Remote sensing image target detection method based on angle priority strategy

Abstract

The present invention provides a remote sensing image target detection method based on an angle priority strategy, including: S1, marking the types of targets in a remote sensing image data set, and forming a training set and a verification set; S2, encoding the target direction through a multi-parameter angle encoding method; S3, using an angle priority strategy to build a target detection network; S4, using the training set and the verification set to train and test the target detection network; S5, using the trained target detection network to perform target detection on the remote sensing image to be detected. The present invention adopts a direction priority strategy to build a target detection network, which can detect the position, size, and direction of the target in the remote sensing image and identify the target type; adopts a multi-parameter angle encoding method to represent the target direction, which can avoid the boundary discontinuity problem during angle regression; adopts an angle channel attention mechanism, so that the target detection network can extract the directional features of the target in a targeted manner.

Description

Technical Field

[0001] The invention relates to the technical field of remote sensing image processing, and in particular to a remote sensing image target detection method based on an angle priority strategy. Background Art

[0002] In recent years, with the vigorous development of satellite technology, commercial satellites have gradually entered the public eye and gained great attention. The resulting massive satellite remote sensing data has huge application potential. Target detection technology on remote sensing images plays an important role in many fields such as urban and rural construction, environmental monitoring, military reconnaissance, and economic planning. It is very meaningful to make full use of remote sensing big data to improve the intelligence of target detection technology.

[0003] With the continuous development of deep learning technology, its application in the field of target detection has led to great development of target detection. In the target detection method based on neural network, the smallest rectangular box that can contain the target is used to locate the target. In the bird's-eye view used in remote sensing images, ground objects have arbitrary directionality. In order to better locate directional targets, directional rectangular boxes with arbitrary angles are used in remote sensing image target detection. However, there are still the following problems in the directional box target detection network that have not been solved:

[0004] First, the direction of the target is arbitrary, which makes it very challenging to detect. The ability of the existing directional frame target detection network needs to be improved. In the target detection network based on anchor frames, it is difficult for the anchor frame to correctly match the real target. Presetting a single anchor frame will lead to a lower matching rate, and presetting multiple anchor frames with different angles will cause serious redundancy of anchor frames.

[0005] Secondly, the direction of the target changes periodically and is continuous in any direction. When the directional target detector uses angles to represent the target direction, the upper and lower boundaries of the angle value range used are not continuous. This results in a huge difference in the angle values ​​on both sides of the range boundary, while the directions they represent are similar, which to some extent affects the learning effect of the detection network on the target direction.

[0006] Finally, the key to the detection network correctly detecting the direction of the target is to be able to effectively extract features of the target in any direction. Most existing detection networks use horizontal convolution kernels to extract features, and do not build targeted network modules for directional feature extraction, which limits the detection network's ability to extract directional features. Summary of the invention

[0007] The purpose of the present invention is to overcome the defects of the prior art, propose a remote sensing image target detection method based on an angle priority strategy, and adopt a new strategy to construct a direction box target detection network to improve the detection capability.

[0008] To achieve the above object, the present invention adopts the following specific technical solutions:

[0009] The remote sensing image target detection method based on the angle priority strategy provided by the present invention comprises the following steps:

[0010] S1. Label the types of objects in the remote sensing image dataset and build a training set and a validation set.

[0011] S2, encoding the target direction through a multi-parameter angle encoding method;

[0012] S3, using angle priority strategy to build target detection neural network;

[0013] S4, using the training set and the validation set to train and test the target detection neural network;

[0014] S5. Use the trained target detection network to perform target detection on the remote sensing image to be detected.

[0015] Preferably, step S1 specifically includes the following sub-steps:

[0016] Acquire a remote sensing image data set, and label the types of targets in each remote sensing image; wherein, the target is labeled using the minimum bounding rectangle containing the target, and the labeling information is the coordinates of the four vertices of the minimum bounding rectangle, and the center point position of the minimum bounding rectangle, the side length of the minimum bounding rectangle, and the inclination angle of the minimum bounding rectangle are extracted from the four vertex coordinates;

[0017] 75% of the remote sensing images in the labeled remote sensing image dataset are taken as training sets to train the target detection neural network, and the other 25% of the remote sensing images are taken as verification sets to verify the detection ability of the trained target detection neural network.

[0018] Preferably, step S2 specifically includes the following sub-steps:

[0019] Obtain the angle value of the target in each remote sensing image according to the annotation information of the remote sensing image dataset;

[0020] Use the multi-parameter angle encoding method to encode the angle value of the target into N parameters, N ≥ 3;

[0021] The entire angle detection range with a length of T is divided into N intervals, with angles θ1, θ2…θ N+1 is the boundary; where θ N+1 and θ1 are the upper and lower limits of the angle detection range respectively;

[0022] Let N parameters x1, x2…x N Respectively with θ1+nT, θ2+nT…θN +nT corresponds to;

[0023] When encoding the angle value θ, first determine which interval of the N intervals the angle value θ is in, that is, θ∈[θ a ,θ b ), where (a, b)∈{(1, 2), (2, 3), ..., (N, N+1)}, and then according to the angle value θ and the two boundary values ​​θ of the interval a ,θ b The difference between the two corresponding parameter values ​​x a 、x b ,θ a ,θ b The larger the difference with θ, the smaller the corresponding two parameter values ​​x a 、x b The smaller the x a 、x b The sum of is 1, and the parameter value x a 、x b Determined by the following formula:

[0024]

[0025] Divide the parameter value x a 、x b In addition, set the remaining parameter values ​​to 0, thus completing the angle encoding.

[0026] Preferably, the angle priority strategy means that the target detection neural network first detects the target direction, predicts the angle value of the target at the positions of all feature points within the target range, and then presets an anchor frame with the predicted angle at each feature point, and the anchor frame is easier to match with the real target frame.

[0027] Preferably, step S3 specifically includes the following sub-steps:

[0028] According to the angle priority strategy, OFRDet, a target detection neural network, is constructed. ResNet50FPN is used as the backbone network to extract features from the input image and obtain multi-scale feature maps.

[0029] Construct an angle coding regression module, with the input as the multi-scale feature map extracted by the backbone network and the output as the coding map of N channels, where each channel corresponds to a coding parameter; in the angle coding regression module, the angle coding value is predicted at each feature point through ordinary convolution, and the predicted angle coding value is decoded into an angle value. A direction anchor frame is preset according to the angle value, and the target detection neural network regresses and adjusts the anchor frame in the rough detection stage;

[0030] Construct a coarse detection stage. In the coarse detection stage, the multi-scale feature map is input into the angle channel attention module. The angle channel attention module has N angle channels with the same number of encoding parameters. An angle value is used in each angle channel. The angle value used in each angle channel is the lower boundary angle value θ1, θ2…θ of the N intervals in the encoding process. N ; In each angle channel, a rotation convolution kernel with a specific angle value performs a convolution operation on the multi-scale feature map to obtain N new feature maps, and the N new feature maps are weighted summed using the encoding map output by the angle encoding regression module to generate the feature map output by the angle channel attention module; finally, after ordinary convolution, the target bounding box and its classification result after regression adjustment of the preset anchor box are obtained;

[0031] A refined detection stage is constructed. In the refined detection stage, the multi-scale feature map is used as input. After AlignConv, ORConv and ordinary convolution, the bounding box obtained in the coarse detection stage is refined and adjusted. At the same time, the target is classified again to obtain the final detection result.

[0032] Preferably, step S4 specifically includes the following sub-steps:

[0033] The training set is input into the target detection neural network to train the target detection neural network, output the detection results and calculate the loss; during the training process, the complete loss function consists of three parts, namely angle encoding regression loss, coarse stage detection loss and fine stage detection loss, which are defined as follows:

[0034]

[0035] Among them, λ1 represents the balance coefficient of the first angle encoding regression loss, N E represents the total number of encoding values ​​with target values, m represents each feature point with target encoding value, n represents each encoding parameter on a feature point, L AER represents smooth L1 loss, represents the encoding value obtained by the target detection neural network, represents the target encoding value, λ2 and λ3 represent the loss balance coefficients, and N C 、N R Respectively represent the number of positive samples in the coarse detection stage and the refined detection stage, i represents each sample, L c represents the classification loss, represents the classification prediction of sample i in the coarse detection stage and the refined detection stage, represents the classification label of sample i, Iverson brackets indicate equations, L r represents the regression loss, represents the position prediction of sample i in the coarse detection stage and the refined detection stage, Represents the position label of sample i;

[0036] The network parameters are updated according to the loss function. After each preset round of network parameter update iterations, the verification set is used to test the detection accuracy of the target detection neural network. When the accuracy of the verification set detection no longer increases or reaches the preset upper limit of the training round, the training ends.

[0037] Preferably, before inputting the training set into the target detection neural network, the following steps are also included:

[0038] Multi-scale scaling, random flipping, and random angle rotation methods are used to operate remote sensing images and the type annotation of targets on remote sensing images to improve the diversity of targets in remote sensing image datasets.

[0039] The present invention can achieve the following technical effects:

[0040] 1. Using a multi-parameter angle encoding method to represent the target direction can avoid the boundary discontinuity problem during angle regression;

[0041] 2. Adopting the angle channel attention mechanism, the target detection neural network can extract the directional features of the target in a targeted manner;

[0042] 3. The present invention is verified on two internationally credible and challenging remote sensing image datasets. The mAPs tested on the DOTA and HRSC2016 datasets are 79.56% and 96.29%, respectively, which are sufficient to prove the performance advantages of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a flowchart of a remote sensing image target detection method based on an angle priority strategy provided in an embodiment of the present invention;

[0044] Figure 2 is an example diagram of a multi-parameter angle encoding method provided according to an embodiment of the present invention;

[0045] Figure 3 is a schematic diagram of the overall logical structure of a target detection neural network according to an embodiment of the present invention;

[0046] Figure 4 is a schematic diagram of the structure of an angle encoding regression module and an angle channel attention module in a target detection neural network according to an embodiment of the present invention;

[0047] Figure 5 is a schematic diagram of detection results on the DOTA data set using the remote sensing image target detection method based on the angle priority strategy according to an embodiment of the present invention;

[0048] Figure 6 It is a schematic diagram of the detection results of the remote sensing image target detection method based on the angle priority strategy according to an embodiment of the present invention on the HRSC2016 data set. DETAILED DESCRIPTION

[0049] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. In the following description, the same modules are represented by the same reference numerals. In the case of the same reference numerals, their names and functions are also the same. Therefore, the detailed description thereof will not be repeated.

[0050] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and do not constitute a limitation of the present invention.

[0051] Figure 1 The flowchart of the remote sensing image target detection method based on the angle priority strategy provided according to an embodiment of the present invention is shown.

[0052] like Figure 1 As shown, the remote sensing image target detection method based on the angle priority strategy provided by the embodiment of the present invention includes the following steps:

[0053] S1. Label the types of objects in the remote sensing image dataset and build training and validation sets.

[0054] There are two ways to determine the target types in the remote sensing image to be detected. The first is to use a public remote sensing image dataset that has been annotated with the target types to be detected, and the second is to take a large number of remote sensing images and manually annotate the target types to be detected.

[0055] The annotation needs to use a rectangular box with a direction, and the rectangular box is the smallest circumscribed rectangular box that can contain the target. The annotation information is the coordinates of the four vertices of the rectangular box, from which the center point position of the rectangular box, the side length of the rectangular box, and the tilt angle of the rectangular box can be extracted.

[0056] After obtaining the remote sensing image dataset with target type annotations, 75% of the remote sensing images are used as the training set to train the target detection neural network, and the other 25% of the remote sensing images are used as the verification set to verify the detection ability of the trained target detection neural network.

[0057] S2. Encode the target direction through a multi-parameter angle encoding method.

[0058] Determine the angle detection range required to represent the target direction, and obtain the angle value of each target according to the annotation information of the remote sensing image data set in step S1. Use the multi-parameter angle encoding method to encode the angle value into N parameters, and the number of parameters N ≥ 3. Divide the entire angle detection range with a length of T into N intervals, with angles θ1, θ2, ... θ N+1 is the boundary, where θ N+1 The upper limit of the angle detection range, θ1 is the lower limit of the angle detection range, and the difference is T, which represents the same direction in target detection. Considering the periodicity of the direction and the subsequent decoding operation, let N parameters x1, x2…x N and θ1+nT, θ2+nT…θ N +nT (n is an arbitrary integer) correspond to each other one by one. When encoding an angle value θ, first determine which interval of the N intervals the angle value belongs to, that is, θ∈[θ a ,θ b ), where (a, b)∈{(1, 2), (2, 3), ..., (N, N+1)}, and then according to θ and the two boundary values ​​θ of the interval a ,θ b The difference between them determines the two corresponding parameter values ​​x a 、x b ,θ a ,θ b The larger the difference with θ, the smaller the corresponding two parameter values ​​x a 、x b The smaller the x a 、x b The sum of is 1, and the parameter value x a 、x b Determined by the following formula:

[0059]

[0060] Except for these two parameters, set the remaining parameters to 0, thus completing the angle encoding.

[0061] The target direction information is encoded into multiple parameters through the multi-parameter angle encoding method and provided to the target detection neural network for learning. Then the encoded values ​​learned by the target detection neural network need to be decoded into angle values ​​for subsequent processing and result display.

[0062] According to the above encoding process and x a 、x b The definition formula of θ can be obtained by the following formula:

[0063] θ=x a θ a +x b θ b ;

[0064] Therefore, when decoding, we only need to find the interval where the encoded angle is located, and then use the corresponding encoding parameters to weight the sum of the two boundary angle values ​​of the interval to get the angle. Ideally, only two or one of the encoded values ​​is non-zero, and it is easy to find the corresponding angle interval. However, in actual situations, the results learned by the target detection neural network will be biased, resulting in N values ​​being non-zero.

[0065] The decoding idea is as follows: two adjacent N coding values ​​are added in sequence, and the two with the largest sum are considered to be ideal non-zero coding values; according to the angle corresponding to the ideal non-zero coding value, the angle interval with it as the boundary is found in the angle detection range; the two boundary angle values ​​are weightedly summed using the coding parameters corresponding to each of them, to obtain the decoded angle value.

[0066] Figure 2 An example of a multi-parameter angle encoding method according to an embodiment of the present invention is shown. In the example, the number of encoding parameters is 4, the angle detection range is [-π / 2, π / 2), the range size T is π, and the upper and lower boundaries are π / 2 and -π / 2, respectively, which are divided into 4 intervals with -π / 2, -π / 4, 0, π / 4, and π / 2 as boundaries. The angle values ​​corresponding to the four parameters x1, x2, x3, and x4 are -π / 2+nπ, -π / 4+nπ, 0+nπ, and π / 4+nπ, respectively, where n is an arbitrary integer. Figure 2 The middle angle θ is between [-π / 2, -π / 4), and the difference between it and the two boundary angle values ​​of -π / 2 and -π / 4 in this interval is α and β respectively. The encoding result of this angle composed of four parameters is as follows Figure 2 Shown in the middle.

[0067] S3. Use angle priority strategy to build target detection neural network.

[0068] The angle priority strategy means that the neural network first detects the target direction information and predicts the angle value of the target at the position of all feature points within the target range. After the target detection neural network obtains the angle value, an anchor box with a predicted angle is preset on each feature point, and the anchor box is easier to match with the real target box, making the subsequent bounding box regression more accurate.

[0069] In the detection network OFRDet built according to the angle-first strategy, ResNet50 FPN is used as the backbone network to extract features from the input image and obtain multi-scale feature maps.

[0070] Construct an angle encoding regression module, with the input as the feature map extracted by the backbone network and the output as the encoding map of N channels, where each channel corresponds to an encoding parameter. In this module, the angle encoding value is predicted at each feature point through ordinary convolution. The angle should be the angle of the target to which the feature point belongs. If the feature point does not belong to any target, the encoding value predicted for the feature point has no target value and will not be constrained by the loss function. The encoding value predicted by the angle encoding regression module is decoded into an angle value, and a direction anchor frame is preset according to the angle value. The neural network regresses and adjusts the anchor frame in the rough detection stage.

[0071] Construct a coarse detection stage, in which the multi-scale feature map is input into the angle channel attention module, which has N angle channels with the same number of encoding parameters in S2. Each angle channel uses an angle value, which is the lower boundary angle value θ1, θ2…θ of the N intervals in the encoding process of step S2. N In each angle channel, the feature map is convolved by the rotation convolution kernel with the above specific angle value to obtain N new feature maps. After the N new feature maps are weighted and summed using the encoding map output by the angle encoding regression module, the feature map output by the angle channel attention module is generated. Finally, after ordinary convolution, the target bounding box and its classification result after regression adjustment of the preset anchor box are obtained.

[0072] A refined detection stage is constructed. In the refined detection stage, the multi-scale feature map is used as input. After AlignConv, ORConv and ordinary convolution, the bounding box obtained in the coarse detection stage is refined and adjusted. At the same time, the target is classified again to obtain the final detection result.

[0073] Figure 3 The overall logical structure of the target detection neural network according to an embodiment of the present invention is shown.

[0074] like Figure 3 As shown in the figure, the entire target detection neural network uses ResNet50 FPN as the backbone network to extract features and generate multi-scale feature maps. Each scale feature map is followed by a detection head for target detection. In the detection head, Figure 3There are three branches in total, namely the angle encoding regression module (AERM), the coarse detection stage, and the refined detection stage. First, the AERM obtains the encoding value to predict the angle, and uses the predicted angle to preset the direction anchor box on each feature point for use in the coarse detection stage, and then sends the encoding value to the angle channel attention module (ACAM) as its weight; the coarse detection stage uses the four angle channels in ACAM to extract the directional features in a targeted manner, and then performs bounding box regression and target classification respectively through ordinary convolution operations; the refined detection stage uses AlignConv to perform feature alignment, and then extracts the directional features through ORConv, and finally obtains the final detection result through the ordinary convolution layer.

[0075] Figure 4 The logical structure of the angle encoding regression module and the angle channel attention module in the target detection neural network according to an embodiment of the present invention is shown.

[0076] like Figure 4 As shown, the number of angle encoding parameters is 4, and the number of angle channels is also 4. Figure 4 The upper middle part is the Angle Encoding Regression Module (AERM), which takes as input the feature map extracted by the backbone network. After three convolutions, a 4-channel encoding map is obtained. The encoding map is compared with the ideal encoding map generated by multi-parameter angle encoding to calculate the loss, so that the network can learn the angle encoding process. The encoding map is also used as an attention map in the Angle Channel Attention Module (ACAM). Figure 4 The lower middle part is ACAM, and its input is also the feature map extracted by the backbone network. This module has 4 angle channels, and each angle channel uses an angle value, which is the lower boundary angle value of the 4 intervals during the encoding process. In the angle channel, the input feature map is convolved by the rotation convolution kernel with the above specific angle value to generate 4 new feature maps. After the 4 new feature maps are weighted and summed using the corresponding attention maps, the feature map output by the angle channel attention module is generated.

[0077] S4. Use the training set and validation set to train and test the target detection neural network.

[0078] The training set is input into the target detection neural network to train the target detection neural network, output the detection results and calculate the loss; during the training process, the complete loss function consists of three parts, namely angle encoding regression loss, coarse stage detection loss and fine stage detection loss, which are defined as follows:

[0079]

[0080] Among them, in the first angle coding regression loss, λ1 represents the balance coefficient of the first angle coding regression loss, N Erepresents the total number of encoding values ​​with target values, m represents each feature point with target encoding value, n represents each encoding parameter on a feature point, L AER represents smooth L1 loss, represents the encoding value obtained by the target detection neural network, represents the target encoding value. The detection losses in both the coarse detection stage and the refined detection stage are composed of the target classification loss and the position regression loss, and the position regression loss is only obtained from positive samples. In the second and third terms of the above formula, λ2 and λ3 represent the loss balance coefficients, respectively, and N C 、N R Respectively represent the number of positive samples in the coarse detection stage and the refined detection stage, i represents each sample, L c represents the classification loss, represents the classification prediction of sample i in the coarse detection stage and the refined detection stage, represents the classification label of sample i, It represents the Iverson bracket indicator equation, that is, when i is a positive sample, the value is 1, L r represents the regression loss, represents the position prediction of sample i in the coarse detection stage and the refined detection stage, Represents the position label of sample i;

[0081] The network parameters are updated according to the loss function. As the loss function gradually decreases, the neural network's ability to detect targets becomes stronger and stronger. After each preset round of network parameter update iterations, the target detection neural network's detection accuracy is tested using the validation set. When the accuracy of the validation set detection no longer increases or reaches the preset upper limit of the training round, the training ends.

[0082] In one example of the present invention, before the training set is input into the target detection neural network, multi-scale scaling, random flipping, and random angle rotation methods are used to operate on the remote sensing images and the type labeling of the targets on the remote sensing images to improve the diversity of targets in the remote sensing image data set.

[0083] S5. Use the trained target detection network to perform target detection on the remote sensing image to be detected.

[0084] The remote sensing image to be detected is sent as input to the target detection neural network whose parameters have been trained in step S4 for target detection, and the required target information can be obtained in the output.

[0085] Figure 5 The detection results of the remote sensing image target detection method based on the angle priority strategy according to the embodiment of the present invention on the DOTA data set are shown.

[0086] like Figure 5As shown in the figure, different types of targets are marked with boxes of different colors, a total of 15 types of targets, namely airplanes (PL), baseball fields (BD), bridges (BR), sports fields (GTF), small vehicles (SV), large vehicles (LV), ships (SH), tennis courts (TC), basketball courts (BC), storage tanks (ST), football fields (SBF), roundabouts (RA), ports (HA), swimming pools (SP) and helicopters (HC). Figure 5 It can be seen that the remote sensing image target detection method provided by the present invention is capable of detecting multiple types of objects with arbitrary directions and dense distribution in a complex background, achieving satisfactory results.

[0087] Figure 6 The detection results of the remote sensing image target detection method based on the angle priority strategy on the HRSC2016 dataset are shown.

[0088] like Figure 6 As shown in the figure, there is only one type of target detected, namely the ship, which is marked with a directional rectangular box. Figure 6 It can be seen that the remote sensing image target detection method provided by the present invention can complete the detection of ships of different scales and directions with high quality at different resolutions.

[0089] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0090] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

[0091] The above specific implementations of the present invention do not constitute a limitation on the protection scope of the present invention. Any other corresponding changes and modifications made based on the technical concept of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A remote sensing image target detection method based on an angle priority strategy, characterized in that: The steps include: S1. Label the types of objects in the remote sensing image dataset and build a training set and a validation set. S2. Encode the target direction by a multi-parameter angle encoding method; Step S2 specifically includes the following sub-steps: Obtaining an angle value of a target in each remote sensing image according to the annotation information of the remote sensing image data set; Use the multi-parameter angle encoding method to encode the angle value of the target into N parameters, N ≥ 3; The entire angle detection range with a length of T is divided into N intervals, with angles θ1, θ2…θ N+1 is the boundary; where θ N+1 and θ1 are the upper and lower limits of the angle detection range respectively; Let N parameters x1, x2…x N Respectively with θ1+nT, θ2+nT…θ N +nT corresponds to; When encoding the angle value θ, first determine which interval of the N intervals the angle value θ is in, that is, θ∈[θ a ,θ b ), where (a, b)∈{(1, 2), (2, 3), ..., (N, N+1)}, and then according to the angle value θ and the two boundary values ​​θ of the interval a ,θ b The difference between the two corresponding parameter values ​​x a 、x b ,θ a ,θ b The larger the difference with θ, the smaller the corresponding two parameter values ​​x a 、x b The smaller the x a 、x b The sum of is 1, and the parameter value x a 、x b Determined by the following formula: ； Divide the parameter value x a 、x b In addition, set the other parameter values ​​to 0, and the angle encoding is completed. S3. Constructing a target detection neural network using an angle priority strategy; the angle priority strategy means that the target detection neural network first detects the target direction, predicts the angle value of the target at the positions of all feature points within the target range, and then presets an anchor frame with a predicted angle at each feature point, and the anchor frame is more easily matched with the real target frame; S4, using the training set and the verification set to train and test the target detection neural network; S5. Use the trained target detection neural network to perform target detection on the remote sensing image to be detected.

2. The remote sensing image target detection method based on the angle priority strategy as claimed in claim 1, characterized in that: Step S1 specifically includes the following sub-steps: Acquire a remote sensing image data set, and mark the types of targets in each remote sensing image; wherein, the target is marked with a minimum bounding rectangle containing the target, and the marking information is the coordinates of the four vertices of the minimum bounding rectangle, and the center point position of the minimum bounding rectangle, the side length of the minimum bounding rectangle, and the inclination angle of the minimum bounding rectangle are extracted from the four vertex coordinates; 75% of the remote sensing images in the labeled remote sensing image data set are taken as training sets to train the target detection neural network, and the other 25% of the remote sensing images are taken as verification sets to verify the detection ability of the trained target detection neural network.

3. The remote sensing image target detection method based on the angle priority strategy as claimed in claim 1, characterized in that: Step S3 specifically includes the following sub-steps: According to the angle priority strategy, OFRDet is constructed to detect objects, and ResNet50 FPN is used as the backbone network to extract features from the input image to obtain a multi-scale feature map. Construct an angle coding regression module, the input of which is the multi-scale feature map extracted by the backbone network, and the output is a coding map of N channels, each channel corresponding to a coding parameter; in the angle coding regression module, an angle coding value is predicted at each feature point by ordinary convolution, and the predicted angle coding value is decoded into an angle value, and a direction anchor frame is preset according to the angle value, and the target detection neural network regresses and adjusts the anchor frame in the rough detection stage; A coarse detection stage is constructed, in which the multi-scale feature map is input into an angle channel attention module, the angle channel attention module has N angle channels with the same number as the encoding parameters, and an angle value is used in each angle channel. The angle value used in each angle channel is the lower boundary angle value θ1, θ2...θ of the N intervals in the encoding process. N ; In each angle channel, a rotation convolution kernel with a specific angle value performs a convolution operation on the multi-scale feature map to obtain N new feature maps, and the N new feature maps are weighted summed using the encoding map output by the angle encoding regression module to generate the feature map output by the angle channel attention module; finally, after ordinary convolution, the target bounding box and its classification result after regression adjustment of the preset anchor box are obtained; A refined detection stage is constructed, in which the multi-scale feature map is used as input, and the bounding box obtained in the coarse detection stage is refined and adjusted through AlignConv, ORConv and ordinary convolution, and the target is classified again to obtain the final detection result.

4. The remote sensing image target detection method based on the angle priority strategy as claimed in claim 3, characterized in that: Step S4 specifically includes the following sub-steps: The training set is input into the target detection neural network to train the target detection neural network, the detection result is output and the loss is calculated; wherein, during the training process, the complete loss function consists of three parts, namely, angle encoding regression loss, coarse stage detection loss and fine stage detection loss, which are defined as follows: ； in, represents the balance coefficient of the first angle encoding regression loss, represents the total number of coded values ​​with target value, Represents each feature point with a target encoding value, Represents each encoding parameter on a feature point, represents smooth L1 loss, represents the encoding value obtained by the target detection neural network, represents the target encoding value, , They represent the loss balance coefficients, , Respectively represent the number of positive samples in the coarse detection stage and the refined detection stage, Represents each sample, represents the classification loss, , Indicates the coarse detection stage and the refined detection stage for the sample Classification prediction of Representation sample The classification label of Iverson brackets indicate equations, represents the regression loss, , Indicates the coarse detection stage and the refined detection stage for the sample The location prediction, Representation sample The location tag; The network parameters are updated according to the loss function. After each preset round of network parameter update iterations, the verification set is used to test the target detection accuracy of the target detection neural network. When the accuracy of the verification set detection no longer increases or reaches the preset upper limit of the training rounds, the training is terminated.

5. The remote sensing image target detection method based on angle priority strategy as claimed in claim 4, characterized in that: Before inputting the training set into the target detection neural network, the following steps are also included: The remote sensing image and the type labeling of the target on the remote sensing image are operated by adopting the methods of multi-scale scaling, random flipping and random angle rotation, so as to improve the diversity of the targets in the remote sensing image data set.

Images