A training method and corresponding apparatus for a semi-supervised rotating target detection model

Abstract

This invention discloses a training method and corresponding apparatus for a semi-supervised rotating target detection model, comprising: training a second model using labeled samples, and calculating a supervised loss L based on the prediction results of the second model and the sample labels. S Different degrees of enhancement were applied to the unlabeled samples to obtain the first enhanced sample I. w Second Enhanced Sample I s Through the first enhanced sample I W Train the first model and use the second augmented sample I s The second model is trained; based on the prediction results of the first and second models, the adaptive weighted loss L for rotation sensing is obtained respectively. R and global consistency loss L G According to supervised loss L S Adaptive weighted loss L R and global consistency loss L G The first and second models are trained iteratively until the models converge.

Description

Technical Field

[0001] This invention belongs to the field of computer vision, and more specifically, relates to a training method and corresponding apparatus for a semi-supervised rotating target detection model. Background Technology

[0002] Remote sensing target detection is a crucial research topic in remote sensing analysis and Earth observation. It aims to detect targets of interest (such as vehicles and ships) in remote sensing image data and output their category and location. Remote sensing images refer to ground images captured at high and very high altitudes by sensors mounted on platforms such as aircraft and satellites, with resolutions up to half a meter. As a fundamental problem in remote sensing analysis, remote sensing target detection technology has broad and practical application prospects, relying on it in many real-world applications such as urban management, precision agriculture, emergency rescue, and disaster monitoring. Furthermore, remote sensing target detection technology is also the foundation for subsequent tasks such as target tracking, scene classification, and image understanding.

[0003] Existing remote sensing target detection methods primarily employ supervised learning paradigms, relying on large amounts of labeled data. However, for remote sensing scenarios, due to the small scale, dense distribution, and diverse orientations of targets, constructing large-scale labeled datasets requires significant human and material resources. Compared to the high cost of labeling, remote sensing data acquisition is relatively inexpensive, and massive amounts of unlabeled data are readily available. Therefore, researching the use of large amounts of unlabeled data to further improve the performance of remote sensing target detection algorithms has significant application value. Summary of the Invention

[0004] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a training method and corresponding apparatus for a semi-supervised rotating target detection model. This method can simultaneously utilize labeled data and a large amount of unlabeled data to further improve the performance of remote sensing target detection methods, thereby alleviating the technical problems of high labeling costs and difficulty in obtaining labeled data in the field of remote sensing target detection.

[0005] To achieve the above objectives, according to one aspect of the present invention, a method for training a semi-supervised rotating target detection model is provided, comprising:

[0006] The first and second models are constructed based on a predetermined rotating target detector;

[0007] A predetermined number of labeled samples are randomly sampled from the labeled image data. The second model is trained using these labeled samples. The supervised loss L is calculated based on the prediction results of the second model and the sample labels. S ;

[0008] A predetermined number of unlabeled samples are randomly sampled from the unlabeled image data, and the unlabeled samples are enhanced to different degrees to obtain the first enhanced sample I. w Second Enhanced Sample I s Through the first enhanced sample I w The first model is trained, and the second augmented sample I is used. s The second model is trained;

[0009] Based on the prediction results of the first model and the second model, the adaptive weighted loss L for rotation sensing is obtained respectively. R and global consistency loss L G ;

[0010] According to the supervised loss L S Adaptive weighted loss L R and global consistency loss L G The first and second models are trained iteratively until the models converge.

[0011] Furthermore, the construction of the first and second models based on the predetermined rotating target detector includes:

[0012] Select a rotating target detector, and construct the first model and the second model according to the structure of the rotating target detector, so as to ensure that the structure of the first model is the same as that of the second model, and the parameters of the first model and the second model are not shared.

[0013] The parameters of the second model are updated by backpropagation through a neural network, and the parameters of the first model are updated by the second model using an exponential moving average.

[0014] Furthermore, based on the prediction results of the first model and the prediction results of the second model, the adaptive weighted loss L for rotation sensing is obtained respectively. R and global consistency loss L G include:

[0015] The first model is obtained based on the first augmented sample I. w The first prediction matrix P obtained is T And obtain the second model based on the second enhanced sample I s The second prediction matrix P obtained from the prediction S Wherein, the first prediction matrix P T and the second prediction matrix P S Each element contains a prediction result, and each prediction result includes bounding box regression and bounding box classification.

[0016] For the first prediction matrix P T The prediction results are processed to obtain prediction boxes, which are then projected onto the first prediction matrix P. T A predetermined number of predicted samples are selected within the projected area to form a pseudo-label set S. pse ={y i}, i = 1, 2, ..., n, where n is the number of pseudo-labels;

[0017] In the second prediction matrix P S Select samples from the corresponding positions as the prediction set S prd ={x i}, i = 1, 2, ..., n;

[0018] Each pseudo-label y i The corresponding prediction result x i Construct a pseudo-label-prediction pair z i =(x i ,y i The loss is calculated for each pseudo-label-prediction pair.

[0019] Loss for each pseudo-label-prediction pair Constructing an adaptive weighted loss L based on rotation sensing R .

[0020] Furthermore, the loss for each pseudo-label-prediction pair Constructing an adaptive weighted loss L based on rotation sensing R include:

[0021] Determine the pseudo-label y i Predicted rotation angle Determine the prediction result x i Predicted rotation angle

[0022] Based on the predicted rotation angle and predict rotation angle Calculate the modulation factor According to the modulation factor Loss for each pseudo-label-prediction pair We perform weighted summation to obtain the adaptive weighted loss L for rotation sensing. R ;

[0023] Adaptive weighted loss L R The calculation formula is:

[0024]

[0025] Modulation factor The calculation formula is:

[0026]

[0027]

[0028] Here, α is set as a hyperparameter.

[0029] Furthermore, based on the prediction results of the first model and the prediction results of the second model, the adaptive weighted loss L for rotation sensing is obtained respectively. R and global consistency loss L G include:

[0030] Let the category prediction score of the first model be s. t ∈R n×K , and the category prediction score of the second model is s s ∈R n×K Where n is the number of pseudo-label-prediction pairs, and K is the number of classes;

[0031] The predicted score based on the category of the first model is s. t The category prediction score of the second model is s s Construct the optimal matching cost formula, find its minimum value, and then obtain the global consistency loss L. G .

[0032] Further, the score predicted based on the category of the first model is s t The category prediction score of the second model is s s Construct the optimal matching cost formula, find its minimum value, and then obtain the global consistency loss L. G include:

[0033] Calculate the category prediction score s t and category prediction scores s The global distribution is given by the following formula:

[0034]

[0035]

[0036] in, This represents the index of the category with the highest confidence level in the i-th pseudo-label;

[0037] The global distribution of the first model respectively and the global distribution of the second model After normalization, the discrete distribution of the first model is obtained. and the discrete distribution of the second model The calculation formula is as follows:

[0038]

[0039]

[0040] Where ||·||1 represents the 1-norm;

[0041] Based on the discrete distribution of the first model and the discrete distribution of the second model Construct the optimal matching cost formula;

[0042] Find the minimum value of the optimal matching cost formula, and then obtain the global consistency loss L. G .

[0043] Furthermore, the discrete distribution based on the first model and the discrete distribution of the second model The global consistency loss L is obtained. G include:

[0044] The optimal matching cost formula between two distributions is established according to the following calculation formula:

[0045]

[0046] in, Represent all possible transport schemes from the first model to the second model, C∈R n×n This represents the cost matrix between the two.

[0047] Solving the optimal matching cost formula yields an approximate solution λ. * and μ * This leads to the global consistency loss L. G ,in, λ * and μ * This is an approximate solution obtained through iteration.

[0048] Furthermore, the training method also includes:

[0049] Calculate the global consistency loss L G Relative to d s The partial derivative of the target gradient is obtained, and the target gradient is used to constrain the global consistency between the first and second models. The target gradient is used in the backpropagation process of the neural network.

[0050] Furthermore, the supervised loss L... S Adaptive weighted loss LR and global consistency loss L G Iterative training of the first and second models until convergence includes:

[0051] Calculate the total loss function L for training according to the following formula, and iteratively train the first and second models based on the total loss function L until the total loss function L reaches the predetermined value.

[0052] L = w s L S +w u L U L U =L R +L G ;

[0053] Among them, w s and w u These represent the weights for supervised and unsupervised losses, respectively.

[0054] According to another aspect of the present invention, a training apparatus for a semi-supervised rotating target detection model is provided, comprising at least one processor and a memory, wherein the at least one processor and the memory are connected via a data bus, and the memory stores instructions that can be executed by the at least one processor, wherein the instructions, after being executed by the processor, are used to complete the training method for the semi-supervised rotating target detection model described in the first aspect.

[0055] Overall, the technical solutions conceived by this invention have the following beneficial effects compared with the prior art: the embodiments of this invention can mine semantic knowledge from a large amount of unlabeled data, and together with labeled data, improve the performance of existing rotating target detection algorithms. In addition, based on unlabeled data, an adaptive weighted loss and a global consistency loss for rotation perception are proposed. During the model training process, the adaptive weighted loss and the global consistency loss are taken into account together, which can improve the accuracy of model detection.

[0056] Furthermore, to address the issue of diverse target orientations, the system cleverly utilizes the key characteristic of the target's rotation angle to dynamically measure the distance between the prediction results of the first and second models. By using a weighted approach, this prediction distance is more flexibly reflected in the calculation of the loss function, helping the network to discover hard-to-distinguish samples.

[0057] Furthermore, to address the issue of small target size and dense distribution, a method is proposed to construct the constraint relationship between the first and second models from a global perspective. This innovatively models the global constraint as an optimal allocation problem, capturing the many-to-many relationship between the prediction results of the first and second models, and characterizing global correlations, such as the quantity and density of the target distribution, thus introducing richer prior information into the model. Attached Figure Description

[0058] Figure 1 This is a schematic flowchart of a training method for semi-supervised rotating target detection proposed in an embodiment of the present invention;

[0059] Figure 2 This is a schematic diagram of the overall framework of a semi-supervised rotating target detection model proposed in an embodiment of the present invention;

[0060] Figure 3 This is a schematic diagram illustrating the results of unlabeled data after different data augmentations, as proposed in an embodiment of the present invention.

[0061] Figure 4 This is provided by the embodiments of the present invention. Figure 1 A detailed flowchart of step 40 is shown below;

[0062] Figure 5 This is a schematic diagram illustrating the visualization effect of projecting the prediction results of the first model according to an embodiment of the present invention.

[0063] Figure 6 This is a schematic diagram illustrating the effect of global consistency constraints provided in an embodiment of the present invention;

[0064] Figure 7 This is a schematic diagram of the training model of a semi-supervised rotating target detection model provided in an embodiment of the present invention. Detailed Implementation

[0065] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0066] The following description, in conjunction with the accompanying drawings and specific embodiments, further illustrates a semi-supervised method for multi-orientation target detection according to the present invention. It should be noted that the following description is exemplary and intended to further illustrate the present invention, rather than limiting the implementation.

[0067] Example 1:

[0068] The method in this embodiment is to use a deep neural network to extract features and complete training and inference to obtain a trained model, and then use the trained model for detection. This embodiment mainly explains how to train to obtain the required model.

[0069] This invention can mine semantic knowledge from a large amount of unlabeled data and combine it with labeled data to improve the performance of existing rotating target detection algorithms. In addition, based on the lack of standard data, this embodiment also proposes two loss functions based on consistency constraints, hereinafter referred to as rotation-aware adaptive weighted loss and global consistency loss. During the model training process, the adaptive weighted loss and global consistency loss are taken into account together, which can improve the accuracy of model detection.

[0070] Combination Figure 1 and Figure 2 This embodiment provides a training method for a semi-supervised rotating target detection model, including:

[0071] Step 10: Construct the first model and the second model based on the predetermined rotating target detector.

[0072] In this embodiment, a rotating target detector is selected, and the first model and the second model are constructed according to the structure of the rotating target detector to ensure that the structure of the first model is the same as that of the second model, and the parameters of the first model and the second model are not shared. During the entire training process, the parameters of the second model are updated by backpropagation of the neural network, and the parameters of the first model are updated by the second model by exponential moving average (EMA).

[0073] The aforementioned EMA update will begin after the second model has been trained a certain number of times on labeled data. Specifically, the training process in this embodiment is divided into three stages: a) training the second model using only labeled image data; b) still performing only supervised training, but simultaneously updating the parameters of the first model through EMA; c) while performing supervised training, training the second model using unlabeled image data, and continuously updating the first model through EMA. Specific training methods are described below.

[0074] In an optional embodiment, the rotating target detector is the rotate-FCOS rotating target detector. The second model is trained based on 10% of the DOTA-V1.5 data as labeled data, according to step 20 below; the remaining data is trained as unlabeled data, and the first and second models are trained according to steps 30 and 40 below.

[0075] Step 20: Randomly sample a predetermined number of labeled samples from the labeled image data, train the second model using the labeled samples, and calculate the supervised loss L based on the prediction results of the second model and the sample labels. S .

[0076] In this embodiment, a certain number of samples are randomly sampled from the labeled image data for the supervised training phase. These labeled samples are used to train a second model, following a given loss function for a rotating object detector. The supervised loss L is calculated based on the prediction results of the second model and the sample labels. S .

[0077] In an optional embodiment, with supervised loss L S In the calculation process, the loss function of rotate-FCOS is followed. The supervised loss calculated includes three parts: classification, regression and centerness, which are calculated by focal loss function, rotated intersection-union loss function and cross-entropy loss function, respectively.

[0078] Step 30: Randomly sample a preset number of unlabeled samples from the unlabeled image data, and enhance the unlabeled samples to different degrees to obtain the first enhanced sample I. w Second Enhanced Sample I s Through the first enhanced sample I w The first model is trained, and the second augmented sample I is used. s The second model is then trained.

[0079] A certain number of samples are randomly sampled from unlabeled image data for the unsupervised stage. These unlabeled samples are then subjected to strong data augmentation and weak data augmentation, respectively, and the augmented samples are denoted as I. s and I w Among them, the first enhanced sample I w Provided to the first model, the second augmented sample I s Provided to the second model.

[0080] Specifically, weak data augmentation includes random flipping (horizontal, vertical, and diagonal) and random scaling; the aforementioned strong data augmentation refers to relatively weak data augmentation, which further incorporates random color transformation, random image grayscale transformation, random blurring, and random masking. Visualizations after different data augmentations are shown below. Figure 3 As shown.

[0081] Wherein, the first enhanced sample I w The enhancement level is lower than that of the second enhanced sample I. s The degree of enhancement. The first enhanced sample I w The second augmented sample I is a sample after weak data augmentation. s This is a sample after strong data augmentation.

[0082] Step 40: Based on the prediction results of the first model and the prediction results of the second model, obtain the adaptive weighted loss L for rotation sensing. R and global consistency loss L G .

[0083] Step 50: Based on the supervised loss L S Adaptive weighted loss L R and global consistency loss L G The first and second models are trained iteratively until the models converge.

[0084] In this embodiment, both the first model and the second model are trained simultaneously. During training, the second model updates the parameters of the first model using an exponential moving average, thereby assisting in updating the first model to obtain the desired first model. In the usage phase, the first model is used for detection.

[0085] Compared with current methods, the beneficial effects of the embodiments of the present invention include at least the following:

[0086] This paper proposes a semi-supervised learning approach for rotating target detection, especially for remote sensing imagery scenes. Compared with general semi-supervised target detection frameworks, this approach fully considers the characteristics of dense target distribution and diverse orientations in remote sensing scenes. It also innovatively makes a series of targeted designs, which effectively improves the performance of existing rotating target detectors and reduces the annotation burden of rotating target detection.

[0087] Combination Figure 2 and Figure 4 To further explain step 40 above, the adaptive weighted loss L is obtained as follows: R :

[0088] Step 401: Obtain the first model based on the first augmented sample I w The first prediction matrix P obtained is T And obtain the second model based on the second enhanced sample I s The second prediction matrix P obtained from the prediction S Wherein, the first prediction matrix P T and the second prediction matrix P S Each element contains a prediction result, and each prediction result includes bounding box regression and bounding box classification.

[0089] The first and second models respectively predict the two-dimensional output matrix P. T ,P S ∈R h×w, where h and w are the values ​​of the original image size scaled up by a certain ratio, and each element in the matrix contains three prediction results: regression, classification, and centerness. The first prediction matrix P T Second prediction matrix P S The dimensions are consistent.

[0090] Step 402: For the first prediction matrix P T The prediction results are processed to obtain prediction boxes, which are then projected onto the first prediction matrix P. T A predetermined number of predicted samples are selected within the projected area to form a pseudo-label set S. pse ={y i}, i = 1, 2, ..., n, where n is the number of pseudo-labels.

[0091] Step 403: In the second prediction matrix P S Select samples from the corresponding positions as the prediction set S prd ={x i}, i = 1, 2, ..., n.

[0092] In this process, the prediction results of the first model are post-processed, including nonmaximum suppression, to obtain the final prediction box, which is then projected onto the first prediction matrix P of the first model through a linear transformation. T 25% of the predicted samples are randomly selected within the projection area to form a dense pseudo-label set S. pse ={y i}, i = 1, 2, ..., n, where n is the number of pseudo-labels, and simultaneously, in the second prediction matrix P of the second model S Samples are selected at the corresponding positions as the prediction set S prd ={x i}, i = 1, 2, ..., n. The visualization of the screening process is as follows: Figure 5 As shown, the rectangles represent the final detection results, the highlighted areas indicate high confidence in the prediction results, and the dots represent filtered sample examples. This filtering strategy effectively suppresses interference from erroneous predictions in the second model.

[0093] Step 404: Transfer each pseudo-label y i The corresponding prediction result x i Construct a pseudo-label-prediction pair z i =(x i ,y i The loss is calculated for each pseudo-label-prediction pair.

[0094] The specific process is as follows: Based on the correspondence of pixel positions, each pseudo-label y is... i The corresponding prediction result x iConstruct a pseudo-label-prediction pair z i =(x i ,y i The loss is calculated for each pseudo-label-prediction pair. The L1 loss function is used for the regression part, and the binary cross-entropy loss function is used for the other parts.

[0095] Step 405: Loss for each pseudo-label-prediction pair Constructing an adaptive weighted loss L based on rotation sensing R .

[0096] In constructing the rotation-aware adaptive weighted loss L R In the process, to address the issue of diverse target orientations, the key characteristic of the target's rotation angle is cleverly utilized to dynamically measure the distance between the prediction results of the first and second models. By using a weighted approach, this prediction distance is more flexibly reflected in the calculation of the loss function, helping the network to discover hard-to-distinguish samples.

[0097] The specific process is as follows: Determine the pseudo-label y i Predicted rotation angle Determine the prediction result x i Predicted rotation angle Based on the predicted rotation angle and predict rotation angle Calculate the modulation factor According to the modulation factor Loss for each pseudo-label-prediction pair We perform weighted summation to obtain the adaptive weighted loss L for rotation sensing. R .

[0098] Among them, the adaptive weighted loss L R The calculation formula is:

[0099]

[0100] Modulation factor The calculation formula is:

[0101]

[0102]

[0103] Here, α is set as a hyperparameter.

[0104] In an optional embodiment, α is 1.

[0105] To address the issue of small target size and dense distribution, this paper proposes constructing a constraint relationship between the first and second models from a global perspective. It innovatively models the global constraint as an optimal allocation problem, capturing the many-to-many relationship between the prediction results of the first and second models, and characterizing global correlations such as the quantity and density of target distribution, thus introducing richer prior information into the model. Further explanation of step 40 is provided, in which the global consistency loss L is obtained as follows: G :

[0106] First, let s be the category prediction score of the first model. t ∈R n×K , and the category prediction score of the second model is s s ∈R n×K Where n is the number of pseudo-label-prediction pairs, and K is the number of classes;

[0107] Then, the predicted score is s based on the category of the first model. t The category prediction score of the second model is s s Construct the optimal matching cost formula, find its minimum value, and then obtain the global consistency loss L. G Specifically, calculate the category prediction score s. t and category prediction scores s The global distribution is given by the following formula:

[0108]

[0109]

[0110] in, This represents the index of the category with the highest confidence level in the i-th pseudo-label;

[0111] Furthermore, the global distribution of the first model is analyzed separately. and the global distribution of the second model After normalization, the discrete distribution of the first model is obtained. and the discrete distribution of the second model The calculation formula is as follows:

[0112]

[0113]

[0114] Where ||·||1 represents the 1-norm;

[0115] Then, based on the discrete distribution of the first model and the discrete distribution of the second model Construct the optimal matching cost formula; solve for the minimum value of the optimal matching cost formula, and then obtain the global consistency loss L. G .

[0116] In an optional embodiment, the optimal matching cost formula between the two distributions is established according to the following calculation formula:

[0117]

[0118] in, Represent all possible transport schemes from the first model to the second model, C∈R n×n This represents the cost matrix between the two.

[0119] Solving the optimal matching cost formula yields an approximate solution λ. * and μ * This leads to the global consistency loss L. G ,in, λ * and μ * This is an approximate solution obtained through iteration.

[0120] In practical applications, the training method further includes: calculating the global consistency loss L. G Relative to d s The partial derivative of the target gradient is obtained, and the target gradient is used to constrain the global consistency between the first and second models. The target gradient is used in the backpropagation process of the neural network.

[0121] To better measure the correlation between the prediction results of the first and second models, a cost matrix C∈R is constructed based on the spatial distance and predicted value distance between matching pairs. n×n :

[0122] A cost matrix C∈R is constructed based on the spatial distance and predicted distance between matching pairs. n×n :

[0123]

[0124]

[0125]

[0126] in, and These represent the two-dimensional pixel coordinates of the i-th sample in the first model and the j-th sample in the second model, respectively.

[0127] In this embodiment, step 50 specifically includes: calculating the total loss function L of the training according to the following formula, and iteratively training the first model and the second model based on the total loss function L until the total loss function L reaches a predetermined value.

[0128] The specific calculation method for the total loss function L is as follows:

[0129] L = w s L s +w u L U L U =L R +L G ;

[0130] Among them, w s and w u These represent the weights for supervised and unsupervised losses, respectively.

[0131] In an optional embodiment, w s and w u All values ​​are set to 1, and the unsupervised loss L is... U It consists of two parts: rotation-aware adaptive weighted loss and global consistency loss.

[0132] After training the model in the aforementioned manner, the first model is directly used for evaluation, and the detection results of rotate-FCOS under supervised conditions are compared. The detection results of this embodiment of the invention have a lower rate of missed detections, false detections, and inaccurate detections compared to the detection results of other methods.

[0133] Example 2:

[0134] Based on the training method of Embodiment 1 above, this embodiment uses rotate-FCOS as the detector and presents the performance achieved on the mainstream remote sensing target detection dataset DOTA-V1.5.

[0135] The DOTA-V1.5 dataset contains 2,806 large-format remote sensing images with 402,089 labeled rotating targets. This dataset is further divided into three subsets: DOTA-V1.5-Train, DOTA-V1.5-Val, and DOTA-V1.5-Test, containing 1,411, 458, and 937 images respectively. The dataset covers 16 target categories: aircraft, baseball fields, bridges, athletic fields, small vehicles, large vehicles, ships, badminton courts, basketball courts, storage tanks, football fields, traffic roundabouts, ports, swimming pools, helicopters, and container cranes.

[0136] To fully verify the effectiveness of this embodiment, it includes two experimental settings: partially labeled data and full labeled data. For the partially labeled data setting, this embodiment randomly samples 10%, 20%, and 30% of the data from DOTA-V1.5-Train as labeled samples, using the remaining data as unlabeled samples, and evaluates the results on DOTA-V1.5-Val. For the full labeled data setting, this embodiment uses DOTA-V1.5-Train as the labeled sample and DOTA-V1.5-Test as the unlabeled sample, evaluating the results on DOTA-V1.5-Val.

[0137] The evaluation metric used is mAP, where the threshold for crossover-union ratio is 0.5.

[0138] With partially labeled data, the evaluation results of this embodiment are shown in Table 1. Compared with the supervised baseline model results, this embodiment can effectively mine the potential semantic information in unlabeled data and effectively improve the target detection performance of the model.

[0139] Table 1 shows the results under partially labeled data settings.

[0140]

[0141] With all labeled data set, the evaluation results of this embodiment are shown in Table 2. This shows that this embodiment can further improve the performance ceiling of the existing detection model without introducing additional labeling costs, and effectively alleviate the labeling pressure in the field of rotating target detection.

[0142] Table 2 shows the results under the setting of full labeled data.

[0143] method mAP Δ Baseline model 65.46 - This embodiment 67.70 +2.24

[0144] Example 3:

[0145] The effectiveness of the two consistency constraints proposed in Example 1 is analyzed. Under experimental settings with partially labeled data, the performance of the adaptive weighted loss (local constraint) and the global consistency loss (global constraint) proposed in this example with and without the use of these constraints is compared. The results are shown in Table 3. The results show that both constraints improve performance under different settings, and the optimal performance is achieved when both constraints are used simultaneously.

[0146] also, Figure 6The visualization results further demonstrate the effectiveness of the Global Consistency (GC) constraint. In the scatter plot, the vertical axis represents the consistency between pseudo-label and prediction pairs; lower values ​​indicate closer similarity between the predictions of the first and second models. The horizontal axis represents the index of the pseudo-label and prediction pairs. The visualization results show that introducing the GC constraint helps improve the consistency between the first and second models, thereby improving the model's detection performance and reducing false negatives and missed detections.

[0147] Table 3 Comparison of the effects of the two consistency constraints

[0148]

[0149] Example 4:

[0150] Based on the training method for the semi-supervised rotating target detection model provided in Embodiment 1 above, this invention also provides a training apparatus for a semi-supervised rotating target detection model that can be used to implement the above method, such as... Figure 7 The diagram shown is a schematic representation of the device architecture according to an embodiment of the present invention. The device in this embodiment includes one or more processors 21 and a memory 22. Figure 7 Take a processor 21 as an example.

[0151] The processor 21 and the memory 22 can be connected via a bus or other means. Figure 7 Taking the example of a connection between China and Israel via a bus.

[0152] The memory 22 is a non-volatile computer-readable storage medium serving as a training method for a semi-supervised rotating target detection model. It can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules, such as the training method for the semi-supervised rotating target detection model in Embodiment 1. The processor 21 executes various functional applications and data processing of the device by running the non-volatile software programs, instructions, and modules stored in the memory 22, thereby implementing the training method for the semi-supervised rotating target detection model in the embodiment.

[0153] The memory 22 may include high-speed random access memory and non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory 22 may optionally include memory remotely located relative to the processor 21, and these remote memories may be connected to the processor 21 via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0154] The program instructions / modules are stored in the memory 22 and, when executed by one or more processors 21, execute the training method of the semi-supervised rotating target detection model in the above embodiments.

[0155] Those skilled in the art will understand that all or part of the steps in the various methods of the embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, which may include: read-only memory (ROM), random access memory (RAM), magnetic disk or optical disk, etc.

[0156] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A training method for a semi-supervised rotating target detection model, characterized in that, include: The first and second models are constructed based on a predetermined rotating target detector; A predetermined number of labeled samples are randomly sampled from the labeled image data. The second model is trained using these labeled samples. The supervised loss is calculated based on the prediction results of the second model and the sample labels. ; A predetermined number of unlabeled samples are randomly sampled from unlabeled image data, and the unlabeled samples are enhanced to different degrees to obtain the first enhanced sample. Second Enhanced Sample Through the first enhanced sample Train the first model and use the second augmented samples. The second model is trained; Based on the prediction results of the first model and the second model, the adaptive weighted loss for rotation perception is obtained respectively. and global consistency loss ;Including: obtaining the first model based on the first augmented sample The first prediction matrix obtained And obtain the second model based on the second augmented sample. The second prediction matrix obtained The first prediction matrix and the second prediction matrix Each element of the first prediction matrix contains a prediction result, and each prediction result includes bounding box regression and bounding box classification; The prediction results are processed to obtain prediction boxes, which are then projected onto the first prediction matrix. A predetermined number of predicted samples are selected within the projected area to form a pseudo-label set. ,in, The number of pseudo-labels; in the second prediction matrix Samples are selected from the corresponding positions to form the prediction set. ; each pseudo-tag Corresponding prediction results Construct a pseudo-label-prediction pair And calculate the loss for each pseudo-label-prediction pair. ; Identify pseudo-labels Predicted rotation angle Determine the prediction results Predicted rotation angle ; Based on the predicted rotation angle and predict rotation angle Calculate the modulation factor According to the modulation factor Loss for each pseudo-label-prediction pair We perform weighted summation to obtain the adaptive weighted loss for rotation sensing. Adaptive weighted loss The calculation formula is: Modulation factor The calculation formula is: in, Set as hyperparameter According to the supervised loss Adaptive weighted loss and global consistency loss The first and second models are trained iteratively until the models converge.

2. The training method for the semi-supervised rotating target detection model according to claim 1, characterized in that, The construction of the first and second models based on the predetermined rotating target detector includes: Select a rotating target detector, and construct the first model and the second model according to the structure of the rotating target detector, so as to ensure that the structure of the first model is the same as that of the second model, and the parameters of the first model and the second model are not shared. The parameters of the second model are updated by backpropagation through a neural network, and the parameters of the first model are updated by the second model using an exponential moving average.

3. The training method for the semi-supervised rotating target detection model according to claim 1, characterized in that, The adaptive weighted loss for rotation perception is obtained based on the prediction results of the first model and the prediction results of the second model. and global consistency loss include: The category prediction score of the first model is set to And the category prediction score of the second model is ,in, K represents the number of pseudo-label-prediction pairs, and K represents the number of classes. The predicted score based on the category of the first model is: The category prediction scores of the second model are Construct the optimal matching cost formula, find its minimum value, and then obtain the global consistency loss. .

4. The training method for the semi-supervised rotating target detection model according to claim 3, characterized in that, The predicted score based on the category of the first model is The category prediction scores of the second model are Construct the optimal matching cost formula, find its minimum value, and then obtain the global consistency loss. include: Calculate category prediction scores and category prediction scores The global distribution is given by the following formula: in, Indicates the first The index of the category with the highest confidence among the pseudo-labels; The global distribution of the first model respectively and the global distribution of the second model After normalization, the discrete distribution of the first model is obtained. and the discrete distribution of the second model The calculation formula is as follows: in, Represents the 1-norm; Based on the discrete distribution of the first model and the discrete distribution of the second model Construct the optimal matching cost formula; Find the minimum value of the optimal matching cost formula, and then obtain the global consistency loss. .

5. The training method for the semi-supervised rotating target detection model according to claim 4, characterized in that, The discrete distribution based on the first model and the discrete distribution of the second model Gaining global consistency loss include: The optimal matching cost formula between two distributions is established according to the following calculation formula: ； in, This represents all possible transmission schemes that map from the first model to the second model. This represents the cost matrix between the two. Solving the optimal matching cost formula yields an approximate solution. and This leads to a loss of global consistency. , and This is an approximate solution obtained through iteration.

6. The training method for the semi-supervised rotating target detection model according to claim 5, characterized in that, The training method also includes: Calculate global consistency loss Compared to The partial derivative of the target gradient is obtained, and the target gradient is used to constrain the global consistency between the first and second models. The target gradient is used in the backpropagation process of the neural network.

7. The training method for the semi-supervised rotating target detection model according to any one of claims 1 to 6, characterized in that, The supervised loss Adaptive weighted loss and global consistency loss Iterative training of the first and second models until convergence includes: Calculate the total loss function L for training according to the following formula, and iteratively train the first and second models based on the total loss function L until the total loss function L reaches the predetermined value. ； in, and These represent the weights for supervised and unsupervised losses, respectively.

8. A training device for a semi-supervised rotating target detection model, characterized in that, It includes at least one processor and a memory, which are connected via a data bus. The memory stores instructions that can be executed by the at least one processor. After being executed by the processor, the instructions are used to complete the training method of the semi-supervised rotating target detection model according to any one of claims 1-7.

Images