Rotating object detection method based on bidirectional feature fusion network

Abstract

The application discloses a rotating target detection method based on a bidirectional feature fusion network, and specifically comprises the following steps: step 1, obtaining a preprocessed image Input; step 2, performing feature extraction, adaptive feature fusion and double-branch feature information decoupling on the image X obtained in step 1 to construct a rotating target detection network model based on a bidirectional feature fusion network; step 3, training the model constructed in step 3 by using the data set preprocessed in step 1 to obtain a trained rotating target detection model based on the bidirectional feature fusion network; and step 4, placing the test set image processed in step 1 into the model trained in step 3, and the rotating target detection model based on the bidirectional feature fusion network is obtained. The application overcomes the problems of large scale gap of the rotating target in the aerial remote sensing image and small target detection, and further accurately detects the ground rotating target.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology and relates to a rotating target detection method based on a bidirectional feature fusion network. Background Technology

[0002] With the rapid development of remote sensing technology, the application scope of remote sensing images is expanding, playing a crucial role in scientific research and practical applications. However, because small targets in remote sensing images occupy an extremely small scale, they are easily lost in deeper features during multiple downsampling, leading to missed detections. Furthermore, compared to natural images, aerial remote sensing images exhibit a much wider range of target scale variations. This is not only reflected in the significant scale changes of the same target at different altitudes (resolutions), but also in the scale differences between different types of targets at the same altitude (resolution), such as small-scale cars and intersections, large-scale playgrounds and stadiums, and uniquely elongated shapes like airport runways and bridges. Therefore, the large scale differences of targets in remote sensing images and the detection of small targets remain extremely challenging problems. Accurately detecting and identifying these rotating targets with excessively large scale differences and small scales is of great significance for applications in agriculture, urban planning, and environmental monitoring.

[0003] Remote sensing image target features are typically manifested in various aspects, such as color, texture, shape, and context. A single feature is often insufficient to adequately describe the overall appearance of a target. Therefore, the appropriate selection and fusion of features can improve the ability of rotating target detection algorithms to identify small targets in different poses and targets of different scales.

[0004] Numerous scholars have conducted extensive research on feature fusion strategies. However, these studies suffer from limitations, focusing only on feature fusion within a single network architecture. Specifically, the local perception mechanism of CNNs restricts their receptive field, while the attention mechanism between spatial regions in Transformers struggles to guarantee local detail and efficiency. Therefore, feature fusion from a single network architecture may fail to balance the model's global perspective, precise detail, and model efficiency. In conclusion, achieving accurate detection of rotating targets in remote sensing images remains a significant challenge, especially given the large scale differences and the impact of detecting small targets. Summary of the Invention

[0005] The purpose of this invention is to provide a rotating target detection method based on a bidirectional feature fusion network. This method can still accurately detect rotating ground targets even when there are problems such as large differences in target scale or small targets.

[0006] The technical solution adopted in this invention is a rotating target detection method based on a bidirectional feature fusion network, which specifically includes the following steps:

[0007] Step 1: Preprocess the images in the pre-divided training and test sets of the DIOR-R dataset to obtain the preprocessed training and test set images (Input).

[0008] Step 2: Construct a rotating target detection network model based on a bidirectional feature fusion network. This model is used to perform feature extraction, adaptive feature fusion, and decoupling of bi-branch feature information on the image Input obtained in Step 1.

[0009] Step 3: Use the training set preprocessed in Step 1 to train the model constructed in Step 2, and obtain the trained rotating target detection model based on the bidirectional feature fusion network.

[0010] Step 4: Input the test set images preprocessed in Step 1 into the model trained in Step 3 to finally obtain the aerial remote sensing images after target detection.

[0011] The invention is further characterized by:

[0012] The image preprocessing process in step 1 is as follows: the image is randomly flipped, randomly cropped, and color transformed in sequence to finally obtain the preprocessed training set and test set images Input.

[0013] The specific process of step 2 is as follows:

[0014] Step 2.1: Use ResNet50 as the backbone network and load the network weights pre-trained on the ImageNet dataset. Then, use the pre-processed image input as the input to the backbone network to extract multi-scale features P{P2,P3,P4,P5} of local details. The sizes of P2, P3, P4, and P5 are 1 / 4, 1 / 8, 1 / 16, and 1 / 32 of the original image size, respectively. The process is shown in formula (1).

[0015] P{P2,P3,P4,P5}=Backbone(Input)(1)

[0016] Step 2.2: The multi-scale features P{P2,P3,P4,P5} of local details obtained in Step 2.1 are used as input to TransBlock to obtain the corresponding features T{T2,T3,T4,T5} that aggregate global features and deep semantic information, as shown in Formula (2); The features T{T2,T3,T4,T5} that aggregate global features and deep semantic information and the multi-scale features P{P2,P3,P4,P5} obtained in Step 2.1 are used as input to obtain the gain features F{F2,F3,F4,F5} through the multi-scale feature dynamic fusion network DFM, as shown in Formula (3).

[0017] T{T2,T3,T4,T5}=TransBlock(P)(2)

[0018] F{F2,F3,F4,F5}=DFM(P,T) (3)

[0019] Step 2.3: The gain feature F obtained in step 2.2 is used as the input of RPN to obtain the accurate rotation target detection box. The feature map of gain feature F is adjusted by 1×1 convolution, and then the channel weights are re-divided using SEBlock to obtain the accurate rotation target detection box (x,y,w,h,α,β), as shown in formula (4):

[0020] (x,y,w,h,α,β)=SEBlock(Conv 1×1 (Feature Map) (4)

[0021] Where (x, y) represents the coordinates of the center point of the anchor frame, w and h represent the width and height of the anchor frame, and α and β are the offsets of the top midpoint and right midpoint of the initial anchor frame.

[0022] Step 2.4: The gain features from Step 2.2 and the detection boxes extracted in Step 2.3 are used as inputs. The information is decoupled through a dual-branch detection head to obtain accurate target detection information. In the detection head, the detection box regression branch uses two convolutions with a stride of 1 and a kernel size of 3×3 to adjust the information. Then, the detection boxes are accurately regressed through a fully connected layer. The process is shown in Formula (5). The classification branch uses two fully connected layers to decouple the classification information of the Feature Map. At the same time, the information of the regression branch is fused to achieve accurate classification of the target features. The process is shown in Formula (6). In Formula (6), the value of j ranges from 1 to 4, which corresponds to the feature map number after downsampling the original image by 4 times, 8 times, 16 times, and 32 times. DF(A,B) shows the dynamic fusion DF process. A and B represent the regression features and the initial feature map features after passing through two linear layers in the DF operation. θ j Representative fusion factor:

[0023] Regression = Linear(Conv) 3×3 (Conv 3×3 (Feature Map))) (5)

[0024]

[0025] The specific process of step 2.2 is as follows:

[0026] Step 2.2.1: The multi-scale features P{P2,P3,P4,P5} extracted from the backbone features are input into the two TSC modules in TransBlock to extract the features T{T2,T3,T4,T5} that aggregate global features and deep semantic information. The process is shown in the following formula (7). The TSC module contains two route branches. The first branch first flattens and reshapes the input features. After passing through the fully connected Linear layer to obtain the correlation between the pairs of features, the flattened features are restored to their original shape. Then, they are processed through the depthwise separable convolution DConv to obtain the information-adjusted features A. i As shown in formula (8) below; the second branch is similar to the first branch, and the feature V is obtained after flattening and reshaping, full-connected linear layer and recovery and reshaping. i The process is shown in formula (9); then, for feature A... i With V i Multiplying the aggregated global features and deep semantic features yields Zaggregate. i The process is shown in formula (10);

[0027] Furthermore, Z i Output feature P connected to the residual i The summation yields the final output feature T. i The process is shown in formula (11):

[0028] T{T2,T3,T4,T5}=TSC(TSC(P)) (7)

[0029] A i =DConv(Reshape(Linear(Reshape(P i (8)

[0030] V i =Reshape(Linear(Reshape(P)) i ))) (9)

[0031] Z i =A i ⊙V i (10)

[0032]

[0033] The value of i in the above formula ranges from 2 to 5, corresponding to the subscript numbers in P2, P3, P4, and P5;

[0034] Step 2.2.2: Multiply the multi-scale features P of local details and the multi-scale features T of global semantic information by their corresponding dynamic feature weights, and then add them together to obtain the gain feature F. i The process is shown in formula (12) below:

[0035]

[0036] Where, λ i As the fusion factor, the value of i in the above formula ranges from 2 to 5, corresponding to the subscript numbers in P2, P3, P4, and P5.

[0037] The loss function used in step 3 of the training process is as follows:

[0038] Smooth L1 Loss Function: The Smooth L1 loss function proposes to be a piecewise function, where x1 is the numerical difference between the predicted bounding box and the ground truth bounding box.

[0039]

[0040] CE loss function: CELows is the cross-entropy loss, where y is the label, x2 is the probability of the predicted class output by the softmax function, n is the number of samples, and the learning objective of the function is to make the output x2 approximate the label y.

[0041]

[0042] The total loss function is:

[0043] Loss = smooth L1 +L CE (15).

[0044] The beneficial effects of this invention are that the proposed rotating target detection method based on a bidirectional feature fusion network addresses the problems of large scale differences in rotating targets and the detection of small targets in images. The detector first achieves a wider receptive field by combining the advantages of CNN and Transformer through a bidirectional feature fusion network, effectively handling the scale variation problem of rotating targets. Then, a dynamically fused dual-branch detection head designed for information decoupling is used to independently process classification and regression information, and dynamic fusion technology enhances the classification and detection performance for small targets. Verification on public datasets demonstrates that the method of this invention has high metrics and excellent performance. Attached Figure Description

[0045] Figure 1 This is a flowchart illustrating the rotating target detection method based on a bidirectional feature fusion network according to the present invention.

[0046] Figure 2This is a structural diagram of the rotating target detection network model based on bidirectional feature fusion network constructed in the rotating target detection method based on bidirectional feature fusion network of the present invention;

[0047] Figure 3 This is a structural diagram of the TransBlock module in the rotating target detection method based on bidirectional feature fusion network of the present invention;

[0048] Figure 4 This is a structural diagram of the dual-branch dynamic fusion feature module in the rotating target detection method based on a bidirectional feature fusion network of the present invention;

[0049] Figure 5 This is a structural diagram of the dual-branch detection head module in the rotating target detection method based on bidirectional feature fusion network of the present invention;

[0050] Figure 6 The image shows the result of small target detection using the rotating target detection method based on bidirectional feature fusion network of this invention;

[0051] Figures 7(a) to 7(d) The image shows the detection results of targets with large scale differences in the rotating target detection method based on bidirectional feature fusion network of this invention. Detailed Implementation

[0052] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0053] Example 1

[0054] This invention relates to a rotating target detection method based on a bidirectional feature fusion network, the process of which is as follows: Figure 1 As shown, the specific steps include the following:

[0055] Step 1: Input the DIOR-R dataset, which contains 23,463 images and 192,518 instances. The image size is 800×800, and the spatial resolution is 0.5 meters to 30 meters. 11,725 ​​images are used for training and serve as the training set, while 11,738 images are used for testing and serve as the test set.

[0056] The images in the pre-divided training and test sets of the DIOR-R dataset are preprocessed by randomly flipping, randomly cropping, and color transforming the images to obtain the preprocessed training and test set images as input.

[0057] Step 2: Perform feature extraction, adaptive feature fusion, and decoupling of dual-branch feature information on the image Input obtained in Step 1 to construct a rotating target detection network model based on a bidirectional feature fusion network.

[0058] Figure 2The constructed model consists of a backbone network, a Dynamic Fusion Module (DFM), an Enhanced Region Proposal Network (ERPN), and a Double-Head detection head. Based on... Figure 2 A rotating target detection network model based on a bidirectional feature fusion network is constructed by using Backbone for feature extraction from the image input, DFM for feature fusion, ERPN for bounding box generation, and Double-Head for feature decoupling. The specific steps are as follows:

[0059] Step 2.1: Use ResNet50 as the backbone network, and load the network weights pre-trained on the ImageNet dataset. Then, use the pre-processed image input as the input to the backbone network to extract multi-scale features P{P2,P3,P4, ...

[0060] P5}, as shown in formula (1), has dimensions that are 1 / 4, 1 / 8, 1 / 16 and 1 / 32 of the original image size, respectively.

[0061] P{P2,P3,P4,P5}=Backbone(Input) (1);

[0062] Step 2.2, as follows Figure 3 As shown ( Figure 3 In this context, H, W, and C represent the height, width, and channels of the feature map, respectively. The multi-scale features P{P2,P3,P4, ..., P4} of the local details obtained in step 2.1 are then processed.

[0063] P5} is used as input to TransBlock, resulting in the corresponding feature T{T2,T3,T4,T5} (as shown in Equation 2), which aggregates global features and deep semantic information. The feature map size is consistent with the feature map size corresponding to the same index previously. Then, as... Figure 4 As shown, the features T{T2,T3,T4,T5} that aggregate global features and deep semantic information, and the multi-scale features P{P2,P3,P4,P5} obtained according to step 2.1 are used as inputs, so that the above two sets of feature maps are processed by the multi-scale feature dynamic fusion network DFM to obtain the gain features F{F2,F3,F4,F5} (as shown in Equation 3):

[0064] T{T2,T3,T4,T5}=TransBlock(P) (2);

[0065] F{F2,F3,F4,F5}=DFM(P,T) (3);

[0066] Step 2.3: The gain feature F obtained in Step 2.2 is used as the input to the RPN to obtain the accurate rotation target detection box. The feature map of the gain feature F undergoes a 1×1 convolution to adjust the channels, and then SEBlock is used to re-divide the channel weights to obtain the accurate rotation target detection box (x, y, w, h, α, β) (as shown in Equation 4), where x and y represent the coordinates of the anchor box's center point, w and h represent the width and height of the anchor box, and α and β are the offsets of the initial anchor box's top and right center points.

[0067] (x,y,w,h,α,β)=SEBlock(Conv 1×1 (Feature Map) (4);

[0068] Step 2.4, as follows Figure 4 As shown, the gain features from step 2.2 and the detection boxes extracted in step 2.3 are used as inputs, and the information is decoupled through a dual-branch detection head to obtain accurate target detection information. In the detection head, the detection box regression branch uses two convolutions with a stride of 1 and a kernel size of 3×3 to adjust the information, and then the detection boxes are accurately regressed through a fully connected layer (as shown in formula (5)); while the classification branch uses two fully connected layers to decouple the classification information of the feature map, and at the same time fuses the information of the regression branch to achieve accurate classification of the target features (as shown in formula (6)). In the formula, the value of j ranges from 1 to 4, which corresponds to the feature map number after downsampling the original image by 4 times, 8 times, 16 times, and 32 times; DF(A, B) shows the dynamic fusion (DF) process (see Figure 5 ), where A and B represent the regression features and the initial feature map features after two linear layers in the DF operation, and θ j Represents the fusion factor.

[0069] Regression = Linear(Conv) 3×3 (Conv 3×3 (Feature Map))) (5);

[0070]

[0071] Step 3: Use the training set preprocessed in Step 1 to train the model constructed in Step 2, and obtain the trained rotating target detection model based on the bidirectional feature fusion network.

[0072] The model is trained using the training set preprocessed in step 1: the network model obtained in step 2 is constrained using the Smooth L1 loss function and the CE loss function, and then the parameters are updated by backpropagation. After 12 rounds of training, where 1 round refers to training all the preprocessed images once, a trained rotating target detection model based on a bidirectional feature fusion network is finally obtained.

[0073] Step 4: Input the test set images processed in Step 1 into the model trained in Step 3, and output the detection results.

[0074] Example 2

[0075] The specific process of step 2.2 is as follows:

[0076] Step 2.2.1: For the multi-scale features P{P2,P3,P4,P5} extracted from the backbone features, they are input into N TSC (Transformer-Style ConvModule) modules (N=2) in TransBlock to extract features T{T2,T3,...} that aggregate global features and deep semantic information.

[0077] T4, T5} (as shown in formula (7)). The TSC module contains two branching paths. The first branch first flattens (reshapes) the input features, which then pass through a fully connected layer (Linear).

[0078] After obtaining the pairwise relationships between features, the flattened features are reshaped back to their original shape, and then subjected to depthwise separable convolution (DConv) to obtain information-adjusted features A. i (As in formula (8)), the second branch is similar to the first branch, and the feature V is obtained after flattening (Reshape), fully connected layer (Linear) and recovery (Reshape). i (as in formula (9)), then for feature A i With V i Multiplying the aggregated global features and deep semantic features yields Zaggregate. i (as in formula (10)), and then add it to the output feature connected to the residual to obtain the final output feature T. i (as in formula (11)), the value range of i in the above formula is 2 to 5, which corresponds to the sequence number of the multi-scale feature.

[0079] T{T2,T3,T4,T5}=TSC(TSC(P)) (7);

[0080] A i =DConv(Reshape(Linear(Reshape(P i)))) (8);

[0081] V i =Reshape(Linear(Reshape(P)) i ))) (9);

[0082] Z i =A i ⊙V i (10);

[0083]

[0084] Step 2.2.2: Based on the multi-scale features P of local details and the multi-scale features T of global semantic information obtained in Step 2.2.1, these are used as inputs. Each feature is then multiplied by its corresponding dynamic feature weight to achieve appropriate feature selection. Finally, the two features are added together to obtain the gain feature F. i (As shown in Formula 12), the value of i in the above formula ranges from 2 to 5, corresponding to the sequence number of the multi-scale feature; λ is the fusion factor.

[0085]

[0086] The loss function used in step 3 is shown below:

[0087] Smooth L1 Loss Function (Regression Loss): The Smooth L1 loss proposes to set the loss function as a piecewise function, making the loss more robust to outliers, where x1 is the numerical difference between the predicted box and the ground truth box.

[0088]

[0089] CE loss function L CE (Classification Loss): CELows is Cross Entropy Loss, suitable for multi-class classification problems, where y is the label (y∈{0,1}), x2 is the probability of the predicted class output by the softmax function (x2∈[0,1]), n is the number of samples, and the learning objective of the function is to make the output x2 approximate the label y.

[0090]

[0091] Therefore, the total loss function of this invention is:

[0092]

[0093] Example 3

[0094] The rotating target detection network model based on bidirectional feature fusion network constructed in this invention was tested. The parameters were set as follows: batch size during training was set to 2, initial learning rate was 0.005, momentum coefficient was 0.9, weight decay coefficient was 0.0001, 50% of the training images were randomly flipped during training, the input image size was 800×800, and the remaining parameters were the default parameters of the MMRorate open-source tool in OpenMMLab. The final aerial remote sensing image after target detection was obtained. The detection results for small targets are shown below. Figure 6 As shown ( Figure 6 The small targets are outlined in the bold black rectangle. In aerial remote sensing images, small targets are generally considered to be less than 10 pixels in both width and height. Figure 6 Small targets within the model were all detected; targets with large scale differences were detected as follows: Figures 7(a) to 7(d) As shown in Figures 7(a) to 7(d), the targets are outlined in bold black rectangles. In aerial remote sensing images, the industry considers a scale difference of more than twice the size of the target to be a relatively large scale difference, i.e., the length (width) of one target is more than twice the length (width) of another target. Figures 7(a) and 7(b) show that the same target (bridge) at different altitudes was correctly identified. That is, the bridge in Figures 7(a) and 7(b) has a large scale difference in width, but they were still correctly identified. Figures 7(c) and 7(d) show that different targets (various sports fields) at the same altitude were also correctly identified. That is, the basketball court, badminton court, and car in Figure 7(c) have large scale differences in length and width, and the track and field, badminton court, and car in Figure 7(d) have large scale differences in length and width, but they were still correctly identified, demonstrating the excellent effect of the present invention.

Claims

1. A rotating target detection method based on a bidirectional feature fusion network, characterized in that: Specifically, the steps include the following: Step 1: Preprocess the images in the pre-divided training and test sets of the DIOR-R dataset to obtain the preprocessed training and test set images (Input). Step 2: Construct a rotating target detection network model based on a bidirectional feature fusion network. This model is used to perform feature extraction, adaptive feature fusion, and decoupling of bi-branch feature information on the image Input obtained in Step 1. The specific process of step 2 is as follows: Step 2.1: Using ResNet50 as the backbone network, extract multi-scale features P{ of local details. P 2, P 3, P 4, P 5}; Step 2.2, the multi-scale features P{ of the local details obtained in Step 2.1 are... P 2, P 3, P 4, P 5} is used as input to TransBlock to obtain the corresponding feature T{ which aggregates global features and deep semantic information. T 2, T 3, T 4, T 5}, which aggregates global features and deep semantic information into feature T{ T 2, T 3, T 4, T 5} and the multi-scale feature P{ obtained according to step 2.1 P 2, P 3, P 4, P 5} is used as input, so that the above two sets of features T and feature P are passed through the multi-scale feature dynamic fusion network DFM to obtain the gain feature F{ F 2, F 3, F 4, F 5}, which aggregates global features and deep semantic information into feature T{ T 2, T 3, T 4, T 5} and the multi-scale feature P{ obtained according to step 2.1 P 2, P 3, P 4, P 5} is used as input, so that the above two sets of features T and feature P are passed through the multi-scale feature dynamic fusion network DFM to obtain the gain feature F{ F 2, F 3, F 4, F 5}; The specific process of step 2.2 is as follows: Step 2.2.1, extract the multi-scale features P{ from the backbone features. P 2, P 3, P 4, P 5}, the input is fed into two TSC modules in TransBlock to extract the feature T{ which aggregates global features and deep semantic information. T 2, T 3, T 4, T 5}, where the TSC module contains two branching paths. The first branch first flattens and reshapes the input features, then passes them through a fully connected Linear layer to obtain the pairwise relationships between features, and then restores the flattened features to their original shape. Finally, it passes them through a depthwise separable convolution (DConv) to obtain the information-adjusted features A. i The second branch, like the first, goes through flattening and reshaping, a fully connected linear layer, and restoration reshaping to obtain the feature V. i Next, targeting feature A i With V i Multiplying the aggregated global features and deep semantic features yields Zaggregate. i And thus Z i Output characteristics connected to residuals P i The final output feature T is obtained by adding the features together. i ; Step 2.2.2: Multiply the multi-scale features P of local details and the multi-scale features T of global semantic information by their corresponding dynamic feature weights, and then add them together to obtain the gain feature. F i ; Step 2.3: Use the gain feature F obtained in step 2.2 as the input of RPN to obtain the accurate rotating target detection box; Step 2.4: Using the gain features from Step 2.2 and the rotated target detection box extracted in Step 2.3 as input, the information is decoupled through a dual-branch detection head to obtain accurate target detection information; Step 3: Use the training set preprocessed in Step 1 to train the model constructed in Step 2, and obtain the trained rotating target detection model based on the bidirectional feature fusion network. Step 4: Input the test set images preprocessed in Step 1 into the model trained in Step 3 to finally obtain the aerial remote sensing image after target detection.

2. The rotating target detection method based on a bidirectional feature fusion network according to claim 1, characterized in that: The image preprocessing process in step 1 is as follows: the image is randomly flipped, randomly cropped, and color transformed in sequence to finally obtain the preprocessed training set and test set images Input.

3. The rotating target detection method based on a bidirectional feature fusion network according to claim 2, characterized in that: In step 2.1, the network weights pre-trained on the ImageNet dataset are loaded, and then the pre-processed image input is used as the input to the backbone network. P 2. P 3. P 4. P The dimensions are successively 1 / 4, 1 / 8, 1 / 16 and 1 / 32 of the original image size, as shown in formula (1): （1） In step 2.2, the feature T{ that aggregates global features and deep semantic information is obtained. T 2, T 3, T 4, T The process of 5} is shown in formula (2); the gain characteristic F{ is obtained. F 2, F 3, F 4, F The process of 5} is shown in formula (3): （2） （3） In step 2.3, the feature map of the gain feature F undergoes a 1×1 convolution to adjust the channels, and then SEBlock is used to re-divide the channel weights to obtain an accurate rotating target detection box. As shown in formula (4): （4） in, Represents the coordinates of the center point of the anchor frame. Represents the width and height of the anchor frame. This is the offset of the top midpoint and the right midpoint of the initial anchor frame; In step 2.4, the regression branch of the detection head uses two convolutions with a stride of 1 and a kernel size of 3×3 to adjust the information, and then achieves accurate regression of the detection box through a fully connected layer, as shown in formula (5); while the classification branch uses two fully connected layers to decouple the classification information of the feature map, and at the same time fuses the information of the regression branch to achieve accurate classification of the target features, as shown in formula (6), where formula (6) j The value range is 1~4, corresponding to the feature map numbers after downsampling the original image by 4, 8, 16, and 32 times; DF(A,B) shows the dynamic fusion DF process, where A and B represent the regression features and the features of the initial feature map after two linear layers in the DF operation. Representative fusion factor: （5） （6）。 4. The rotating target detection method based on a bidirectional feature fusion network according to claim 3, characterized in that: In step 2.2.1, the aggregated global features and deep semantic information features T{ are obtained. T 2, T 3, T 4, T The process of 5} is shown in the following formula (7); obtain information to adjust feature A i As shown in the following formula (8); the characteristic V is obtained. i The process is shown in formula (9); the aggregated global features and deep semantic features Z are obtained. i The process is shown in formula (10); The final output feature T is obtained. i The process is shown in formula (11): （7） （8） （9） （10） （11） In the above formula i The value range is 2~5, and... P 2. P 3. P 4. P The subscript numbers in 5 correspond to the subscript numbers; Step 2.2.2, obtain the gain characteristics. F i The process is shown in the following formula (12): （12） in, This is the fusion factor.

5. The rotating target detection method based on a bidirectional feature fusion network according to claim 4, characterized in that: The loss function used in step 3 of the training process is as follows: Smooth L1 Loss Function: The Smooth L1 loss method proposes setting the loss function as a piecewise function, where... x 1 represents the numerical difference between the predicted bounding box and the ground truth bounding box: （13） CE loss function: CELouss is the cross-entropy loss, where y It's a tag. x 2 represents the probability of the predicted class output by the softmax function. n It is the number of samples, and the learning objective of the function is to make the output... x 2. Approximate the label y : （14） The total loss function is: （15）。

Images