A rotating object detection method based on binary classification embedding space guided angle regression

Abstract

The present invention relates to a rotating target detection method using a binary classification embedding space to guide angle regression, and belongs to the technical field of remote sensing rotating target detection. First, the method decouples the angle prediction period of the remote sensing target into two mirror-symmetrical binary classification embedding spaces, and performs a preliminary category judgment on the direction of the target. Then, under the guidance of the binary classification subspace, the angle of the rotating target is further accurately predicted in the subspace. Finally, the true angle is calculated using the classification results obtained in the binary classification embedding space and the angle prediction in the subspace. This method effectively solves the influence of the angle regression period discontinuity problem on the network convergence during network training, improves the angle prediction accuracy of remote sensing rotating targets, and has good practical application value.

Description

Technical Field

[0001] The present invention relates to a remote sensing rotating target detection method, in particular to a rotating target detection method based on binary classification embedding space guided angle regression, and belongs to the technical field of remote sensing image target detection. Background Art

[0002] Target detection technology has become a primary method for processing remote sensing images. Target detection extracts effective information from wide-format remote sensing images and predicts the precise location and classification of different target categories. This technology supports the intelligent application of remote sensing data and plays a vital role in areas such as intelligent transportation, disaster prediction, port trade management, emergency search and rescue, and military movement detection.

[0003] As the spatial resolution of remote sensing images continues to increase, the shape and contour information of remote sensing targets becomes richer. Traditional horizontal box target detection methods are no longer able to accurately locate remote sensing targets with arbitrary angles and irregular aspect ratios. This is because the positioning of horizontal boxes for remote sensing targets with large aspect ratios and arbitrary angles will include a large amount of background interference, which will affect target recognition. Furthermore, in scenarios where remote sensing targets are densely packed, the use of horizontal boxes can easily lead to the non-maximum suppression algorithm filtering out the bounding boxes of positive samples due to the high overlap between the horizontal boxes of densely packed targets, resulting in missed detections.

[0004] Therefore, accurately detecting rotating remote sensing targets from wide-width, high-resolution remote sensing images has become a very important research topic. Currently, most rotating frame target detection methods generally choose to directly add an additional angle prediction vector based on the horizontal frame detector to predict the angle information of the rotating target. However, such methods have the problem of angle regression periodicity discontinuity. For example, when the target angle is at the edge of periodic change, slight angle offsets can easily cause a sudden change in the angle regression loss value, resulting in unstable network convergence and affecting the prediction accuracy of the rotating target angle. Although some researchers have solved the problem of angle periodicity discontinuity, it often requires complex conversion formulas and post-processing operations, which affects the processing efficiency of the network. Summary of the Invention

[0005] The purpose of the present invention is to address the shortcomings and defects of the existing technology and to solve the problem of limited angle prediction accuracy caused by the discontinuous regression period of the angle of remote sensing rotating targets. A rotating target detection method with binary classification embedded space guided angle regression is creatively proposed.

[0006] The method of the present invention is achieved through the following technical solutions.

[0007] A rotation target detection method based on binary classification embedding space guided angle regression includes the following steps:

[0008] Step 1: Build an encoder-decoder network based on convolutional neural network;

[0009] Step 2: After the encoder-decoder network, four prediction networks are constructed in parallel, including the target center prediction network, the target width and height prediction network, the binary embedding space category prediction network, and the subspace angle value prediction network;

[0010] Step 3: Input the remote sensing image into the encoder-decoder network based on convolutional neural network obtained in step 1 to obtain the deep semantic features of the remote sensing image.

[0011] Step 4: The deep semantic features of the remote sensing image obtained in step 3 are input in parallel into the target center point prediction network, target width and height prediction network, binary classification embedding space category prediction network and subspace angle value prediction network constructed in step 2 to obtain the target center point prediction value, target width and height prediction value, binary classification embedding space category prediction value and subspace angle value prediction value.

[0012] Step 5: Process the true rotation bounding box in the remote sensing image to obtain four supervisory signals for supervising network training, including the target center point coordinate value, target width and height value, target binary classification embedding space category value, and target binary classification subspace angle value;

[0013] Step 6: The target center coordinate value, target width and height value, target binary embedding space category value and target binary subspace angle value obtained in step 5 are respectively combined with the target center point prediction value, target width and height prediction value, binary embedding space category prediction value and subspace angle prediction value obtained in step 4 to calculate the loss value through the loss function, and then the network optimization training is performed through the back propagation algorithm until the network training converges to obtain the trained network model;

[0014] Step 7: Input the remote sensing image to be detected into the network model trained in step 6. The network model will output the target center point prediction value, target width and height prediction value, binary embedding space category prediction value and subspace angle value prediction value. Then, through post-processing operations, the four prediction values ​​are restored to the target rotation bounding box, thereby completing the rotation target detection of the binary embedding space guided angle regression.

[0015] The method of constructing a codec network based on a convolutional neural network in step 1 is as follows:

[0016] Step 1.1: Select a general backbone network as the encoder part of the encoder-decoder network.

[0017] Step 1.2: After the encoder part in step 1.1, three transposed convolutions with an upsampling factor of 2 are constructed in series. This combination of transposed convolution layers serves as the decoder part of the encoder-decoder network.

[0018] In step 2, the center point prediction network, the target width and height prediction network, the binary embedding space category prediction network, and the subspace angle prediction network are constructed as follows:

[0019] The center point prediction network uses two convolutional layers: the first layer uses a parameter-independent 3×3 convolutional layer with a stride of 1, a padding of 1, and 64 output channels; the second layer uses a parameter-independent 1×1 convolutional layer with a stride of 1, a padding of 0, and an output channel equal to the number of categories.

[0020] The width and height prediction network uses two convolutional layers: the first layer uses a parameter-independent 3×3 convolutional layer with a stride of 1, a padding of 1, and an output channel of 64; the second layer uses a parameter-independent 1×1 convolutional layer with a stride of 1, a padding of 0, and an output channel of 2.

[0021] The binary embedding spatial category prediction network uses two convolutional layers: the first layer uses a parameter-independent 3×3 convolutional layer with a stride of 1, a padding of 1, and an output channel of 64; the second layer uses a parameter-independent 1×1 convolutional layer with a stride of 1, a padding of 0, and an output channel of 2.

[0022] The subspace angle prediction network uses two convolutional layers: the first layer uses a parameter-independent 3×3 convolutional layer with a stride of 1, a padding of 1, and an output channel of 64; the second layer uses a parameter-independent 1×1 convolutional layer with a stride of 1, a padding of 0, and an output channel of 1.

[0023] The method of obtaining the deep semantic features of the remote sensing image in step 3 is as follows:

[0024] The remote sensing image is input into the encoder-decoder network based on convolutional neural network, and the convolution passes through the convolution layers of the encoder-decoder network in sequence. The output result of the last convolution layer is regarded as the deep semantic feature of the remote sensing image.

[0025] In step 4, the method for obtaining the target center point prediction value, the target width and height prediction value, the binary embedding space category prediction value, and the subspace angle value prediction value is as follows:

[0026] The target center point prediction value is obtained by the output of the last convolutional layer of the target center point prediction network; the target width and height prediction value is obtained by the output of the last convolutional layer of the target width and height prediction network; the binary embedding space category prediction value is obtained by the output of the last convolutional layer of the binary embedding space category prediction network; the subspace angle value prediction value is obtained by the output of the last convolutional layer of the subspace angle value prediction network;

[0027] For step 5 mentioned above, the method for processing the true rotation bounding box in the remote sensing image is as follows:

[0028] Step 5.1: Reorder the four coordinate points of the true annotation box of the remote sensing target in the dataset to ensure that the abscissa of the first coordinate point, that is, the x value, is the smallest among the four coordinate points, and then sort the remaining three coordinate points in clockwise order. Denote the sorted coordinate points as (x i , y i ), where i ∈ [0, 3].

[0029] Step 5.2: Calculate the center point of the target and the width and height of the target bounding box according to the four coordinate points in step 5.1.

[0030] Step 5.3: Calculate the category of the target in the binary classification embedding space according to the four coordinate points in step 5.1. The specific operation is as Figure 3 shown. Calculate the distances between two groups of coordinate points (x1, y1) and (x2, y2), (x2, y2) and (x3, y3), as shown in the following formulas (1) and (2):

[0031]

[0032]

[0033] Then, judge the magnitudes of d2 and d1. If d1 > d2, it is determined that the long side falls into the first and third quadrants of the Cartesian coordinate system, and the binary classification embedding space category d is determined to be 0; if d1 < d2, it is determined that the long side falls into the second and fourth quadrants, and the binary classification embedding space category d is determined to be 1.

[0034] Step 5.4: Calculate the subspace angle value of the target according to the four coordinate points in step 5.1. As Figure 3 shown, denote the two endpoint coordinates of the long side as (x a , y a ) and (x b , y b ), and then calculate the angle α in the embedding subspace through the following formula (3).

[0035]

[0036] For step 7 mentioned above, the post-processing operation of restoring the predicted values of the target center point, target width and height, binary classification embedding space category, and subspace angle value to the target rotation bounding box is as follows:

[0037] Step 7.1: Process the predicted values of the binary classification embedding space category and subspace angle value according to formula (4) to obtain the true rotation angle of the target rotation box.

[0038]

[0039] Among them, α′ represents the true angle value in the 180-degree representation; α represents the angle prediction value in the subspace; and d represents the category prediction value in the binary embedding space.

[0040] Step 7.2: Combine the calculated angle α′ with the predicted center point coordinates and the target width and height prediction values ​​to calculate the four-point coordinate values ​​of the remote sensing target Complete the prediction of the rotated object bounding box.

[0041] Beneficial effects

[0042] Compared with the existing technology, this method has the following advantages:

[0043] This method can effectively solve the impact of the angle regression period discontinuity problem on network convergence during network training, improve the accuracy of angle prediction, and at the same time, compared with other methods for solving the period discontinuity problem, it is simple and efficient.

[0044] The method of the present invention first decouples the angle prediction period of the remote sensing target into two mirror-symmetric binary classification embedding spaces, performs preliminary classification judgment on the direction of the target, and converts the angle periodic regression problem into a binary classification problem. Then, through the guidance of the binary classification embedding space, the angle of the rotating target is accurately predicted in the embedded subspace, avoiding the problem of loss value mutation due to slight angle changes at the angle regression period boundary during network training. Finally, the true angle of the remote sensing rotating target is calculated based on the classification results of the binary classification embedding space and the angle prediction results in the subspace. The calculation method is simple and does not generate cumbersome post-processing operations.

[0045] The present invention relates to a rotating target detection method using a binary classification embedding space to guide angle regression, and belongs to the technical field of remote sensing rotating target detection. First, the method decouples the angle prediction period of the remote sensing target into two mirror-symmetrical binary classification embedding spaces, and performs a preliminary category judgment on the direction of the target. Then, under the guidance of the binary classification subspace, the angle of the rotating target is further accurately predicted in the subspace. Finally, the true angle is calculated using the classification results obtained in the binary classification embedding space and the angle prediction in the subspace. This method effectively solves the influence of the angle regression period discontinuity problem on the network convergence during network training, improves the angle prediction accuracy of remote sensing rotating targets, and has good practical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 The overall flow chart of this method is shown in Figure 2.

[0047] Figure 2 Schematic diagram for comparing the binary classification embedding space guided angle regression method proposed in this method with other methods.

[0048] Figure 3 Schematic diagram for obtaining angle information in the binary embedding space and embedding subspace in this method. DETAILED DESCRIPTION

[0049] The method of the present invention is further described in detail below with reference to the accompanying drawings and examples.

[0050] Example

[0051] A rotation target detection method based on binary classification embedding space guided angle regression. The overall process is as follows Figure 1 As shown, the specific steps include:

[0052] Step 1: Build an encoder-decoder network based on convolutional neural network;

[0053] The present invention provides a specific implementation method, which includes the following steps:

[0054] Step 1.1: Select the ResNet-50 backbone network as the encoder part of the encoder-decoder network.

[0055] Step 1.2: After the encoder part in step 1.1, three transposed convolutions with an upsampling factor of 2 are constructed in series. This combination of transposed convolution layers serves as the decoder part of the encoder-decoder network.

[0056] The following example illustrates the construction method of transposed convolution with an upsampling multiple of 2:

[0057] First, construct a transposed convolution with an input channel of 2048, an output channel of 256, a convolution kernel size of 4, a convolution step of 2, and a padding of 1; then, construct a transposed convolution with both input and output channels of 256, a convolution kernel size of 4, a convolution step of 2, and a padding of 1; finally, construct a transposed convolution with both input and output channels of 256, a convolution kernel size of 4, a convolution step of 2, and a padding of 1.

[0058] Step 2: After the encoder-decoder network, four prediction networks are constructed in parallel, including the target center prediction network, the target width and height prediction network, the binary embedding space category prediction network, and the subspace angle prediction network.

[0059] The present invention provides a specific implementation method:

[0060] The center point prediction network uses two convolutional layers: the first layer uses a parameter-independent 3×3 convolutional layer with a stride of 1, a padding of 1, and 64 output channels; the second layer uses a parameter-independent 1×1 convolutional layer with a stride of 1, a padding of 0, and an output channel equal to the number of categories.

[0061] The width and height prediction network uses two convolutional layers: the first layer uses a parameter-independent 3×3 convolutional layer with a stride of 1, a padding of 1, and an output channel of 64; the second layer uses a parameter-independent 1×1 convolutional layer with a stride of 1, a padding of 0, and an output channel of 2.

[0062] The binary embedding spatial category prediction network uses two convolutional layers: the first layer uses a parameter-independent 3×3 convolutional layer with a stride of 1, a padding of 1, and an output channel of 64; the second layer uses a parameter-independent 1×1 convolutional layer with a stride of 1, a padding of 0, and an output channel of 2.

[0063] The subspace angle prediction network uses two convolutional layers: the first layer uses a parameter-independent 3×3 convolutional layer with a stride of 1, a padding of 1, and an output channel of 64; the second layer uses a parameter-independent 1×1 convolutional layer with a stride of 1, a padding of 0, and an output channel of 1.

[0064] Step 3: Input the remote sensing image into the encoder-decoder network based on convolutional neural network to obtain the deep semantic features of the remote sensing image.

[0065] The present invention provides a specific implementation method, which includes the following steps:

[0066] Step 3.1: Feed the remote sensing image into the ResNet-50 encoder network constructed in step 1.1 for feature extraction to obtain the intermediate feature F.

[0067] The following examples illustrate this.

[0068] The optical remote sensing image S with a size of 640×640 A Input into the ResNet-50 encoder network, the output of the last layer of convolutional network is the intermediate feature F, whose size is 20×20 and the number of channels is 2048.

[0069] Step 3.2: Input the intermediate feature F into the decoder network constructed in step 1.2 to improve the resolution of the intermediate feature.

[0070] The following examples illustrate this.

[0071] After the output of the encoder network with a size of 20×20 and a channel number of 2048 is input into the decoder, it undergoes three transposed convolutions with an upsampling factor of 2 to finally obtain the deep semantic features F d, its dimension is 256×160×160, that is, the size is 160×160 and the number of channels is 256.

[0072] Step 4: The deep semantic features of the remote sensing image obtained in step 3 are input into the target center point prediction network, the target width and height prediction network, the binary embedding space category prediction network and the subspace angle value prediction network in parallel, thereby obtaining the target center point prediction value, the target width and height prediction value, the binary embedding space category prediction value and the subspace angle value prediction value.

[0073] The present invention provides a specific implementation method, which includes the following steps:

[0074] Step 4.1: Deep semantic features F d Input into the target center point prediction network described in step 2, and the last convolution layer outputs the target center point prediction value Its dimension is C×160×160, where C represents the number of categories.

[0075] Step 4.2: Deep semantic features F d Input into the target width and height prediction network described in step 2, and the last convolution layer outputs the target center point prediction value wh p , whose dimensions are 2×160×160, where 2 represents both width and height.

[0076] Step 4.3: Deep semantic features F d Input into the binary classification embedding space category prediction network described in step 2, and the last convolutional layer outputs the target binary classification embedding space category prediction value Its dimension is 2×160×160, where 2 represents the number of categories in the binary embedding space.

[0077] Step 4.4: Deep semantic features F d Input into the subspace angle value prediction network described in step 2, and the last convolution layer outputs the subspace angle value prediction value α p , whose dimensions are 1×160×160, where 1 represents the angle value in the subspace.

[0078] Step 5: Process the true rotation bounding box in the remote sensing image to obtain four supervisory signals for supervising network training, including the target center point coordinate value, target width and height value, target binary classification embedding space category value, and target binary classification subspace angle value;

[0079] The present invention provides a specific implementation method, which includes the following steps:

[0080] Step 5.1: Reorder the four coordinate points of the true annotation box of the remote sensing target in the dataset to ensure that the abscissa of the first coordinate point, i.e., the x value, is the smallest among the four coordinate points, and then sort the remaining three coordinate points in clockwise order. Denote the sorted coordinate points as (x i , y i ), where i ∈ [0, 3].

[0081] Step 5.2: Calculate the center point of the target and the width and height of the target bounding box according to the four coordinate points in Step 5.1.

[0082] Step 5.3: Calculate the category of the target in the binary classification embedding space according to the four coordinate points in Step 5.1, and then calculate the angle in the embedding subspace.

[0083] To avoid the problem of angle mutation at the boundary of the angle regression period (as shown in Figure 2 (a) and (b)), first decouple the entire angle period according to the orientation of the target, divide it into two mirror-symmetric subspaces, and then calculate the angle values within the subspaces.

[0084] The following is an example for illustration.

[0085] As shown in Figure 3 , calculate the distances between two groups of coordinate points (x1, y1) and (x2, y2), (x2, y2) and (x3, y3), as shown in the following formulas (5) and (6):

[0086]

[0087]

[0088] Then, judge the magnitudes of d2 and d1. If d1 > d2, it is determined that the long side falls into the first and third quadrants of the Cartesian coordinate system, and the category d of the binary classification embedding space of the target is determined to be 0; if d1 < d2, it is determined that the long side falls into the second and fourth quadrants, and the category d of the binary classification embedding space of the target is determined to be 1.

[0089] Next, denote the two endpoint coordinates of the long side as (x a , y a ) and (x b , y b ), and then calculate the angle α within the subspace through the following formula (7).

[0090] [[ID=四十八]]

[0091] [[ID=四十九]]After decoupling the angle period by the above method, the regression period of angle prediction is shown in Figure 2 (c), which can avoid the problem of angle mutation at the boundary of the angle regression period.

[0092] Step 6: Calculate the loss value by using the loss function with the target center point coordinate value, target width and height value, target binary embedding space category value and target binary subspace angle value obtained in step 5 and the target center point prediction value, target width and height prediction value, binary embedding space category prediction value and subspace angle prediction value obtained in step 4, and then perform network optimization training by back propagation algorithm until the network training converges to obtain the trained network model;

[0093] The present invention provides a specific implementation method:

[0094] Based on the supervision signal prepared in step 5, the network is supervised and trained using the loss function.

[0095] The following example illustrates this.

[0096] For the target center point prediction task, the modified Focal Loss is used, as shown in the following formula (8):

[0097]

[0098] Among them, C represents the category, N represents the number of positive samples; H and W represent the width and height of the output feature map; Represents the predicted value at the (x, y) position on the channel of the cth category; Y xyc =1 is the true value at the (x, y) position on the channel of the cth category; α and β represent weight adjustment factors, whose values ​​are 2 and 4 respectively.

[0099] Then, for the binary embedding space prediction task, a loss function similar to the center point prediction task is used, as shown in the following formula (9):

[0100]

[0101] Where D represents the number of categories in the binary embedding space, M represents the number of positive samples; H and W represent the width and height of the output feature map; Represents the predicted value at the (x, y) position on the d-th category channel; R xyd Represents the true value at the (x,y) position on the d-th category channel.

[0102] In addition, the loss functions for width, height, and angle in the subspace all adopt the Smooth-L1 paradigm loss function, as shown in the following formula (10):

[0103]

[0104] Where x represents the difference between the true value and the predicted value, that is, h p-h g , w p -w g and α p -α g , where h, w and α represent the width, height and angle in the subspace respectively.

[0105] Step 7: Input the remote sensing image to be detected into the network model trained in step 6. The network model will output the target center point prediction value, target width and height prediction value, binary embedding space category prediction value and subspace angle value prediction value. Then, through post-processing operations, the four prediction values ​​are restored to the target rotation bounding box, thereby completing the rotation target detection of the binary embedding space guided angle regression.

[0106] The present invention provides a specific implementation method, which includes the following steps:

[0107] Step 7.1: Restore the true angle of the rotated target bounding box.

[0108] The following examples illustrate this.

[0109] The binary classification result d and the subspace angle α are post-processed and calculated according to formula (11) to obtain the true angle α′. It should be noted that α′ represents the true angle value in the 180-degree representation.

[0110]

[0111] Step 7.2: Combine the calculated angle α′ with the predicted center point coordinates (x, y) and the target width w and height h to calculate the four-point coordinate values ​​of the remote sensing target. Complete the prediction of the rotated object bounding box.

[0112] Several improvements can be made without departing from the principles of the present invention, and these should also be considered to fall within the scope of protection of the present invention.

Claims

1. A rotating target detection method based on binary classification embedding space guided angle regression, characterized by The following steps are involved: Step 1: Build an encoder-decoder network based on convolutional neural network; Step 2: Construct the target center prediction network, target width and height prediction network, binary embedding space category prediction network and subspace angle value prediction network; Step 3: Input the remote sensing image into the encoder-decoder network based on the convolutional neural network constructed in step 1 to obtain the deep semantic features of the remote sensing image; Step 4: Input the deep semantic features of the remote sensing image obtained in step 3 into the target center prediction network constructed in step 2 to obtain the target center prediction value; Input the deep semantic features of the remote sensing image obtained in step 3 into the target width and height prediction network constructed in step 2 to obtain the target width and height prediction values; Input the deep semantic features of the remote sensing image obtained in step 3 into the binary classification embedding spatial category prediction network constructed in step 2 to obtain the binary classification embedding spatial category prediction value; Input the deep semantic features of the remote sensing image obtained in step 3 into the subspace angle value prediction network constructed in step 2 to obtain the subspace angle prediction value; Step 5: Process the true rotation bounding box in the remote sensing image to obtain the target center coordinate value, target width and height value, target binary classification embedding space category value and target binary classification subspace angle value for supervised network training; Step 6: Calculate the loss value by using the loss function with the target center point coordinate value, target width and height value, target binary embedding space category value and target binary subspace angle value obtained in step 5 and the target center point prediction value, target width and height prediction value, binary embedding space category prediction value and subspace angle prediction value obtained in step 4, and then perform network optimization training by back propagation algorithm until the network training converges to obtain the trained network model; Step 7: Input the remote sensing image to be detected into the network model trained in step 6. The network model outputs the target center point prediction value, target width and height prediction value, binary embedding space category prediction value and subspace angle value prediction value. Then, through post-processing operations, the four prediction values ​​are restored to the target rotation bounding box, thereby completing the rotation target detection of the binary embedding space guided angle regression.

2. The rotating target detection method using binary classification embedding space guided angle regression according to claim 1 is characterized by: The method of constructing a codec network based on a convolutional neural network in step 1 is as follows: Step 1.1: Select the skeleton network as the encoder part of the encoder-decoder network; Step 1.2: After the encoder part in step 1.1, three transposed convolutional layers with an upsampling factor of 2 are constructed in series. This combination of transposed convolutional layers serves as the decoder part of the encoder-decoder network.

3. The rotating target detection method using binary classification embedding space guided angle regression according to claim 2 is characterized by: The method for constructing the combination of transposed convolutional layers is as follows: First, construct a transposed convolution with an input channel of 2048, an output channel of 256, a convolution kernel size of 4, a convolution stride of 2, and a padding of 1; then, construct a transposed convolution with an input and output channel of 256, a convolution kernel size of 4, a convolution stride of 2, and a padding of 1; finally, construct a transposed convolution with an input and output channel of 256, a convolution kernel size of 4, a convolution stride of 2, and a padding of 1.

4. The rotating target detection method using binary classification embedding space guided angle regression according to claim 1 is characterized by: In step 2, the method for constructing the center point prediction network is as follows: The center point prediction network uses two convolutional layers: the first layer uses a parameter-independent 3×3 convolutional layer with a stride of 1, a padding of 1, and 64 output channels; the second layer uses a parameter-independent 1×1 convolutional layer with a stride of 1, a padding of 0, and an output channel equal to the number of categories.

5. The rotating target detection method using binary classification embedding space guided angle regression according to claim 1, characterized in that: In step 2, the method for constructing the width and height prediction network is as follows: The width and height prediction network uses two convolutional layers: the first layer uses a parameter-independent 3×3 convolutional layer with a stride of 1, a padding of 1, and an output channel of 64; the second layer uses a parameter-independent 1×1 convolutional layer with a stride of 1, a padding of 0, and an output channel of 2.

6. The rotating target detection method using binary classification embedding space guided angle regression according to claim 1, characterized in that: In step 2, the method for constructing a binary classification embedding space category prediction network is as follows: The binary embedding spatial category prediction network uses two convolutional layers: the first layer uses a parameter-independent 3×3 convolutional layer with a stride of 1, a padding of 1, and an output channel of 64; the second layer uses a parameter-independent 1×1 convolutional layer with a stride of 1, a padding of 0, and an output channel of 2; In step 2, the method for constructing the subspace angle prediction network is as follows: The subspace angle prediction network uses two convolutional layers: the first layer uses a parameter-independent 3×3 convolutional layer with a stride of 1, a padding of 1, and an output channel of 64; the second layer uses a parameter-independent 1×1 convolutional layer with a stride of 1, a padding of 0, and an output channel of 1.

7. The rotating target detection method using binary classification embedding space guided angle regression according to claim 1, characterized in that: The method of obtaining the deep semantic features of the remote sensing image in step 3 is as follows: The remote sensing image is input into the convolutional neural network based encoder-decoder network, and passes through all the convolutional layers of the encoder-decoder network in sequence. The output result of the last convolutional layer is regarded as the deep semantic feature of the remote sensing image.

8. The rotating target detection method using binary classification embedding space guided angle regression according to claim 1, characterized in that: In step 4, the method for obtaining the target center point prediction value, the target width and height prediction value, the binary embedding space category prediction value, and the subspace angle value prediction value is as follows: The predicted value of the target center point is obtained from the output of the last convolutional layer of the target center point prediction network; the predicted value of the target width and height is obtained from the output of the last convolutional layer of the target width and height prediction network; the predicted value of the binary classification embedding space category is obtained from the output of the last convolutional layer of the binary classification embedding space category prediction network; the predicted value of the subspace angle value is obtained from the output of the last convolutional layer of the subspace angle value prediction network.

9. A rotation target detection method for binary classification embedding space-guided angle regression according to claim 1, characterized in that: In the step 5, the method for processing the true rotation bounding box in the remote sensing image is: Step 5.1: Reorder the coordinates of the four points of the real annotation box of the remote sensing target in the dataset to ensure that the horizontal coordinate of the first coordinate point, that is, the x value, is the smallest among the four coordinate points, and then sort the remaining three coordinate points in clockwise order. The sorted coordinate points are recorded as (x i ,y i ),i∈[0,3]; Step 5.2: Calculate the center point of the target and the width and height of the target bounding box according to the four coordinate points in step 5.1; Step 5.3: Calculate the category of the binary classification embedding space where the target is located according to the four coordinate points in step 5.1, and calculate the distances between two groups of coordinate points (x1, y1) and (x2, y2), (x2, y2) and (x3, y3), as shown in the following formulas (1) and (2): Then, judge the magnitudes of d2 and d1. If d1 > d2, it is determined that the long side falls into the first and third quadrants of the Cartesian coordinate system, and the binary classification embedding space category d of the target is determined to be 0; if d1 < d2, it is determined that the long side falls into the second and fourth quadrants, and the binary classification embedding space category d of the target is determined to be 1; Step 5.4: Calculate the subspace angle value of the target based on the four coordinate points in step 5.1, and mark the coordinates of the two endpoints of the long side of the target as (x a ,y a ) and (x b ,y b ), and then calculate the angle α in the embedded subspace by the following formula (3); 10. A rotation target detection method for binary classification embedding space-guided angle regression according to claim 1, characterized in that: In the step 7, the post-processing operation of restoring the predicted value of the target center point, the predicted value of the target width and height, the predicted value of the binary classification embedding space category and the predicted value of the subspace angle value to the target rotation bounding box is: Step 7.1: Process the predicted value of the binary classification embedding space category and the predicted value of the subspace angle value according to formula (4) to obtain the true rotation angle of the target rotation box; Among them, α′ represents the true angle value in the 180-degree representation method; α represents the predicted value of the angle in the subspace; d represents the predicted value of the binary classification embedding space category; Step 7.2: Combine the calculated angle α′ with the predicted center point coordinates and the target width and height prediction values ​​to calculate the four-point coordinate values ​​of the remote sensing target Complete the prediction of the rotated object bounding box.

Images