for directional target detection in remote sensing images based on rotation feature perception
By integrating a rotation feature perception module into the feature fusion network, the model enhances sensitivity to rotational invariant features, addressing the insufficiency in current models and improving target detection accuracy in remote sensing images.
Patent Information
- Application Number
- CN202410484441.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-04-22
- Publication Date
- 2024-12-20
- Estimated Expiration
- 2044-04-22
AI Technical Summary
Current convolutional neural network-based remote sensing image target detection models lack sensitivity to rotation invariance characteristics, leading to insufficient perception of rotational invariant features, which hampers accurate detection and localization of oriented targets.
Integrate a rotation feature perception module into the feature fusion network, comprising modules like normalization, multi-layer perceptron, depth separation convolution, group convolution, self-attention, and feature reshaping, to enhance the model's sensitivity to rotational invariant features.
The enhanced model effectively extracts and fuses rotational invariant features, improving the accuracy of target detection and localization in remote sensing images, enabling reliable and efficient detection tasks.
Smart Images
Figure CN118298157B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of image processing, and in particular to a method and system for detecting directional targets in remote sensing images based on rotation feature perception. Background Art
[0002] At present, remote sensing images have broad application prospects in the fields of national defense security, environmental monitoring, urban planning, etc. Remote sensing image target detection technology has always been one of the research focuses in the field of remote sensing. Remote sensing images have the characteristics of top-down view, arbitrary rotation angle of the target, and arbitrary direction, which makes directional target detection challenging. Rotation invariance means that the semantic information of the foreground remains unchanged during the image rotation process. The refined processing of rotation invariance can better solve the problem of directional target detection in remote sensing images. The current mainstream convolutional neural network-based target detection model is not sensitive enough to the rotation invariance characteristics of the target. Generally, the model is guided to capture the rotation invariance characteristics of the directional target through the design of symmetric convolution, pooling, loss function, etc., but the effect is often weak. This is because the convolution operator does not have the natural ability to extract rotation transformation features and cannot effectively process the rotation features of the target. In addition, due to the limitation of the size of the convolution kernel, local features can only interact with each other in a relatively narrow receptive field, and the ability to perceive the spatial structure information of the features is poor. Transformer has the ability to perceive the direction of the object. It uses position encoding, that is, the self-attention mechanism to capture the spatial structure information features to remain sensitive to the rotation invariance characteristics of the target.
[0003] Semantic features are of great significance in the field of remote sensing image target detection. They contain high-level abstract information related to objects and scenes in the image. This information covers the category, shape, structure of the objects, and their position and relationship in the environment. They are abstract manifestations of rich rotation invariant features. The basic feature fusion network in the remote sensing image target detection model usually adopts upsampling and feature fusion, ignoring the effective aggregation of rotation invariant information in high-level semantic features. The rotation invariant feature perception ability in the remote sensing image directional target detection model is insufficient, and the quality of semantic features in the feature fusion stage needs to be improved.
[0004] Therefore, we consider introducing Transformer into the feature fusion network and designing a cross-scale semantic feature refinement processing module to guide the model to be sensitive to the rotation invariant features of oriented targets, so as to correctly identify and locate potential target objects in remote sensing images. Summary of the invention
[0005] In view of this, the present invention aims at the problem of insufficient perception ability of rotational invariant features in remote sensing image oriented target detection models, and provides a remote sensing image oriented target detection method and system based on rotational feature perception. By building a feature fusion network combined with a rotational feature perception module to promote the quality of semantic features in the feature fusion stage, the model is further guided to remain sensitive to the rotational invariant features of oriented targets, thereby effectively improving the performance of remote sensing image oriented target detection; it is conducive to achieving accurate remote sensing image oriented target detection task identification; and it provides a more reliable and efficient solution for the field of remote sensing image target detection.
[0006] To achieve the above object, the technical solution adopted by the present invention is:
[0007] In a first aspect, an embodiment of the present invention provides a method for detecting directional targets in remote sensing images based on rotation feature perception, the method comprising the following steps:
[0008] S1, build a rotation feature perception module;
[0009] S2. Building a feature fusion network and its corresponding remote sensing image directional target detection model in combination with the rotation feature perception module;
[0010] S3, performing target detection task training and testing on the remote sensing image directional target detection model;
[0011] S4. Use the trained remote sensing image directional target detection model to perform target detection on the remote sensing image to be tested.
[0012] Furthermore, in step S1, the rotation feature perception module constructed includes: a normalization module, a multi-layer perceptron module, a depth separation convolution module, a group convolution module, a hole convolution module, a self-attention module, a feature splicing module and a feature reshaping module; wherein:
[0013] The normalization module standardizes the input feature map; the multi-layer perceptron module extracts features from the normalized feature map; the deep separation convolution module, the grouped convolution module and the dilated convolution module are three independent branches that use different convolution operations to mine the feature information of the feature map under different receptive fields; the self-attention module extracts the rotation invariant features of different branches; the normalization module also processes the outputs of different branches to obtain feature maps of adapted shapes; the feature splicing module aggregates the feature maps of different branches along the channel dimension to obtain multi-dimensional rotational feature information; the feature reshaping module changes the shape of the spliced feature map to conform to the feature maps of other levels.
[0014] Furthermore, in the multi-layer perceptron module, the number of neurons in the input layer is the number of channels C of the input feature map, the number of neurons in the hidden layer is 5×C, and the number of neurons in the output layer is 2×C.
[0015] Furthermore, the calculation process of the self-attention module includes word vector embedding and Softmax calculation.
[0016] Furthermore, the feature reshaping module modifies the size of the feature map to match the feature maps of other levels through sampling / interpolation strategies.
[0017] Furthermore, in step S2, the remote sensing image directional target detection model constructed includes: a backbone feature extraction network, a feature fusion network based on the optimization and improvement of the rotation feature perception module, and a feature decoupling head.
[0018] Furthermore, the backbone feature extraction network is composed of a plurality of convolution modules stacked together, each of which includes: a basic convolution operator, a normalization module and an activation function, wherein:
[0019] The basic convolution operator calculates the channel information of the input remote sensing image at a pixel-by-pixel level; the normalization module normalizes the calculated channel results to the standard range; the activation function activates and inhibits the feature matrix of the potential target area.
[0020] Furthermore, in step S3, the focus loss and the rotation frame intersection-over-union loss are used to train the remote sensing image directional target detection model for target detection tasks.
[0021] In a second aspect, an embodiment of the present invention further provides a remote sensing image directional target detection system based on rotation feature perception, which applies the above-mentioned remote sensing image directional target detection method based on rotation feature perception to achieve accurate remote sensing image directional target detection task recognition, and the system includes:
[0022] A target detection model building module is used to construct a rotation feature perception module, and to build a feature fusion network and its corresponding remote sensing image directional target detection model in combination with the rotation feature perception module;
[0023] A target detection model training module is used to perform target detection task training and testing on the remote sensing image directional target detection model;
[0024] The target detection model application module is used to use the trained remote sensing image directional target detection model to perform target detection on the remote sensing image to be tested.
[0025] Furthermore, the constructed rotation feature perception module includes: a normalization module, a multi-layer perceptron module, a depth separation convolution module, a group convolution module, a hole convolution module, a self-attention module, a feature splicing module and a feature reshaping module; wherein:
[0026] The normalization module is used to standardize the input feature map; the multi-layer perceptron module is used to extract features from the normalized feature map; the deep separation convolution module, the grouped convolution module and the dilated convolution module are used as three independent branches to use different convolution operations to mine the feature information of the feature map under different receptive fields; the self-attention module is used to extract the rotation invariant features of different branches; the normalization module is also used to process the outputs of different branches to obtain feature maps of adapted shapes; the feature splicing module is used to aggregate the feature maps of different branches along the channel dimension to obtain multi-dimensional rotational feature information; the feature reshaping module is used to change the shape of the spliced feature map to conform to the feature maps of other levels.
[0027] Compared with the prior art, the present invention has at least the following beneficial effects:
[0028] 1. The present invention provides a method and system for remote sensing image directional target detection based on rotation feature perception, wherein a rotation feature perception module is constructed, and a feature fusion network optimized and improved in combination with the rotation feature perception module and its corresponding remote sensing image directional target detection model are constructed. The constructed rotation feature perception module can effectively extract the rotation invariant features in the high-level semantic information in the feature map, guide the model to be sensitive to the rotation invariant features of the directional target, facilitate the correct identification and positioning of potential target objects in the remote sensing image, and help the remote sensing image directional target detection model to more accurately understand the image content and correctly identify and locate potential targets in the remote sensing image.
[0029] 2. The feature fusion network optimized based on the rotation feature perception module in the present invention can fully fuse the rotation invariance information of semantic feature information at different levels in the feature fusion stage, effectively solving the problem of insufficient perception of rotation invariance information by current target detection methods, thereby realizing accurate recognition of remote sensing image directional target detection tasks.
[0030] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description and the accompanying drawings.
[0031] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS
[0032] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.
[0033] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.
[0034] Figure 1 A flowchart of a method for directional target detection in remote sensing images based on rotation feature perception is provided in an embodiment of the present invention.
[0035] Figure 2 A schematic diagram of the structure of a rotation feature perception module provided in an embodiment of the present invention.
[0036] Figure 3 A schematic diagram of the structure of a remote sensing image directional target detection model based on a rotation feature perception module provided in an embodiment of the present invention. DETAILED DESCRIPTION
[0037] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments.
[0038] In the description of the present invention, it should be noted that: in some processes described in the specification and drawings of this application, multiple operations appearing in a specific order are included, but it should be clearly understood that these operations may not be performed in the order in which they appear in this document or may be performed in parallel. In addition, various serial numbers are only used for descriptive purposes and cannot be understood as indicating or implying relative importance.
[0039] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0040] See also Figure 1 As shown, the present invention provides a method for detecting directional targets in remote sensing images based on rotation feature perception, and the method mainly includes the following steps:
[0041] S1, build a rotation feature perception module;
[0042] S2. Combine the rotation feature perception module to build a feature fusion network and its corresponding remote sensing image directional target detection model;
[0043] S3, training and testing the remote sensing image directional target detection model for target detection tasks;
[0044] S4. Use the trained remote sensing image directional target detection model to perform target detection on the remote sensing image to be tested.
[0045] The specific implementation methods of the method of the present invention are described in detail below:
[0046] In this embodiment, a feature fusion network based on a rotation feature perception module is designed to improve the quality of semantic features in the feature fusion stage, and the model is further guided to be sensitive to the rotation invariant features of oriented targets, so as to effectively improve the performance of oriented target detection in remote sensing images and achieve accurate recognition of oriented target detection tasks in remote sensing images. The specific implementation of this method is as follows:
[0047] 1. Build a rotation feature perception module
[0048] In this embodiment, the structure of the rotation feature perception module is as follows: Figure 2 As shown in the figure, its core modules include multi-layer perceptron module, normalization module, depth separation convolution module, group convolution module, hole convolution module, self-attention module, feature splicing module, feature reshaping module, etc.
[0049] The following is a detailed description of the working process of the rotation feature perception module in the order of module calculation:
[0050] First, the normalization module normalizes the input feature map, which makes it easier for the multi-layer perceptron module to extract features from the input feature map. The calculation process can be formally expressed as:
[0051]
[0052] in, is the input feature map, is the output feature map. represents the feature normalization operation, Represents a multilayer perceptron module, and the number of neurons in its input layer is Number of channels , the number of neurons in the hidden layer is 5×C, and the number of neurons in the output layer is 2×C.
[0053] Furthermore, different convolution operations of the three independent branches are used to mine feature information under different receptive fields, and the calculation process can be formally expressed as:
[0054]
[0055]
[0056]
[0057] in, Represents deep separation convolution, which aims to mine the correlation characteristics of feature maps in spatial and channel dimensions. Represents grouped convolution, which aims to mine the correlation features between different channel layers of the feature map. Represents a dilated convolution, which aims to mine the global features of the feature map from the perspective of a larger convolution receptive field. Represents the feature maps obtained after calculation of different branches.
[0058] Furthermore, attention calculation is introduced to extract rotation invariant features of different branches. The calculation process can be formally expressed as:
[0059]
[0060] in, Represents the self-attention module, whose calculation process includes word vector embedding and Softmax calculation, etc., with the feature map of the i-th layer Taking the feature map as input as an example, the calculation process can be formally expressed as:
[0061]
[0062] in, Represents the word vector embedding process, which transforms the feature map into Expand to The size is 1×1, and the number of channels is The word vector .
[0063] Furthermore, the self-attention module will The calculation process can be formally expressed as:
[0064]
[0065] Among them, Q represents the query matrix, K represents the key-value matrix, and V represents the value matrix. The three are composed of the input word vector Obtained through linear mapping. Represents the dimension of its feature map. The self-attention mechanism maps the input vector to the probability of a given output word vector to obtain a prediction of the content of each sequence.
[0066] Furthermore, the feature normalization module is used to process the outputs of different branches to obtain feature maps of the adapted shape. The calculation process can be formally expressed as:
[0067]
[0068] in, Represents the feature map after feature normalization.
[0069] Furthermore, the feature concatenation operation of the feature concatenation module is used to aggregate feature information of different branches. The calculation process can be formally expressed as:
[0070]
[0071] in, Represents the feature concatenation operation, which aggregates the feature maps of different branches along the channel dimension to obtain multi-dimensional rotation feature information. Represents the feature map obtained after splicing, and its number of channels is .
[0072] Furthermore, in order to facilitate the fusion with feature maps of different levels, the feature reshaping operator of the feature reshaping module is used to change the shape of the feature map to adapt to the feature maps across other levels. The calculation process can be formally expressed as:
[0073]
[0074] in, Represents the feature reshaping operator, which modifies the size of the feature map to match the feature maps of other levels through sampling / interpolation strategies.
[0075] At this point, the rotation feature perception module has been built.
[0076] 2. Build a feature fusion network optimized and improved by the rotation feature perception module and its corresponding remote sensing image directional target detection model:
[0077] Most traditional convolutional neural network-based target detection models use a backbone feature extraction network-feature fusion network-feature decoupling head structure as the basic model architecture. The remote sensing image directional target detection model based on the rotation feature perception module constructed in the present invention follows this paradigm and introduces the rotation feature perception module constructed in step 1 to optimize the feature fusion network to achieve the optimization of multi-scale semantic feature information, and further guide the model to be sensitive to the rotation invariant features of the directional target. The structure of the remote sensing image directional target detection model is as follows: Figure 3 The following will introduce in detail the construction method of the remote sensing image directional target detection model based on the rotation feature perception module of the present invention.
[0078] In this embodiment, first, the remote sensing image and its label information are input into the model through the data loader. The calculation process of the data loader is expressed as:
[0079]
[0080] in, Indicates the size ,Input remote sensing image with 3 channels; Annotation information for the target in the image, including location and category information; It is a data loader function used to load data and enhance data;
[0081] Then the lightweight backbone feature extraction network extracts features from imgs. The calculation process can be formally expressed as:
[0082]
[0083] in, Represents the feature map obtained after feature extraction, with a size of , and there are channels; It is a backbone feature extraction network, which obtains the feature information of the input remote sensing image by stacking convolution modules in layers. Typical backbone feature extraction networks include: ResNet, VGG, EfficientNet, etc. The method shown in the embodiment of the present invention takes the basic convolution module as an example, but it should be noted that through simple model replacement, the remote sensing image directional target detection model based on rotation feature perception of the present invention can be adapted to the general basic backbone feature extraction network.
[0084] In a specific embodiment, the backbone feature extraction network It is composed of stacked convolution modules. Each basic convolution module contains a basic convolution operator (Conv), a normalization module (Batch Normalization) and an activation function (Relu). The basic convolution operator calculates the channel information of the input remote sensing image at a pixel-by-pixel level to achieve the purpose of feature extraction. The normalization module normalizes the calculated channel results to the standard range to facilitate subsequent calculations while avoiding abnormal gradient optimization problems. The activation function activates and inhibits the feature matrix of the potential target area to help the model distinguish between foreground and background information. The calculation process of the operation layer of each module can be formally expressed as:
[0085]
[0086] in, represents the basic convolution operation, is a normalization module, which keeps the number of channels it receives and outputs features unchanged. To input the original remote sensing image, It is the output feature map obtained after feature extraction. The activation function is responsible for activating the potential target area of the output feature map. Definition It is a convolution module that includes the basic convolution operator, normalization module and activation function. Therefore, the deep feature extraction process of the deep neural network can be formally expressed as:
[0087]
[0088] in, represents the convolutional module of the nth layer, is the input image, is the output feature map. Based on this, we can obtain n feature maps of remote sensing images after deep feature extraction.
[0089] In this embodiment, the feature fusion network is responsible for receiving the feature graph of the backbone network for fusion to obtain multi-scale feature information of the potential target. Extracted feature map Pass it to the feature fusion network for feature fusion; the calculation process of the feature fusion network is expressed as:
[0090]
[0091] in, It is a feature fusion network used for feature fusion. It first receives the feature map of the backbone feature extraction network. Then, the rotation feature perception module in step 1 is used to optimize the quality of feature fusion, and the shape of the feature map is further changed through the convolution operator to achieve fusion to obtain the fused feature map , the calculation process of the feature fusion network is expressed as:
[0092]
[0093]
[0094]
[0095] in, represents the feature concatenation operator, represents the convolution operator, RFC represents the rotation feature perception module, Represents the feature reshaping module.
[0096] Finally, the fused feature map is decoupled using the feature decoupling head (detection head). The feature decoupling head network consists of a basic convolution module and a prediction function. The convolution module is used to receive the input remote sensing image features and process them in appropriate shapes and channel dimensions for subsequent calculations. The prediction function is used to predict the location of the potential target area of the feature map, i.e., the category information, and output the prediction result. The calculation process of the detection head is expressed as:
[0097]
[0098]
[0099] in, Classification information obtained for prediction; Location information obtained for prediction; The feature map output by the feature fusion network; is the classification function, is the rectangular box prediction function.
[0100] 3. Training and testing of target detection tasks
[0101] In one embodiment, after building a remote sensing image directional target detection model based on a rotation feature perception module, the network is trained for the remote sensing image directional target detection task using focal loss and rotated IoU loss, which is formally expressed as:
[0102]
[0103] in, Represents focal loss, which is used in the classification prediction branch of remote sensing image object detection task. Represents the rotation box IoU loss function, which is used to calculate the distance between the predicted box and the target true value label corresponding to the input remote sensing image. It indicates the total loss used in network training. When the network is trained until the loss no longer decreases, it means that the network training has reached stability, the training process is over, and the trained remote sensing image directional target detection model combined with the rotation feature perception module is obtained.
[0104] Furthermore, the trained remote sensing image directional target detection model combined with the rotation feature perception module is used to test the remote sensing image to be tested, which is formally expressed as:
[0105]
[0106] in, represents the trained remote sensing image directional target detection model, and They respectively represent the remote sensing images to be tested and the corresponding target detection results.
[0107] From the description of the above embodiments, those skilled in the art can know that in order to solve the problem of insufficient perception ability of rotational invariant features in the remote sensing image directional target detection model, the present invention proposes an innovative and efficient remote sensing image directional target detection method. The method adopts a strategy based on rotational feature perception, designs an efficient rotational feature perception module, and integrates it into the feature fusion network; the rotational feature perception module proposed in the present invention can effectively extract the rotational invariant features in the high-level semantic information in the feature map, and guide the model to remain sensitive to the rotational invariant features of the directional target, which is crucial for correctly identifying and locating potential target objects in remote sensing images. Help the target detection model to understand the image content more accurately, and correctly identify and locate potential targets in remote sensing images. The feature fusion network based on the rotational feature perception module proposed in the present invention can fully fuse the rotational invariant information of semantic feature information at different levels in the feature fusion stage, effectively solving the problem of insufficient perception of rotational invariant information in the current target detection method, thereby realizing accurate recognition of remote sensing image directional target detection tasks.
[0108] The method of the present invention provides a more reliable and efficient solution for the field of directional target detection in remote sensing images, and provides strong support for future remote sensing application fields; it not only has broad application prospects in the fields of national defense security, environmental monitoring, urban planning, etc., but also provides a more feasible and accurate tool for remote sensing image interpretation, which is helpful to promote the further development of remote sensing technology in practical applications.
[0109] Furthermore, the present invention also provides a remote sensing image directional target detection system based on rotation feature perception, which is applied to a remote sensing image directional target detection method based on rotation feature perception in the above embodiment to improve the performance of image target detection and facilitate accurate remote sensing image directional target detection task recognition. The system includes:
[0110] The target detection model building module is used to build a rotation feature perception module, and to build a feature fusion network and its corresponding remote sensing image directional target detection model in combination with the rotation feature perception module;
[0111] The target detection model training module is used to train and test the remote sensing image directional target detection model for target detection tasks;
[0112] The target detection model application module is used to use the trained remote sensing image directional target detection model to perform target detection on the remote sensing image to be tested.
[0113] The implementation principle and technical effects of the system provided in the embodiment of the present invention are the same as those of the aforementioned method embodiment. For the sake of brief description, for matters not mentioned in the system embodiment, reference can be made to the corresponding contents in the aforementioned method embodiment, which will not be repeated here.
[0114] In addition, an embodiment of the present invention also provides a storage medium on which one or more programs readable by a computing device are stored, and the one or more programs include instructions. When the instructions are executed by the computing device, the computing device executes a remote sensing image directional target detection method based on rotation feature perception in the above-mentioned embodiment.
[0115] In the embodiment of the present invention, the storage medium may be, for example, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination thereof. More specific examples of storage media (a non-exhaustive list) include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disk read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical encoding device, and any suitable combination thereof.
[0116] Those skilled in the art should understand that the embodiments of the present invention can be provided as methods, systems or computer program products, etc. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects.
[0117] It should be noted that the word "comprising" does not exclude the presence of components or steps not listed in the claims. The word "a" or "an" preceding a component does not exclude the presence of a plurality of such components. The invention can be implemented by means of hardware comprising several distinct components, and by means of a suitably programmed computer.
[0118] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.
[0119] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method for directional target detection in remote sensing images based on rotation feature perception. It is characterized in that The method comprises the following steps: S1, build a rotation feature perception module; S2. Build a feature fusion network optimized and improved by the rotation feature perception module and its corresponding remote sensing image directional target detection model; S3, performing target detection task training and testing on the remote sensing image directional target detection model; S4, using the trained remote sensing image directional target detection model to perform target detection on the remote sensing image to be tested; In step S1, the rotation feature perception module constructed includes: a normalization module, a multi-layer perceptron module, a depth separation convolution module, a group convolution module, a hole convolution module, a self-attention module, a feature normalization module, a feature splicing module and a feature reshaping module; wherein: The normalization module normalizes the input feature map; The multi-layer perceptron module extracts features from the normalized feature map; The deep separation convolution module, grouped convolution module and dilated convolution module are three independent branches that use different convolution operations to mine the feature information of feature maps under different receptive fields. The self-attention module extracts rotation invariant features of different branches; The feature normalization module processes the outputs of different branches to obtain the feature map of the adapted shape. The calculation process can be formally expressed as: iAA =BN(F iA_qkv ), where F iAA ∈R H×W×C represents the feature map after feature normalization, F iA_qkv Represents the result of self-attention calculation; The feature concatenation module aggregates feature maps of different branches along the channel dimension to obtain multi-dimensional rotation feature information; The feature reshaping module changes the size of the concatenated feature map through sampling / interpolation strategies to match the feature maps of other levels.
2. According to claim 1, a method for detecting directional targets in remote sensing images based on rotation feature perception, It is characterized in that In the multi-layer perceptron module, the number of neurons in the input layer is the number of channels C of the input feature map, the number of neurons in the hidden layer is 5×C, and the number of neurons in the output layer is 2×C.
3. According to claim 1, a method for detecting directional targets in remote sensing images based on rotation feature perception, It is characterized in that The calculation process of the self-attention module includes word vector embedding and Softmax calculation.
4. According to claim 1, a method for detecting directional targets in remote sensing images based on rotation feature perception, It is characterized in that In step S2, the remote sensing image directional target detection model constructed includes: a backbone feature extraction network, a feature fusion network based on the optimization and improvement of the rotation feature perception module, and a feature decoupling head.
5. According to claim 4, a method for detecting directional targets in remote sensing images based on rotation feature perception, It is characterized in that The backbone feature extraction network is composed of a plurality of stacked convolution modules, each of which includes: a basic convolution operator, a normalization module and an activation function, wherein: The basic convolution operator calculates the channel information of the input remote sensing image at a pixel-by-pixel level; the normalization module normalizes the calculated channel results to the standard range; the activation function activates and inhibits the feature matrix of the potential target area.
6. The method for detecting directional targets in remote sensing images based on rotational feature perception according to claim 1, It is characterized in that In step S3, the focus loss and the rotation frame intersection-over-union loss are used to train the remote sensing image directional target detection model for target detection tasks.
7. A remote sensing image directional target detection system based on rotation feature perception, It is characterized in that The method for detecting directional targets in remote sensing images based on rotation feature perception as described in any one of claims 1 to 6 is applied to realize accurate recognition of directional target detection tasks in remote sensing images. The system comprises: The target detection model building module is used to construct a rotation feature perception module, and build a feature fusion network optimized and improved by the rotation feature perception module and its corresponding remote sensing image directional target detection model; A target detection model training module is used to perform target detection task training and testing on the remote sensing image directional target detection model; The target detection model application module is used to use the trained remote sensing image directional target detection model to perform target detection on the remote sensing image to be tested; The constructed rotation feature perception module includes: a normalization module, a multi-layer perceptron module, a depth separation convolution module, a group convolution module, a hole convolution module, a self-attention module, a feature normalization module, a feature splicing module and a feature reshaping module; wherein: The normalization module is used to standardize the input feature map; The multi-layer perceptron module is used to extract features from the normalized feature map; The deep separation convolution module, grouped convolution module and dilated convolution module are used as three independent branches to mine the feature information of feature maps under different receptive fields using different convolution operations; The self-attention module is used to extract rotation invariant features of different branches; The feature normalization module processes the outputs of different branches to obtain the feature map of the adapted shape. The calculation process can be formally expressed as: iAA =BN(F iA_qkv ), where F iAA ∈R H×W×C represents the feature map after feature normalization, F iA_qkv Represents the result of self-attention calculation; The feature concatenation module is used to aggregate feature maps of different branches along the channel dimension to obtain multi-dimensional rotation feature information; The feature reshaping module changes the size of the concatenated feature map through sampling / interpolation strategies to match the feature maps of other levels.