Aerial image target detection method based on classification-positioning double-branch interactive distillation

Abstract

The application discloses an aerial image target detection method based on classification-positioning double-branch interactive distillation, which comprises a pre-training teacher network, a student network, an interactive distillation region selection module and a self-adaptive generation distillation module. The interactive distillation region selection module is constructed to enable the knowledge of the pre-training teacher network and the student network detection head to guide each other, and the optimal distillation region is dynamically selected. The self-adaptive generation distillation module is integrated into the classification branch of the student network detection head, the output feature layer of the classification branch is masked, the output feature layer of the pre-training teacher network is adaptively reconstructed, and the learning ability of the student network is further enhanced. Finally, the KL divergence loss is introduced to constrain the distillation region. The application realizes the interaction between the classification and positioning tasks in the distillation process of the pre-training teacher network and the student network, effectively improves the detection precision and speed of the student network while reducing the parameter quantity.

Description

Technical Field

[0001] This invention relates to the field of drone aerial image technology, and more specifically, to a target detection method for aerial images based on classification-localization bi-branch interactive distillation. Background Technology

[0002] Aerial image target detection is an important research area in remote sensing, widely used in military reconnaissance, disaster monitoring, and urban planning. Traditional high-precision aerial image target detection models often come at the cost of increased computational load and parameter count, while edge computing platforms such as satellites and drones have limited memory and computational resources. Therefore, while pursuing high detection accuracy, it is essential to lightweight the model to enable deployment on mobile platforms.

[0003] To overcome these problems, many model lightweighting methods have emerged. Knowledge distillation is a commonly used model compression technique that approximates the performance of a pre-trained, complex teacher network by training a lightweight student network. Through knowledge distillation, the student network can significantly reduce its computational resource requirements with minimal performance loss, making it more suitable for deployment in practical applications.

[0004] In knowledge distillation, not all predictions from the pre-trained teacher network contribute to the learning process of the student network; inappropriate learning areas may even reduce learning effectiveness. The detection head of an object detector typically consists of two independent, parallel branches: a classification branch and a localization branch. Due to the high resolution, complex backgrounds, and multi-scale targets in aerial images, this complexity increases the difficulty of classification and localization tasks, leading to significant misalignment between the two independent branches in understanding and predicting targets. Therefore, cross-firing and distilling the information from the classification and localization branches during knowledge distillation can help the student network better learn the knowledge from the pre-trained teacher network's detection head. Summary of the Invention

[0005] This invention proposes an aerial image target detection method based on classification-localization bi-branch interactive distillation, which effectively mines the knowledge contained in the detection head. By combining interactive learning and adaptive generative distillation, the robustness and detection performance of the model are effectively enhanced, while reducing the computational load of the model.

[0006] The technical solution for implementing this invention is as follows: a target detection method for aerial images based on classification-localization bi-branch interactive distillation, comprising the following steps:

[0007] Step S1: Download the aerial image detection dataset DIOR-R, randomly cut the image into 800×800 blocks, randomly divide the blocks of uniform size into training and test datasets in a 7:3 ratio, perform data augmentation on the training dataset to form the teacher-student learning network training dataset, and proceed to step S2.

[0008] Step S2: Use the teacher-student learning network training dataset to pre-train the teacher network to obtain the pre-trained teacher network, and then proceed to step S3.

[0009] Step S3: Construct a teacher-student learning network using a pre-trained teacher network, student network, adaptive generation distillation module, and interactive distillation region selection module, then proceed to step S4.

[0010] Step S4: Input the teacher-student learning network training dataset into the teacher-student learning network for training, extract the multi-scale feature layers of the pre-trained teacher network and student network and their corresponding output prediction values; construct an interactive distillation region selection module to dynamically select suitable classification distillation regions and localization distillation regions for distillation; construct an adaptive generation distillation module to mask the output feature layer of the student network classification branch, so that it adaptively reconstructs the output feature layer of the pre-trained teacher network, fixes the parameters of the pre-trained teacher network, and updates the parameters of the student network through the overall loss function of the student network, finally obtaining the trained student network, and proceeding to step S5.

[0011] Step S5: Input the test dataset into the trained student network for detection, merge the detection results of each block image, output the category and location of all targets in the test dataset in the image, and obtain the detection accuracy of the trained student network.

[0012] Compared with existing technical solutions, the significant advantages of this invention are:

[0013] (1) Compared with existing aerial image target detection methods, this invention proposes an aerial image target detection method based on classification-localization bi-branch interactive distillation, which fully explores the knowledge contained in the detection head, improves the detection performance of the student network, and effectively reduces the number of model parameters and computational load, thereby enhancing the robustness and generalization ability of the student network.

[0014] (2) This invention proposes an interactive distillation region selection module to dynamically select the best classification distillation region and localization distillation region for the detection head of the student network. This addresses the problem of insufficient interaction between classification and regression branches during the distillation process and significantly improves distillation efficiency. The integration of an adaptive generation distillation module into the detection head of the student network enables the student network to generate output predictions more flexibly and autonomously, rather than completely imitating the output of the teacher network. This promotes the initiative and creativity of the student network in the learning process and enhances the learning ability of the student network. Attached Figure Description

[0015] Figure 1 This is a network model diagram of the aerial image target detection method based on classification-localization dual-branch interactive distillation of the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0017] Combination Figure 1 The aerial image target detection method based on classification-localization bi-branch interactive distillation described in this invention includes the following steps:

[0018] Step S1: Download the aerial image detection dataset DIOR-R, randomly cut the image into 800×800 blocks, randomly divide the blocks of uniform size into training and test datasets in a 7:3 ratio, perform data augmentation on the training dataset to form the teacher-student learning network training dataset, and proceed to step S2.

[0019] Step S2: Use the teacher-student learning network training dataset to pre-train the teacher network to obtain the pre-trained teacher network, and then proceed to step S3.

[0020] Step S3: Construct a teacher-student learning network using a pre-trained teacher network, student network, adaptive generation distillation module, and interactive distillation region selection module, then proceed to step S4.

[0021] Step S4: Input the teacher-student learning network training dataset into the teacher-student learning network for training. Extract the multi-scale feature layers and their corresponding output prediction values ​​of the pre-trained teacher network and student network. Construct an interactive distillation region selection module to dynamically select suitable classification distillation regions and localization distillation regions for distillation. Construct an adaptive generation distillation module to mask the output feature layer of the student network's classification branch, enabling it to adaptively reconstruct the output feature layer of the pre-trained teacher network. Fix the parameters of the pre-trained teacher network and update the parameters of the student network through the overall loss function of the student network to finally obtain the trained student network, as detailed below:

[0022] In result-based distillation algorithms, not all predictions will promote the learning process of the student network. Inappropriate learning regions may reduce the learning effect. Therefore, an interactive distillation region selection module is constructed to select the optimal distillation region for the detection head of the student network, which can effectively improve the learning ability of the student network.

[0023] First, all output features of the pre-trained teacher network and student network detector heads are extracted, and an interactive distillation region selection module is constructed to dynamically select appropriate classification distillation regions and localization distillation regions for distillation.

[0024] The interactive distillation region selection module includes a categorized main distillation region, a positioned main distillation region, a positioned auxiliary categorized distillation region, and a categorized auxiliary positioned distillation region.

[0025] The primary distillation region for classification and the primary distillation region for localization are directly determined by label assignment, i.e., the location of the positive sample during training; the localization-assisted classification distillation region uses the localization branch of the pre-trained teacher network to guide the selection of the classification distillation region of the student network; the classification-assisted localization distillation region uses the classification branch of the pre-trained teacher network to guide the selection of the localization distillation region of the student network.

[0026] For the selection of the categorized master distillation region and the localized master distillation region in the interactive distillation region selection module, calculate the intersection-union matrix RIOU_0 between all anchor frames and the rotated truth frame in the l-th layer. l The formula is:

[0027]

[0028] in, B represents the preset anchor frame of the l-th layer. gt RIoU represents the rotation of the truth box, and RIoU represents the degree of overlap between the two boxes.

[0029] RIOU_0 l The lower limit is set to α. pos α pos This indicates that the positive sample threshold set in the basic framework is greater than α. pos The regions are the classification master distillation region and the localization master distillation region that the student network needs to learn. The masks for these two master distillation regions are denoted as follows: and Loss function of the primary distillation region in the l-th layer classification The calculation formula is:

[0030]

[0031] Loss function for locating the main distillation region at layer l The calculation formula is:

[0032]

[0033] in, This represents the classification prediction value of the detection head at layer l of the student network. This represents the classification prediction value of the l-th layer detector head in the pre-trained teacher network. This represents the regression prediction value of the l-th layer detector head in the student network. L represents the regression prediction value of the l-th layer detector head in the pre-trained teacher network. KL This represents the KL divergence loss.

[0034] Calculate the loss function for the classification master distillation region and the loss function for the localization master distillation region for all feature layers, and finally calculate the loss function L for the classification master distillation region. CMD for:

[0035]

[0036] Loss function L for locating the main distillation region LMD for:

[0037]

[0038] Where L represents the total number of layers in the output feature.

[0039] The main classification distillation region only considers the classification region where the positive sample is located, ignoring the region where there may be potential positive samples. Therefore, a localization-assisted classification distillation region is introduced.

[0040] For the selection of the classification distillation region with localization assistance, the IL of the detection head in the l-th layer of the pre-trained teacher network is first calculated. l The intersection-union matrix RIOU_1 of the predicted bounding boxes and the J ground truth boxes l , The calculation formula is:

[0041]

[0042] in, This represents the output prediction box of the detection head at layer l of the pre-trained teacher network.

[0043] For each truth box, remove the I-value of the j-th truth box from the detection head of the l-th layer of the pre-trained teacher network. l Cross-union matrix of prediction boxes The maximum value in Of the remaining values, calculate the first k RIOU_1 values. lThe sum of the values ​​is denoted as iou_sum, and k takes the value min = min(1000, [iou_sum]), where [iou_sum] represents rounding down iou_sum. Let k be the number of samples that need to be distilled for this truth box. The index value corresponding to the required distilled samples is the classification region that the student network should learn. Regions whose index values ​​are the same as the main classification distillation region that the student network should learn are not distilled. The operation of obtaining the index value is defined as get_index(), and then the mask of the classification distillation region assisted by the l-th layer detector is obtained. The calculation formula is:

[0044]

[0045] The loss function corresponding to the classification distillation region with positioning assistance of the l-th layer detection head. The formula is:

[0046]

[0047] Calculate the loss function corresponding to the localization-aided classification distillation region of each detection head, and finally calculate the loss function L corresponding to the localization-aided classification distillation region. LACD The formula is:

[0048] L LACD =M LACD L KL (P S ,P T ,RIoU_1)

[0049] Among them, P S P represents the classification prediction value of all layers of the student network's detector heads. T M represents the classification prediction values ​​of all layers of the pre-trained teacher network's detection heads. LACD This represents the mask for the classification distillation region that assists in the positioning of all layer detector heads.

[0050] The process of selecting the localization distillation region for classification assistance is similar to that for localization assistance. To avoid the student network learning incorrect regions, the classification prediction values ​​P of all layers of the pre-trained teacher network's detector heads are used. T Combined with the intersection-over-union (IoU) matrix RIOU_0 between all anchor boxes and rotated truth boxes across all layers, a classification-weighted IoU matrix RIOU_2 is defined, calculated as follows:

[0051]

[0052] For each ground truth box, remove the classification-weighted intersection-over-union matrix of the l-th layer detector head of the pre-trained teacher network corresponding to the j-th ground truth box. The maximum value in From the remaining values, calculate the weighted intersection-union matrix of the first k' categories. The sum is denoted as iou_sum', k' = min(1000, [iou_sum']), let k' be the number of samples that need to be distilled for the truth box, the index value corresponding to the required distillation samples is the localization region that the student network needs to learn, and the mask of the localization distillation region assisted by the classification of the l-th layer detector head. The calculation formula is:

[0053]

[0054] The loss function corresponding to the classification-aided localization distillation region of the l-th layer detection head. The formula is:

[0055]

[0056] Among them, RIOU_2 l Let represent the classification weighted intersection-union matrix of the detection head in the l-th layer of the pre-trained teacher network.

[0057] Calculate the loss function corresponding to the classification-aided localization distillation region for each layer of the detection head, and finally calculate the loss function L corresponding to the classification-aided localization distillation region. CALD The formula is:

[0058] L CALD =M CALD L KL (B S B T ,RIoU_2)

[0059] Among them, B S B represents the regression prediction value of all layer detector heads in the student network. T M represents the regression predictions of all layers of the pre-trained teacher network's detector heads. CALD This represents the mask value for the positioning distillation region assisted by the classification of all layers of the detector head.

[0060] The interactive distillation region selection module uses the localization branch of the pre-trained teacher network to guide the selection of classification distillation regions in the student network, and conversely, uses the classification branch of the pre-trained teacher network to guide the selection of localization distillation regions in the student network. This dynamically selects the most suitable distillation regions for classification distillation and localization distillation, which significantly improves distillation efficiency and solves the problem of insufficient interaction between classification and regression branches during the distillation process.

[0061] Construct an adaptive distillation generation module:

[0062] The classification prediction value of the l-th layer of the pre-trained teacher network is The classification prediction value of the l-th layer of the student network is Hl W represents the height of the features in the l-th layer. l K represents the width of the feature in the l-th layer, and K represents the number of channels in the feature.

[0063] Design a random mask to randomly mask the predicted values ​​output by the student network classification branch. The calculation formula is as follows:

[0064]

[0065] in, This represents the mask value at position (m, n) on the feature map of layer l. Let represent the feature value at any position (m,n) in the l-th layer feature map, and ε represent the set mask threshold. If the feature value at position (m,n) is... If the value is greater than or equal to the mask threshold ε, then the mask value at that position is set to 1; otherwise, it is set to 0.

[0066] The classification prediction value of the l-th layer of the student network Mask values ​​at each position in the l-th layer The product is fed into a classification restorer, which consists of two 3×3 convolutions and a ReLU function. The formula for calculating the restored classification features is as follows:

[0067]

[0068] in, This represents the classification prediction value of the l-th layer of the student network. The reconstructed classification prediction value after random masking is transformed to obtain:

[0069]

[0070] in, This represents the classification prediction probability recovered after the student network's layer l is randomly masked. ReShape() represents the dimension transformation function.

[0071] Will The loss function for the main classification distillation region, added to both the classification master distillation region and the localization-assisted classification distillation region, is redefined as L'. CMD :

[0072]

[0073] in, This represents the classification prediction probability of the l-th layer of the pre-trained teacher network;

[0074] The loss function for the localization-assisted classification distillation region is redefined as L' LACD :

[0075] L' LACD =M LACD L KL (p' S ,p T ,RIoU_1)

[0076] Among them, p' s p represents the classification prediction probability recovered after random masking of the classification prediction values ​​of all layers of the student network. T This represents the classification prediction probability of all layers in the pre-trained teacher network.

[0077] The key to the adaptive generative distillation module is to allow the pre-trained teacher network to guide the student network to recover its own classification features, enhance the representational ability of classification features, and allow the student network to maintain independent thinking while absorbing teacher knowledge, thus enabling it to have a more comprehensive learning ability.

[0078] Construct the overall loss function for the student network:

[0079] When training the student network, all parameters of the pre-trained teacher network remain unchanged, while the parameters of the student network are updated using a loss function. The overall loss function L of the student network is:

[0080] L = L task +β1L' CMD +β2L LMD +β3L' LACD +β4L ACD

[0081] Among them, L task β1 represents the original detection task loss of the student network, which includes classification loss and regression loss; β2, β3, and β4 represent the coefficients used to balance the different distillation loss terms.

[0082] By combining adaptive generative distillation and interactive learning, the knowledge contained in the detection head was effectively extracted, comprehensively improving the learning ability of student networks.

[0083] Proceed to step S5.

[0084] Step S5: Input the test dataset into the trained student network for detection, merge the detection results of each block image, output the category and location of all targets in the test dataset in the image, and obtain the detection accuracy of the trained student network.

[0085] Example 1

[0086] The aerial image target detection method based on classification-localization bi-branch interactive distillation described in this invention comprises the following steps:

[0087] Step S1: Download the aerial image detection dataset DIOR-R, randomly cut the image into 800×800 blocks, randomly divide the blocks of uniform size into training and test datasets in a 7:3 ratio, perform data augmentation on the training dataset to form the teacher-student learning network training dataset, and proceed to step S2.

[0088] Step S2: Use the teacher-student learning network training dataset to pre-train the teacher network to obtain the pre-trained teacher network, and then proceed to step S3.

[0089] Step S3: Construct a teacher-student learning network using a pre-trained teacher network, student network, adaptive generation distillation module, and interactive distillation region selection module, then proceed to step S4.

[0090] Step S4: Input the teacher-student learning network training dataset into the teacher-student learning network for training, extract the multi-scale feature layers of the pre-trained teacher network and student network and their corresponding output prediction values; construct an interactive distillation region selection module to dynamically select suitable classification distillation regions and localization distillation regions for distillation; construct an adaptive generation distillation module to mask the output feature layer of the student network classification branch, so that it adaptively reconstructs the output feature layer of the pre-trained teacher network, fixes the parameters of the pre-trained teacher network, and updates the parameters of the student network through the overall loss function of the student network, finally obtaining the trained student network, and proceeding to step S5.

[0091] Step S5: Input the test dataset into the trained student network for detection, merge the detection results of each block image, output the category and location of all targets in the test dataset in the image, and obtain the detection accuracy of the trained student network.

[0092] This invention utilizes Python and the PyTorch framework for experiments, with model training and inference performed on an NVIDIA GeForce RTX 3090 graphics card. In the experiments, the model training employed the SGD strategy for optimization, with a momentum decay of 0.9 and a weight decay of 0.0001. The initial learning rate was 0.0025, which was divided by 10 at epochs 24 and 33. The DIOR-R dataset was used for 36 training epochs, with a batch size of 2. For the hyperparameter settings, the balance coefficients for the four distillation regions were set to β1 = β2 = β3 = β4 = 0.1, and the mask threshold ε was set to 0.65.

[0093] To demonstrate the effectiveness of this invention, several popular distillation algorithms from recent years were selected as comparative models. In the experiment, a high-precision remote sensing image target detection algorithm with a ResNet-101 backbone network was used. 2A-Net was used as the teacher network, and MobilenetV2 was used as the backbone extraction network for the student network. All other parameters remained unchanged. The experimental results are shown in Table 1.

[0094] Table 1. Comparison of experimental results with other distillation algorithms on the DIOR-R dataset.

[0095]

[0096] The experimental results in Table 1 demonstrate the practicality and effectiveness of this invention.

Claims

1. A target detection method for aerial images based on classification-localization bi-branch interactive distillation, characterized in that, Includes the following steps: Step S1: Download the aerial image detection dataset DIOR-R, randomly cut the image into 800×800 blocks, randomly divide the blocks of uniform size into training dataset and test dataset in a 7:3 ratio, perform data augmentation on the training dataset to form the teacher-student learning network training dataset, and proceed to step S2. Step S2: Use the teacher-student learning network training dataset to pre-train the teacher network to obtain the pre-trained teacher network, then proceed to step S3; Step S3: Construct a teacher-student learning network using a pre-trained teacher network, student network, adaptive generation distillation module, and interactive distillation region selection module, then proceed to step S4; Step S4: Input the teacher-student learning network training dataset into the teacher-student learning network for training. Extract the multi-scale feature layers and their corresponding output prediction values ​​of the pre-trained teacher network and student network. Construct an interactive distillation region selection module to dynamically select suitable classification distillation regions and localization distillation regions for distillation. Construct an adaptive generation distillation module to mask the output feature layer of the student network's classification branch, enabling it to adaptively reconstruct the output feature layer of the pre-trained teacher network. Fix the parameters of the pre-trained teacher network and update the parameters of the student network using the overall loss function of the student network to finally obtain the trained student network. The interactive distillation region selection module is constructed as follows: The interactive distillation region selection module includes a categorized main distillation region, a positioned main distillation region, a positioned auxiliary categorized distillation region, and a categorized auxiliary positioned distillation region. The primary distillation region for classification and the primary distillation region for localization are directly determined by label assignment, i.e., the location of the positive sample during training; the localization-assisted classification distillation region uses the localization branch of the pre-trained teacher network to guide the selection of the classification distillation region for the student network; the classification-assisted localization distillation region uses the classification branch of the pre-trained teacher network to guide the selection of the localization distillation region for the student network. Proceed to step S5; Step S5: Input the test dataset into the trained student network for detection, merge the detection results of each block image, output the category and location of all targets in the test dataset in the image, and obtain the detection accuracy of the trained student network.

2. The aerial image target detection method based on classification-localization bi-branch interactive distillation according to claim 1, characterized in that, In step S4, the selection of the main distillation region in the interactive distillation region selection module, specifically the classification of the main distillation region and the location of the main distillation region, is as follows: Extract all output features from the detection heads of the pre-trained teacher and student networks, and calculate the intersection-union matrix (CUC) between all anchor boxes and rotated truth boxes in layer l. The formula is: ； in, This indicates the preset anchor frame of the l-th layer. Indicates a rotated truth box. Indicates the degree of overlap between the two frames; Will The lower limit is set to , This indicates that the positive sample threshold set in the basic framework is greater than... The regions are the classification master distillation region and the localization master distillation region that the student network needs to learn. The masks for these two master distillation regions are denoted as follows: and The loss function of the l-th layer classification master distillation region The calculation formula is: ； Loss function for locating the main distillation region at layer l The calculation formula is: ； in, This represents the classification prediction value of the detection head at layer l of the student network. This represents the classification prediction value of the l-th layer detector head in the pre-trained teacher network. This represents the regression prediction value of the l-th layer detector head in the student network. This represents the regression prediction value of the l-th layer detector head in the pre-trained teacher network. Indicates the KL divergence loss; Calculate the loss function for the classification master distillation region and the loss function for the localization master distillation region for all feature layers, and finally calculate the loss function for the classification master distillation region. for: ； Loss function for locating the main distillation region for: ； in, This indicates the total number of layers in the output feature.

3. The aerial image target detection method based on classification-localization bi-branch interactive distillation according to claim 2, characterized in that, In step S4, the selection of the positioning-assisted classification distillation region and the classification-assisted positioning distillation region in the interactive distillation region selection module is as follows: For the selection of the classification distillation region with localization assistance, the detection head of the l-th layer of the pre-trained teacher network is first calculated. The intersection-union matrix of 1 predicted bounding box and 1 ground truth bounding box. The calculation formula is: ； in, This represents the output prediction box of the detection head at layer l of the pre-trained teacher network; For each truth box, remove the first... The truth boxes and the detection head of the l-th layer of the pre-trained teacher network Cross-union matrix of prediction boxes The maximum value in Of the remaining values, calculate the first n. The sum of values ​​is denoted as 𝑖𝑜𝑢_𝑠𝑢𝑚, and the value of 𝑘 is 𝑘=𝑚𝑖𝑛(1000, [𝑖𝑜𝑢_𝑠𝑢𝑚]), [𝑖𝑜𝑢_𝑠𝑢𝑚] represents rounding down 𝑖𝑜𝑢_𝑠𝑢𝑚. Let 𝑘 be the number of samples that need to be distilled for this truth box. The index value corresponding to the required distilled samples is the classification region that the student network should learn. Regions whose index values ​​are the same as the main classification distillation region that the student network should learn are not distilled. The operation of obtaining the index value is defined as 𝑔𝑒𝑡_𝑖𝑛d𝑒𝑥(), then the mask of the classification distillation region for the localization assistance of the l-th layer detector head. The calculation formula is: ； The loss function corresponding to the classification distillation region with positioning assistance of the l-th layer detection head. The formula is: ； Calculate the loss function corresponding to the localization-aided classification distillation region for each layer of the detection head, and finally calculate the loss function corresponding to the localization-aided classification distillation region. The formula is: ； in, This represents the classification prediction values ​​of all layers of the student network's detector heads. This represents the classification prediction values ​​of all layers of the pre-trained teacher network's detection heads. A mask representing the classification distillation region for all layer detector head positioning assistance; The process of selecting the localization distillation region for classification assistance is similar to that of localization assistance for classification distillation, using the classification prediction values ​​of all layers of the pre-trained teacher network's detector heads. The cross-union matrix between all anchor boxes and rotated truth boxes in all layers By combining these factors, a classification-weighted intersection-union matrix is ​​defined. The calculation formula is: ； For each truth box, remove the first... The classification weighted intersection-over-union matrix of the l-th layer detector head of the pre-trained teacher network corresponding to each ground truth box. The maximum value in From the remaining values, calculate the weighted intersection-union matrix of the first n' categories. The sum is denoted as 𝑖𝑜𝑢_𝑠𝑢𝑚', 𝑘'=𝑚𝑖𝑛(1000, [𝑖𝑜𝑢_𝑠𝑢𝑚']), let 𝑘' be the number of samples required for distillation of the truth box, the index value corresponding to the required distillation samples is the localization region that the student network needs to learn, and the mask of the localization distillation region assisted by the classification of the l-th layer detector head. for: ； The loss function corresponding to the classification-aided localization distillation region of the l-th layer detection head. The formula is: ； in, This represents the classification weighted intersection-union matrix of the detection heads in the l-th layer of the pre-trained teacher network; Calculate the loss function corresponding to the classification-aided localization distillation region for each layer of the detection head, and finally calculate the loss function corresponding to the classification-aided localization distillation region. The formula is: ； in, This represents the regression prediction values ​​of all layer detector heads in the student network. This represents the regression prediction values ​​of all layers of the pre-trained teacher network's detector heads. This represents the mask value for the positioning distillation region assisted by the classification of all layers of the detector head.

4. The aerial image target detection method based on classification-localization bi-branch interactive distillation according to claim 3, characterized in that, In step S4, an adaptive generation distillation module is constructed, as follows: The classification prediction value of the l-th layer of the pre-trained teacher network is The classification prediction value of the l-th layer of the student network is , Indicates the height of the features in the l-th layer. K represents the width of the feature in the l-th layer, and K represents the number of channels in the feature. Design a random mask to randomly mask the predicted values ​​output by the student network classification branch. The calculation formula is as follows: ； in, Indicates the position on the feature map of layer l. The mask value at that location, Represents any position in the feature map of layer l. eigenvalues ​​at that location This indicates the set mask threshold; if the position... eigenvalues ​​at Greater than or equal to the mask threshold If the value is 1, then the mask value at that position is set to 1; otherwise, it is set to 0. The classification prediction value of the l-th layer of the student network Mask values ​​at each position in the l-th layer The product is fed into a classification restorer, which consists of two 3×3 convolutions and a ReLU function. The formula for calculating the restored classification features is as follows: ； in, This represents the classification prediction value of the l-th layer of the student network. The reconstructed classification prediction value after random masking is transformed to obtain: ； in, This represents the classification prediction probability recovered after the student network's layer l is randomly masked. , Indicates a dimension transformation function; Will The loss function for the main classification distillation region, added to both the classification master distillation region and the localization-assisted classification distillation region, is redefined as follows: : ； in, This represents the classification prediction probability of the l-th layer of the pre-trained teacher network; The loss function for localization-assisted classification distillation regions is redefined as : ； in, This represents the classification prediction probability recovered after random masking of the classification prediction values ​​of all layers of the student network. This represents the classification prediction probability of all layers in the pre-trained teacher network.

5. The aerial image target detection method based on classification-localization bi-branch interactive distillation according to claim 4, characterized in that, In step S4, the construction of the overall loss function of the student network is as follows: When training the student network, all parameters of the pre-trained teacher network remain unchanged, while the parameters of the student network are updated using a loss function. The overall loss function L of the student network is: ； in, The original detection task loss of the student network is represented by two parts: classification loss and regression loss. , , , This represents the coefficient used to balance different distillation loss terms.

Images