Characteristic factorization remote sensing target counting model establishment method and device
By using the feature factorization method, combined with the backbone network, multi-order collaborative modules, and dynamic input sensing distillation technology, a remote sensing target counting model was constructed, which solved the problems of robustness and high accuracy in remote sensing target counting and achieved better counting results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NANJING UNIV OF INFORMATION SCI & TECH
- Filing Date
- 2025-07-29
- Publication Date
- 2026-07-07
AI Technical Summary
Existing remote sensing target counting models struggle to achieve robust and high-precision target counting when faced with factors such as small object size, large scale variations, complex background interference, and occlusion.
A feature factorization method is adopted, and a remote sensing target counting model is constructed by using a backbone network, multi-order collaborative modules and dynamic input sensing distillation technology. The multi-order collaborative modules are used to enhance the features of salient areas and the optimal density map is generated by dynamic input sensing distillation to guide the network output.
It improves the accuracy and stability of remote sensing target counting, better captures the intrinsic correlation between features, and enhances the robustness and applicability of the model.
Smart Images
Figure CN120912555B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer vision technology, specifically to a method and apparatus for establishing a remote sensing target counting model based on feature factorization. Background Technology
[0002] The purpose of remote sensing counting is to accurately map pixel-level features in remote sensing images to the number of specific objects present in a scene. This technology automatically and efficiently extracts object counts from large amounts of remote sensing images, providing crucial data support for decision-making in various fields such as urban planning, public safety, ecological monitoring, and agricultural yield estimation, demonstrating significant practical value and broad application prospects. However, achieving robust and high-accuracy counting remains challenging due to factors such as small object size, large scale variations, complex background interference, and frequent object occlusion and congestion. To overcome these obstacles, extensive research has been conducted to improve the robustness, accuracy, and real-world applicability of remote sensing counting models.
[0003] Among these methods, hierarchical feature fusion is a common technique for improving the model's adaptability to scale variations in remotely sensed images. This approach leverages the complementary advantages of features at different depths: shallow features capture fine-grained details such as edges and textures, which are crucial for detecting small targets; while deeper features extract high-level semantic information, such as object structure and contextual cues, making them more effective at identifying large or complex targets. Therefore, effectively integrating multi-level features is essential for building stable counting models. These fusion strategies typically include bidirectional architectures combining bottom-up and top-down paths, as well as parallel multi-branch designs, generally relying on simple linear operations for feature fusion, such as element-wise addition or channel concatenation. However, such operations often fail to capture the intrinsic correlations between features, thus limiting the model's ability to fully utilize feature relationships and improve performance. Summary of the Invention
[0004] The purpose of this invention is to propose a method and apparatus for establishing a remote sensing target counting model based on feature factorization, which can efficiently achieve the target counting task and produce counting results that are superior to those of existing technologies.
[0005] To achieve the above-mentioned technical objectives, the technical solution adopted by the present invention is as follows:
[0006] In a first aspect, the present invention discloses a method for establishing a remote sensing target counting model based on feature factorization, the method comprising the following steps:
[0007] Step S1: Divide the dataset containing several images to be counted into a training dataset, a validation dataset, and a test dataset;
[0008] Step S2: Preprocess the images in the training dataset;
[0009] Step S3: Construct a feature factorization remote sensing target counting model, including a backbone network, a multi-level collaborative module, a dynamic input-aware distillation, and a regression head; the backbone network is used for feature extraction from the input image; the multi-level collaborative module is used to enhance salient region features while suppressing irrelevant representations to improve the model's discriminative ability; the dynamic input-aware distillation is used to generate diverse input variants, dynamically select the optimal density map, and guide the output in the network based on a single sample input; the regression head is used to regress the density map from the output of the multi-level collaborative module through upsampling and convolution operations.
[0010] Step S4: Obtain the preprocessed training data in sequence and train the feature factorization remote sensing target counting model.
[0011] Step S5: Select and save the feature factorization remote sensing target counting model with the highest prediction accuracy from the validation dataset;
[0012] Step S6: Test the saved model using the test dataset to obtain the target counting accuracy of the model;
[0013] Step S7: Complete the training and obtain the trained feature factorization remote sensing target counting model.
[0014] Furthermore, step S2 further includes:
[0015] Step S2-1: Randomly scale the image within a preset ratio range and crop it to a fixed resolution;
[0016] Step S2-2: Randomly flip or randomly grayscale the cropped image with a preset probability to achieve data augmentation.
[0017] Furthermore, the multi-level collaborative module includes a first-order interaction module and a second-order interaction module;
[0018] The first-order interaction module uses a low-rank matrix to replace the traditional full-parameter transformation matrix for feature transformation in the multi-feature fusion process; specifically, the transformation matrix W is initialized for multiple features obtained from the backbone network. i =A i B i ,in and r< <min(m,n);
[0019] The first-order interaction module is implemented using the following formula:
[0020]
[0021] In the formula, l represents the number of interactive features, Xi This represents the i-th same-scale feature obtained from different depths of the backbone network, where m and n represent the matrix A, respectively. i Input dimensions and matrix B i The output dimension is r, where r represents the rank of the matrix;
[0022] The second-order interaction module explicitly models the interactions between pairs of features, enabling the model to identify key feature combinations and enhancing the model's ability to represent complex dependencies.
[0023] The second-order interaction module is implemented using the following formula:
[0024]
[0025] In the formula, α represents the weighting factor. Let represent the i-th learnable matrix.
[0026] Furthermore, the dynamic input-aware distillation adds multiple initialized learnable cues and random noise to the original input image to generate cue-guided input and uncertainty-guided input, respectively, for additional supplementation of the network input. The data streams for cue-guided and uncertainty-guided inputs are defined as input guidance paths, and the data streams directly input to the original image are defined as base paths. For the input guidance path, the optimal density map is selected from the density maps generated based on the input-aware guidance pool by measuring the counting accuracy (MAE), MSE, and structural similarity (SSIM) between the density map and the ground truth density map, guiding the network output of the base path. For the base path, the counting loss, optimal transmission loss, and total change loss are calculated based on the network output of the original input image and the ground truth density map. The spatial distillation loss is calculated based on the network output of the input guidance path and the network output of the original input image. Finally, the Adam optimizer is used to optimize the network parameters.
[0027] Furthermore, the input-aware guidance pool is implemented using the following formula:
[0028] X = {X θ ,X μ}
[0029] X θ =I+θ,X μ =I+μ
[0030] In the formula, θ represents the learnable cue, μ represents the frozen random noise, and X θ Indicates a prompt or guided input, X μ represents the input guided by uncertainty, and I represents the original input image of the network.
[0031] Furthermore, the density map generated based on the input-aware guidance pool is calculated according to the following formula:
[0032]
[0033] In the formula, E(·) represents the backbone network, and R(·) represents the regression head. This represents the density map generated based on the input-aware guided pool for input X. This indicates the density map generated under the guidance of prompts. F1(·) and F2(·) represent the density map generated by uncertainty guidance, respectively. F1(·) and F2(·) represent the function functions of the first-order and second-order interaction modules contained in the multi-order collaborative module.
[0034] Furthermore, the process of selecting the optimal density map from the density maps generated based on the input-aware guided pool includes:
[0035] Calculate MAE using the following formula:
[0036]
[0037] In the formula, y (h,w) It is a true density map. This represents the density map generated based on the input-aware guided pool, where H and W are the height and width of the density map;
[0038] Calculate MSE using the following formula:
[0039]
[0040] SSIM is calculated using the following formula:
[0041]
[0042] In the formula, μ x and μ y σ represents the mean of the real density map and the density map generated based on the input-aware guided pool. x 2 and σ y 2 σ represents the variance between the true density map and the density map generated based on the input-aware guided pool. xy c1 and c2 represent the covariance between the true density map and the density map generated based on the input-aware guided pool, and represent constants.
[0043] The multi-index fusion evaluation function is calculated using the following formula:
[0044]
[0045] In the formula, ω, β, and γ represent hyperparameters; ε represents a constant used to prevent the denominator from being 0.
[0046] The optimal density map is selected from the density maps generated based on the input-aware guidance pool using a multi-index fusion evaluation function.
[0047] Furthermore, step S4 specifically includes:
[0048] Step S4-1: Divide the preprocessed population into a training set and a validation set;
[0049] Step S4-2: Initialize the learning rate;
[0050] Step S4-3: For the input guidance path, a multi-index optimization screening strategy is used to select the density map with the most accurate prediction and the most similar structure to guide the network output of the base path; for the base path, the network output is based on the original input image. Calculate the loss using the true density map y Network output based on input guidance path Network output with the original input image Calculate space distillation loss Finally, based on the total loss of the model... Use the Adam optimizer to optimize network parameters;
[0051] Step S4-4: Use the validation set to select the optimal network parameters;
[0052] Step S4-5: Initialize the feature factorization remote sensing target counting model using the optimal model parameters to complete model training.
[0053] Furthermore, in step S4-3, the process of calculating the total loss of the model includes the following steps:
[0054] S4-3-1: Calculate the counting loss according to the following formula.
[0055]
[0056] In the formula, ||·||1 represents L1 normal form;
[0057] S4-3-2: Calculate the optimal transmission loss according to the following formula.
[0058]
[0059] In the formula, α * and β * The solution represents the optimal transmission cost, and <·, ·> represent the dot product operation;
[0060] S4-3-3: Calculate the total change loss according to the following formula.
[0061]
[0062] S4-3-4: Calculate the space distillation loss according to the following formula.
[0063]
[0064] In the formula, C represents the number of channels in the density map. and These represent spatial distillation losses in the horizontal and vertical directions, respectively.
[0065] S4-3-5: Calculate the total loss of the model according to the following formula.
[0066]
[0067] In the formula, λ1, λ2 and λ3 represent hyperparameters.
[0068] Secondly, the present invention discloses a feature factorization remote sensing target counting model establishment device, including a processor and a computer program stored in a memory and capable of running on the processor, wherein the processor executes the program to implement the method described above.
[0069] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0070] First, the feature factorization remote sensing target counting model establishment method and apparatus of the present invention proposes a multi-level collaborative module, which can enable the network to focus on the importance of individual features and model complex relationships between features.
[0071] Second, the feature factorization remote sensing target counting model establishment method and apparatus of the present invention proposes dynamic input sensing distillation, which can generate diverse input variants, dynamically select the most accurate and structurally similar density map, and guide the network output of the base path. Attached Figure Description
[0072] Figure 1 This is a flowchart illustrating the method for establishing a remote sensing target counting model based on feature factorization provided by the present invention.
[0073] Figure 2 This is a schematic diagram of the structure of the remote sensing target counting model based on the characteristic factorization of the present invention. Detailed Implementation
[0074] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.
[0075] This embodiment provides a method for establishing a remote sensing target counting model based on feature factorization, such as... Figure 1 and Figure 2As shown, the specific steps include the following:
[0076] Step S1: Divide the dataset containing several images to be counted into a training dataset, a validation dataset, and a test dataset;
[0077] Step S2: Preprocess the images in the training dataset;
[0078] The preprocessing methods are as follows:
[0079] Step S2-1: Randomly scale the image within a preset ratio range and crop it to a fixed resolution; in this embodiment, the preset ratio range for random scaling is 0.8 to 1.2, and the fixed resolution is 256*256.
[0080] Step S2-2: Perform data augmentation on the cropped image. The data augmentation method is as follows: S2-2-1: Randomly flip the image horizontally with a preset probability; S2-2-2: Randomly process the image to grayscale with a preset probability. In this embodiment, the preset probabilities for random horizontal flipping and random grayscale processing are 0.5 and 0.1, respectively.
[0081] Step S3: Construct a feature factorization remote sensing target counting model, including a backbone network, a multi-level collaborative module, a dynamic input perceptual distillation, and a regression head; the backbone network is used for feature extraction from the input image; the multi-level collaborative module is used to enhance salient region features while suppressing irrelevant representations, thereby improving the model's discriminative ability; the dynamic input perceptual distillation is used to generate diverse input variants, dynamically select the optimal density map, and guide the output in the network based on a single sample input; the regression head is used to regress the density map from the output of the multi-level collaborative module through upsampling and convolution operations.
[0082] Specifically, the backbone network employs a pre-trained VMamba-base model to capture detailed feature information from different network depths.
[0083] Specifically, the multi-level collaborative module includes:
[0084] First-order interaction module: Addressing the issue of increased model optimization complexity caused by the introduction of numerous additional parameters using full-rank matrices in common feature alignment methods, this module proposes replacing traditional full-parameter transformation matrices with low-rank matrices for feature adaptation during multi-feature fusion. Specifically, a weight matrix W = AB is initialized for each of the multiple features obtained from the backbone network, where... and r << min(m,n).
[0085] The first-order interaction module is implemented using the following formula:
[0086]
[0087] In the formula, l represents the number of interactive features, X i This represents the i-th same-scale feature obtained from different depths of the backbone network, where m and n represent the matrix A, respectively. i Input dimensions and matrix B i The output dimension is denoted by r, which represents the rank of the matrix. The main advantage of this design is its high parameter efficiency. By approximating the transformation matrix, the number of parameters is reduced from mn to (m+n)r, significantly reducing model complexity and thus mitigating the risk of overfitting during fine-tuning, especially on small datasets. Furthermore, it encourages the model to capture only the most relevant information between features, effectively filtering out redundant features and noise, thereby improving generalization performance.
[0088] Second-order interaction modules: First-order interactions independently evaluate the importance of each feature through linear combinations. While they capture the individual contribution of features, they fail to model more complex relationships between features. This limitation becomes particularly pronounced in scenarios requiring collaborative decision-making based on multiple features. To address this issue, second-order interaction modules are introduced to explicitly model interactions between pairs of features. This approach enables the model to identify key feature combinations, enhancing its ability to represent complex dependencies.
[0089] The second-order interaction module is implemented using the following formula:
[0090]
[0091] In the formula, α represents the weighting factor. Let represent the i-th learnable matrix.
[0092] The dynamic input-aware distillation process involves adding multiple initialized learnable cues and random noise to the original input image to generate cue-guided input and uncertainty-guided input, respectively, to supplement the network input. The data streams for cue-guided and uncertainty-guided inputs are defined as input guidance paths, while the data stream directly input to the original image is defined as base paths. For the input guidance path, the optimal density map is selected from the density maps generated based on the input-aware guidance pool by measuring the counting accuracy (MAE), MSE, and structural similarity (SSIM) between the density map and the ground truth density map, guiding the network output of the base path. For the base path, the counting loss, optimal transmission loss, and total change loss are calculated based on the network output of the original input image and the ground truth density map. The spatial distillation loss is calculated based on the network output of the input guidance path and the network output of the original input image. Finally, the Adam optimizer is used to optimize the network parameters.
[0093] The input-aware guidance pool is implemented using the following formula:
[0094] X∈{X θ ,X μ}
[0095] X θ =I+θ,X μ =I+μ
[0096] In the formula, θ represents the learnable cue, μ represents the frozen random noise, and X θ Indicates a prompt or guided input, X μ represents the input guided by uncertainty, and I represents the original input image of the network.
[0097] The density map generated based on the input-aware guide pool is calculated using the following formula:
[0098]
[0099] In the formula, E(·) represents the backbone network, and R(·) represents the regression head. This represents the density map generated based on the input-aware guided pool for input X. This indicates the density map generated under the guidance of prompts. F1(·) and F2(·) represent the density map generated by uncertainty guidance, respectively. F1(·) and F2(·) represent the function functions of the first-order and second-order interaction modules contained in the multi-order collaborative module.
[0100] Specifically, the regression head consists of three 3*3 convolutions and one 1*1 convolution, ultimately generating a density map with 1 channel.
[0101] The process of selecting the optimal density map from the density maps generated based on the input-aware guided pool includes:
[0102] Calculate MAE using the following formula:
[0103]
[0104] In the formula, y (h,w) Represents the true density map, This represents the density map generated based on the input-aware guided pool, where H and W represent the height and width of the density map, respectively.
[0105] Calculate MSE using the following formula:
[0106]
[0107] SSIM is calculated using the following formula:
[0108]
[0109] In the formula, μ xand μ y σ represents the mean of the real density map and the density map generated based on the input-aware guided pool. x 2 and σ y 2 σ represents the variance between the true density map and the density map generated based on the input-aware guided pool. xy c1 and c2 represent the covariance between the true density map and the density map generated based on the input-aware guided pool, and represent constants.
[0110] The multi-index fusion evaluation function is calculated using the following formula:
[0111]
[0112] In the formula, ω, β, and γ represent hyperparameters; ε represents a constant used to prevent the denominator from being 0.
[0113] Among them, the optimal density map is selected from the density map generated based on the input-aware guidance pool according to the multi-index fusion evaluation function.
[0114] Step S4: Obtain the preprocessed training data sequentially and train the feature factorization remote sensing target counting model. The specific training methods include:
[0115] Step S4-1: Divide the preprocessed population into a training set and a validation set;
[0116] Step S4-2: Initialize the learning rate;
[0117] Step S4-3: For the input guidance path, a multi-index optimization screening strategy is used to select the density map with the most accurate prediction and the most similar structure to guide the network output of the base path; for the base path, the network output is based on the original input image. Calculate the loss using the true density map y Network output based on input guidance path Network output with the original input image Calculate space distillation loss Finally, based on the total loss of the model... Use the Adam optimizer to optimize network parameters;
[0118] In step S4-3, the process of calculating the total loss of the model includes the following steps:
[0119] S4-3-1: Calculate the counting loss according to the following formula.
[0120]
[0121] In the formula, ||·||1 represents L1 normal form;
[0122] S4-3-2: Calculate the optimal transmission loss according to the following formula.
[0123]
[0124] In the formula, α * and β * The solution represents the optimal transmission cost, and <·, ·> represent the dot product operation;
[0125] S4-3-3: Calculate the total change loss according to the following formula.
[0126]
[0127] S4-3-4: Calculate the space distillation loss according to the following formula.
[0128]
[0129] In the formula, C represents the number of channels in the density map. and These represent spatial distillation losses in the horizontal and vertical directions, respectively.
[0130] S4-3-5: Calculate the total loss of the model according to the following formula.
[0131]
[0132] In the formula, λ1, λ2 and λ3 represent hyperparameters.
[0133] Step S4-4: Use the validation set to select the optimal network parameters;
[0134] Step S4-5: Initialize the feature factorization remote sensing target counting model using the optimal model parameters to complete model training.
[0135] Step S5: Select and save the feature factorization remote sensing target counting model with the highest prediction accuracy from the validation dataset;
[0136] Step S6: Test the saved model using the test dataset to obtain the target counting accuracy of the model;
[0137] Step S7: Complete the training and obtain the trained feature factorization remote sensing target counting model.
[0138] This embodiment also provides a feature factorization remote sensing target counting model establishment device, including a processor and a computer program stored in a memory and capable of running on the processor. When the processor executes the program, it implements the above-described method.
[0139] The quantitative results of this embodiment on the remote sensing counting dataset RSOC are compared with those of other methods, and the results are shown in Table 1. This embodiment selected two indicators to evaluate the network's counting performance: MAE (Mean Absolute Error) and MSE (Mean Squared Error). A lower MAE indicates more accurate prediction, while a lower MSE indicates more stable prediction results.
[0140] Table 1. Quantitative comparison results of this embodiment with other methods on remote sensing counting datasets.
[0141]
[0142]
[0143] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.
[0144] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.
Claims
1. A method for establishing a remote sensing target counting model based on feature factorization, characterized in that: The method includes the following steps: Step S1: Divide the dataset containing several images to be counted into a training dataset, a validation dataset, and a test dataset; Step S2: Preprocess the images in the training dataset; Step S3: Construct a feature factorization remote sensing target counting model, including a backbone network, a multi-level collaborative module, a dynamic input-aware distillation, and a regression head; the backbone network is used for feature extraction from the input image; the multi-level collaborative module is used to enhance salient region features while suppressing irrelevant representations to improve the model's discriminative ability; the dynamic input-aware distillation is used to generate diverse input variants, dynamically select the optimal density map, and guide the output in the network based on a single sample input; the regression head is used to regress the density map from the output of the multi-level collaborative module through upsampling and convolution operations. Step S4: Obtain the preprocessed training data in sequence and train the feature factorization remote sensing target counting model. Step S5: Select and save the feature factorization remote sensing target counting model with the highest prediction accuracy from the validation dataset; Step S6: Test the saved model using the test dataset to obtain the target counting accuracy of the model; Step S7: Complete the training and obtain the trained feature factorization remote sensing target counting model; The dynamic input-aware distillation process adds multiple initialized learnable cues and random noise to the original input image to generate cue-guided input and uncertainty-guided input, respectively, to supplement the network input. The data streams for cue-guided and uncertainty-guided inputs are defined as input guidance paths, while the data streams directly fed into the original image are defined as base paths. For the input guidance path, the counting accuracy with the true density map is measured. , and structural similarity The optimal density map is selected from the density map generated by the input-aware guided pool to guide the network output of the base path. For the base path, the counting loss, optimal transmission loss and total change loss are calculated based on the network output of the original input image and the real density map. The spatial distillation loss is calculated based on the network output of the input guided path and the network output of the original input image. Finally, the Adam optimizer is used to optimize the network parameters. The input-aware guidance pool is implemented using the following formula: In the formula, Hints indicating that something is worth learning. Indicates frozen random noise, This indicates a prompt or guided input. This indicates input guided by uncertainty. This represents the original input image to the network; The density map generated based on the input-aware guide pool is calculated using the following formula: In the formula, Indicates the backbone network. Indicates the return to the head, This indicates that for the input Density map generated based on input-aware guided pooling. This indicates the density map generated under the guidance of prompts. This represents a density map generated under uncertainty guidance; and These represent the function names of the first-order and second-order interaction modules contained in the multi-order collaborative module, respectively.
2. The method for establishing a remote sensing target counting model based on feature factorization according to claim 1, characterized in that: Step S2 further includes: Step S2-1: Randomly scale the image within a preset ratio range and crop it to a fixed resolution; Step S2-2: Randomly flip or randomly grayscale the cropped image with a preset probability to achieve data augmentation.
3. The method for establishing a remote sensing target counting model based on feature factorization according to claim 1, characterized in that: The multi-level collaborative module includes a first-level interaction module and a second-level interaction module; The first-order interaction module uses a low-rank matrix to replace the traditional full-parameter transformation matrix for feature transformation in the multi-feature fusion process; specifically, the transformation matrix is initialized for multiple features obtained from the backbone network. ,in and , ; The first-order interaction module is implemented using the following formula: In the formula, It is the number of interactive features. This indicates the number of data points obtained from different depths of the backbone network. Features of the same scale and Represent matrices respectively Input dimensions and matrices The output dimension, Describes the rank of a matrix; The second-order interaction module explicitly models the interactions between pairs of features, enabling the model to identify key feature combinations and enhancing the model's ability to represent complex dependencies. The second-order interaction module is implemented using the following formula: In the formula, Indicates the weighting factor. Indicates the first A learnable matrix.
4. The method for establishing a remote sensing target counting model based on feature factorization according to claim 1, characterized in that: The process of selecting the optimal density map from the density maps generated by the input-aware guided pool includes: Calculate according to the following formula : In the formula, Represents the true density map, This represents the density map generated based on the input-aware guided pool. and These represent the height and width of the density map, respectively. Calculate according to the following formula : ； Calculate according to the following formula : In the formula, and This represents the mean of the true density map and the density map generated based on the input-aware guided pool. and This represents the variance between the true density map and the density map generated based on the input-aware guided pool. This represents the covariance between the true density map and the density map generated based on the input-aware guided pool. and Represents a constant; The multi-index fusion evaluation function is calculated using the following formula: In the formula, , and It's a hyperparameter; It is a constant used to prevent the denominator from being zero; The optimal density map is selected from the density maps generated based on the input-aware guidance pool using a multi-index fusion evaluation function.
5. The method for establishing a remote sensing target counting model based on feature factorization according to claim 1, characterized in that: Step S4 specifically includes: Step S4-1: Divide the preprocessed population into a training set and a validation set; Step S4-2: Initialize the learning rate; Step S4-3: For the input guidance path, a multi-index optimization screening strategy is used to select the density map with the most accurate prediction and the most similar structure to guide the network output of the base path; for the base path, the network output is based on the original input image. and true density map Calculate loss Network output based on input guidance path Network output with the original input image Calculate space distillation loss Finally, based on the total loss of the model Use the Adam optimizer to optimize network parameters; Step S4-4: Use the validation set to select the optimal network parameters; Step S4-5: Initialize the feature factorization remote sensing target counting model using the optimal network parameters to complete model training.
6. The method for establishing a remote sensing target counting model based on feature factorization according to claim 1, characterized in that: Step S4-3, the process of calculating the total loss of the model includes the following steps: S4-3-1: Calculate the counting loss according to the following formula. : In the formula, express Paradigm; S4-3-2: Calculate the optimal transmission loss according to the following formula. : In the formula, and This represents the solution with the optimal transmission cost. This represents the dot product operation; S4-3-3: Calculate the total change loss according to the following formula. : ； S4-3-4: Calculate the space distillation loss according to the following formula. : In the formula, This represents the number of channels in the density map. and These represent spatial distillation losses in the horizontal and vertical directions, respectively. S4-3-5: Calculate the total loss of the model according to the following formula. : In the formula, , and This represents hyperparameters.
7. A device for establishing a remote sensing target counting model based on feature factorization, comprising a processor and a computer program stored in a memory and executable on the processor, characterized in that: When the processor executes the program, it implements the method described in any one of claims 1-6.