Conditional frequency domain interactive target counting method and device based on step prior learning

Abstract

The application discloses a kind of condition frequency domain interactive target counting method and device based on step prior learning, first acquisition target image and label establishment training dataset, verification dataset;Condition frequency domain interactive target counting model is designed, and the representation ability of model is enhanced, the condition frequency domain interactive target counting model includes feature extractor, condition frequency domain interactive module and regressor;Training dataset is used to train the condition frequency domain interactive target counting model, in the training process, using step prior learning strategy guides the learning of condition frequency domain interactive target counting model;The best frequency domain self-adapting target counting model is selected and saved using verification dataset screening evaluation index performance;Obtain final frequency domain self-adapting target counting model.The application strengthens the extraction of target counting model feature, and target counting effect is more excellent compared with existing counting model.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a conditional frequency domain interactive target counting method and apparatus based on step prior learning. Background Technology

[0002] Target clustering is prevalent in various scenarios, including business analytics, public safety management, traffic management, plant phenotyping, and pest control. For example, during holidays or promotional periods, the number of customers in shopping malls may increase dramatically, leading to high-density clustering. This can degrade the shopping experience, increase safety risks (such as stampedes), and put pressure on businesses' operations. Every year during migratory bird season, certain locations become habitats for large numbers of birds. Monitoring these birds is beneficial for assessing species diversity and ecological balance. In the face of the aforementioned target clustering phenomena, accurate dense target counting is crucial for risk assessment and the development of effective management measures.

[0003] The dense target counting task aims to automatically identify and count the number of densely distributed targets in images or videos using computer vision techniques. In recent years, a large amount of research has been conducted on this task, yielding fruitful results. A current research focus is on regression using Bayesian loss based on point annotations, achieving excellent performance. However, this loss generates a density map by using points as the regression object, but without explicit density map targets, it leads to distribution drift during the learning process, thus failing to effectively handle irregular targets and occluded targets. Summary of the Invention

[0004] Purpose of the invention: In order to overcome the shortcomings of the prior art, the present invention provides a conditional frequency domain interactive target counting method and apparatus based on step prior learning. By utilizing the step prior learning strategy to construct a robust target counting model, the robustness and adaptability of the model in multiple fields are improved.

[0005] Technical solution: To achieve the above objectives, the technical solution adopted by this invention is as follows:

[0006] A conditional frequency domain interactive target counting method based on step prior learning includes the following steps:

[0007] Step S1: Collect and label target images to establish training and validation datasets.

[0008] Step S2: Design a conditional frequency domain interactive target counting model to enhance the model's representation ability. The conditional frequency domain interactive target counting model includes a feature extractor, a conditional frequency domain interaction module, and a regressor. The feature extractor is used to extract high-level semantic features of the input samples, the conditional frequency domain interaction module is used to enhance global and local features, and the regressor is used to regress the final density map.

[0009] Step S3: Train the conditional frequency domain interactive target counting model described in step S2 using the training dataset. During the training process, the learning of the conditional frequency domain interactive target counting model is guided by a step-prior learning strategy.

[0010] Step S4: Use the validation dataset to select and save the frequency domain adaptive target counting model that performs best on the evaluation metrics.

[0011] Step S5: Return to step S3 until the model training is complete and the final frequency domain adaptive target counting model is obtained.

[0012] Step S6: Acquire the image of the target to be counted, and input the image of the target to be counted into the final frequency domain adaptive target counting model to obtain the target count.

[0013] Preferred method: The specific construction method of the conditional frequency domain interactive target counting model in step 2 is as follows:

[0014] Step S2-1: Extract the features y(x) of the output sample x using the pre-trained network model, denoted as y(x) = φ(x, θ). y ), where φ(·,·) is the feature extractor, and θ y These are network parameters.

[0015] Step S2-2: Transform the features obtained in step S2-1 using a series of convolutional layers to obtain the transformed features. in For a series of convolution operations, θ z These are the parameters of the convolutional layer.

[0016] Step S2-3: Transform the transformed feature z(x) in step S2-2 to the frequency domain using the Fourier transform function, and obtain the real part feature R(x) and the imaginary part feature I(x) corresponding to the transformed feature z(x).

[0017] Step S2-4: Obtain the conditional mapping by passing the output z(x) from step S2-2 through a 1×1 convolutional layer and an adaptive average pooling layer. Further multiplied by a learnable convolution kernel Determine the conditional convolution kernel

[0018] Step S2-5: Obtain the conditional mapping by passing the output z(x) from step S2-2 through a 1×1 convolutional layer and an adaptive average pooling layer. Further multiplied by a learnable convolution kernel Determine the conditional convolution kernel

[0019] Step S2-6: Convolve the real part features R(x) from step S2-3 using the conditional convolution kernel K0 from step S2-4, and then obtain the conditional real part features after passing through a series of convolutional layers.

[0020] Step S2-7: Convolve the real part feature I(x) from step S2-3 using the conditional convolution kernel K1 from step S2-5, and then obtain the conditional real part feature after passing through a series of convolutional layers.

[0021] Step S2-8: The conditional real part features obtained in step S2-6 are... and the conditional real part features obtained in steps S2-7 Perform an inverse Fourier transform, and then add the transformed feature z(x) from step S2-2 to obtain the feature.

[0022] Step S2-9: Output the features from step S2-8 After passing through the regressor, the density prediction map m(x) is obtained.

[0023] Preferred method: The method for training the step-conditional frequency domain interactive target counting model using the training dataset in step S3:

[0024] Step S3-1: Set the parameters required for network training and initialize the prior labels of the samples.

[0025] Step S3-2: Feed the samples in the training dataset into the conditional frequency domain interactive target counting model in batches, and calculate the total loss. Then, the Adam optimizer is used to optimize the model.

[0026]

[0027] Where N represents the number of samples in the current batch, This indicates whether the j-th annotation point in the actual annotation is reliable. This represents the bias generated by the predicted density map at the j-th labeled point, where the bias is calculated by Bayesian loss, and γ represents the weighting factor. This represents the step-prior learning loss.

[0028] Step S3-3: Every t iterations, the trained conditional frequency domain interactive target counting model is used to infer the training dataset described in step S1, generating reliable prior labels to guide the learning of the model output.

[0029] Step S3-4: Validate the conditional frequency domain interactive target counting model using samples from the validation dataset, and save the model parameters with the smallest error.

[0030] Step S3-5: Return to step S3-2 until training is complete.

[0031] Preferred method: In step S3-1, the prior labels of the samples are initialized, specifically as follows:

[0032] For the i-th sample x i The prior label {Ind(x i Den(x) i Let Ind(x) i ) = 0, Den(x) i ) = 0, where Ind(x) i ) represents the current sample x i Does the label have a prior label? Den(x) i ) represents the current sample x i Density map.

[0033] Preferred method: In step S3-3, a reliable prior label is generated, specifically through the following steps:

[0034] Step S3-3-1: Transfer the i-th sample x i density map m(x) i Sum all elements and add them together with g. t (x i By comparing the number of targets, we can obtain the current sample x. i The prior label {Ind(x i Den(x) i The specific calculation formula is as follows:

[0035]

[0036] In the formula, δ is the scaling factor.

[0037] Step S3-3-2: Load the current sample x i Existing prior label {Ind(x i ) old Den(x) i ) old}; if {Ind(x i ) old =0, then save {Ind(x)}i Den(x) i )}, denoted as {Ind(x i ) old Den(x) i ) old}; If Ind(x i ) old =1,Ind(x) i ) = 1 and Then save {Ind(x)} i Den(x) i )}, denoted as {Ind(x i ) old Den(x) i ) old}

[0038] Step S3-3-3: When the iteration number is τ, load the saved historical prior label {Ind(x i ) old Den(x) i ) old The network's learning at the current stage is guided by the following loss function:

[0039]

[0040] In the formula, Φ(·,·) represents the SSIM loss function.

[0041] Preferred method: In step S1, data is collected using existing datasets and real-world scenarios, and unlabeled samples are manually labeled. The labeled data is then divided into training datasets, validation datasets, and test datasets.

[0042] A conditional frequency domain interactive target counting system based on step prior learning comprises an input unit, a conditional frequency domain interactive target counting model unit, a step prior learning strategy unit, and an output unit, wherein:

[0043] The input unit is used to input the target image and label it to establish a training dataset, a validation dataset, and to input the target image to be counted.

[0044] The conditional frequency domain interactive target counting model unit is equipped with a conditional frequency domain interactive target counting model, which includes a feature extractor, a conditional frequency domain interaction module, and a regressor. The feature extractor is used to extract high-level semantic features of the input samples, the conditional frequency domain interaction module is used to enhance global and local features, and the regressor is used to regress the final density map. The conditional frequency domain interactive target counting model is trained using a training dataset guided by a step-prior learning strategy. The best-performing frequency domain adaptive target counting model on the evaluation metrics is selected using a validation dataset and saved to obtain the final frequency domain adaptive target counting model. The target image to be counted is input into the final frequency domain adaptive target counting model to obtain the target count.

[0045] The step-prior learning strategy unit is used to guide the learning of the conditional frequency domain interactive target counting model during the training process using a step-prior learning strategy.

[0046] The output unit is used to output the target count.

[0047] A conditional frequency domain interactive target counting device based on step prior learning includes 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 conditional frequency domain interactive target counting method based on step prior learning.

[0048] Compared with the prior art, the present invention has the following advantages:

[0049] 1. This invention proposes a conditional frequency domain interaction module, which uses the features of the input samples as conditional mappings to guide the generation of learnable convolutional kernels, thereby enhancing the model's ability to extract global and local features.

[0050] 2. This invention proposes a step-by-step prior learning strategy, which uses a reliable intermediate-generated density map to guide the generation of the current density map, reducing the risk of density representation shift caused by regression without a definite density map, and achieving robust density regression. Attached Figure Description

[0051] Figure 1 This is a flowchart illustrating the conditional frequency domain interactive target counting method based on step prior learning provided by the present invention.

[0052] Figure 2 This is a schematic diagram of the overall structure of the conditional frequency domain interactive target counting model of the present invention. Detailed Implementation

[0053] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these examples are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0054] A conditional frequency domain interactive target counting method based on step-prior learning is proposed. First, a training sample library is established, utilizing existing datasets and data collected from real-world scenarios, with unlabeled samples manually labeled. Then, a conditional frequency domain interactive target counting model is designed to enhance its representational ability. Next, the conditional frequency domain interactive target counting model is trained using the training dataset, employing a step-prior learning strategy to guide its learning. Finally, a validation dataset is used to select and save the best-performing frequency domain adaptive target counting model based on evaluation metrics. After completing the model training, the final frequency domain adaptive target counting model is obtained. Figure 1 As shown, the specific steps include the following:

[0055] Step S1: Collect target images and label them to establish a training sample library, thereby obtaining the training dataset and the validation dataset.

[0056] Using existing datasets and data collected from real-world scenarios, unlabeled samples are manually labeled, and the labeled data is divided into training datasets, validation datasets, and test datasets.

[0057] Step S2: Design a conditional frequency domain interactive target counting model to enhance the model's representational ability, such as... Figure 2 As shown, the conditional frequency domain interactive target counting model includes a feature extractor, a conditional frequency domain interactive module, and a regressor. The feature extractor is used to extract high-level semantic features of the input samples, the conditional frequency domain interactive module is used to enhance global and local features, and the regressor is used to regress the final density map.

[0058] like Figure 2 As shown, the specific construction method of the conditional frequency domain interactive target counting model is as follows:

[0059] Step S2-1: Extract the features y(x) of the output sample x using the pre-trained network model, denoted as y(x) = φ(x, θ). y ), where φ(·,·) is the feature extractor, and θ y These are network parameters.

[0060] Step S2-2: Transform the features obtained in step S2-1 using a series of convolutional layers to obtain the transformed features. in For a series of convolution operations, θ z These are the parameters of the convolutional layer.

[0061] Step S2-3: Transform the feature z(x) obtained in step S3-2 to the frequency domain using the Fourier transform function, and obtain the real part feature R(x) and the imaginary part feature I(x) corresponding to the feature z(x).

[0062] Step S2-4: Obtain the conditional mapping by passing the output z(x) from step S2-2 through a 1×1 convolutional layer and an adaptive average pooling layer. Further multiplied by a learnable convolution kernel Determine the conditional convolution kernel

[0063] Step S2-5: Obtain the conditional mapping by passing the output z(x) from step S2-2 through a 1×1 convolutional layer and an adaptive average pooling layer. Further multiplied by a learnable convolution kernel Determine the conditional convolution kernel

[0064] Step S2-6: Convolve the real part features R(x) from step S3-3 using the conditional convolution kernel K0 from step S2-4, and then obtain the conditional real part features after passing through a series of convolutional layers.

[0065] Step S2-7: Convolve the real part feature I(x) from step S3-3 using the conditional convolution kernel K1 from step S2-5, and then obtain the conditional real part feature after passing through a series of convolutional layers.

[0066] Step S2-8: The conditional real part features obtained in step S2-6 are... and the conditional real part features obtained in steps S3-7 Perform an inverse Fourier transform, and then add the transformed feature z(x) from step S2-2 to obtain the feature.

[0067] Step S2-9: Output the features from step S2-8 After passing through the regressor, the density prediction map m(x) is obtained.

[0068] Step S3: Train the conditional frequency domain interactive target counting model described in step S2 using the training dataset. During the training process, the learning of the conditional frequency domain interactive target counting model is guided by a step-prior learning strategy.

[0069] Step S3-1: Set the parameters required for network training and initialize the prior labels of the samples.

[0070] Initialize the prior labels of the samples, specifically as follows:

[0071] For the i-th sample xi The prior label {Ind(x i Den(x) i Let Ind(x) i ) = 0, Den(x) i ) = 0, where Ind(x) i ) represents the current sample x i Does the label have a prior label? Den(x) i ) represents the current sample x i Density map.

[0072] Step S3-2: Feed the samples in the training dataset into the conditional frequency domain interactive target counting model in batches, and calculate the total loss. Then, the Adam optimizer is used to optimize the model.

[0073]

[0074] Where N represents the number of samples in the current batch, This indicates whether the j-th annotation point in the actual annotation is reliable. This represents the bias generated by the predicted density map at the j-th labeled point, where the bias is calculated by Bayesian loss, and γ represents the weighting factor. This represents the step-prior learning loss.

[0075] Step S3-3: Every t iterations, the trained conditional frequency domain interactive target counting model is used to infer the training dataset described in step S1, generating reliable prior labels to guide the learning of the model output.

[0076] Generate reliable prior labels, the specific steps are as follows:

[0077] Step S3-3-1: Transfer the i-th sample x i density map m(x) i Sum all elements and add them together with g. t (x i By comparing the number of targets, we can obtain the current sample x. i The prior label {Ind(x i Den(x) i The specific calculation formula is as follows:

[0078]

[0079] In the formula, δ is the scaling factor.

[0080] Step S3-3-2: Load the current sample x i Existing prior label {Ind(x i )old Den(x) i ) old}; if {Ind(x i ) old =0, then save {Ind(x)} i Den(x) i )}, denoted as {Ind(x i ) old Den(x) i ) old};;If Ind(x i ) old =1,Ind(x) i ) = 1 and Then save {Ind(x)} i Den(x) i )}, denoted as {Ind(x i ) old Den(x) i ) old}

[0081] Step S3-3-3: When the iteration number is τ, load the saved historical prior label {Ind(x i ) old Den(x) i ) old The network's learning at the current stage is guided by the following loss function:

[0082]

[0083] In the formula, Φ(·,·) represents the SSIM loss function.

[0084] Step S3-4: Validate the conditional frequency domain interactive target counting model using samples from the validation dataset, and save the model parameters with the smallest error.

[0085] Step S3-5: Return to step S3-2 until training is complete.

[0086] Step S4: Use the validation dataset to select and save the frequency domain adaptive target counting model that performs best on the evaluation metrics.

[0087] Step S5: Return to step S3 until the model training is complete and the final frequency domain adaptive target counting model is obtained.

[0088] Step S6: Perform performance testing on the model using the test dataset.

[0089] During counting, images of the target to be counted are acquired and input into the final frequency domain adaptive target counting model to obtain the target count.

[0090] In another embodiment, a conditional frequency domain interactive target counting system based on step prior learning is also provided, comprising an input unit, a conditional frequency domain interactive target counting model unit, a step prior learning strategy unit, and an output unit, wherein:

[0091] The input unit is used to input the target image and label it to establish a training dataset, a validation dataset, and to input the target image to be counted.

[0092] The conditional frequency domain interactive target counting model unit is equipped with a conditional frequency domain interactive target counting model, which includes a feature extractor, a conditional frequency domain interaction module, and a regressor. The feature extractor is used to extract high-level semantic features of the input samples, the conditional frequency domain interaction module is used to enhance global and local features, and the regressor is used to regress the final density map. The conditional frequency domain interactive target counting model is trained using a training dataset guided by a step-prior learning strategy. The best-performing frequency domain adaptive target counting model on the evaluation metrics is selected using a validation dataset and saved to obtain the final frequency domain adaptive target counting model. The target image to be counted is input into the final frequency domain adaptive target counting model to obtain the target count.

[0093] The step-prior learning strategy unit is used to guide the learning of the conditional frequency domain interactive target counting model during the training process using a step-prior learning strategy.

[0094] The output unit is used to output the target count.

[0095] In another embodiment, a conditional frequency domain interactive target counting device based on step prior learning is also provided, including a processor and a computer program stored in a memory and executable on the processor, wherein the processor implements the above method when executing the program.

[0096] This invention enhances the extraction of features from the target counting model, resulting in superior target counting performance compared to existing counting models.

[0097] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A conditional frequency domain interactive target counting method based on step prior learning, characterized in that, Includes the following steps: Step S1: Collect and label target images to establish training and validation datasets; Step S2: Design a conditional frequency domain interactive target counting model to enhance the model's representation ability. The conditional frequency domain interactive target counting model includes a feature extractor, a conditional frequency domain interaction module, and a regressor. The feature extractor is used to extract high-level semantic features of the input samples, the conditional frequency domain interaction module is used to enhance global and local features, and the regressor is used to regress the final density map. The specific construction method of the conditional frequency domain interactive target counting model is as follows: Step S2-1: Extract the features y(x) of the output sample x using the pre-trained network model, denoted as y(x) = φ(x, θ). y ), where φ(·,·) is the feature extractor, and θ y For network parameters; Step S2-2: Transform the features obtained in step S2-1 using a series of convolutional layers to obtain the transformed features. in For a series of convolution operations, θ z These are the parameters of the convolutional layer; Step S2-3: Transform the transformed feature z(x) from step S2-2 to the frequency domain using the Fourier transform function, and obtain the real part feature R(x) and the imaginary part feature I(x) corresponding to the transformed feature z(x); Step S2-4: Obtain the conditional mapping by passing the output z(x) from step S2-2 through a 1×1 convolutional layer and an adaptive average pooling layer. Further multiplied by a learnable convolution kernel Determine the conditional convolution kernel Step S2-5: Obtain the conditional mapping by passing the output z(x) from step S2-2 through a 1×1 convolutional layer and an adaptive average pooling layer. Further multiplied by a learnable convolution kernel Determine the conditional convolution kernel Step S2-6: Convolve the real part features R(x) from step S2-3 using the conditional convolution kernel K0 from step S2-4, and then obtain the conditional real part features after passing through a series of convolutional layers. Step S2-7: Convolve the real part feature I(x) from step S2-3 using the conditional convolution kernel K1 from step S2-5, and then obtain the conditional real part feature after passing through a series of convolutional layers. Step S2-8: The conditional real part features obtained in step S2-6 are... and the conditional real part features obtained in steps S2-7 Perform an inverse Fourier transform, and then add the transformed feature z(x) from step S2-2 to obtain the feature. Step S2-9: Output the features from step S2-8 After passing through the regressor, the density prediction map m(x) is obtained; Step S3: Train the conditional frequency domain interactive target counting model described in step S2 using the training dataset. During the training process, the learning of the conditional frequency domain interactive target counting model is guided by a step-prior learning strategy. Step S4: Use the validation dataset to select and save the frequency domain adaptive target counting model that performs best on the evaluation metrics; Step S5: Return to step S3 until the model training is complete and the final frequency domain adaptive target counting model is obtained; Step S6: Acquire the image of the target to be counted, and input the image of the target to be counted into the final frequency domain adaptive target counting model to obtain the target count.

2. The conditional frequency domain interactive target counting method based on step prior learning according to claim 1, characterized in that: The method for training the step-conditional frequency domain interactive target counting model using the training dataset in step S3 is as follows: Step S3-1: Set the parameters required for network training and initialize the prior labels of the samples; Step S3-2: Feed the samples in the training dataset into the conditional frequency domain interactive target counting model in batches, and calculate the total loss. Then the Adam optimizer is used to optimize the model; Where N represents the number of samples in the current batch, This indicates whether the j-th annotation point in the actual annotation is reliable. This represents the bias generated by the predicted density map at the j-th labeled point, where the bias is calculated by Bayesian loss, and γ represents the weighting factor. This indicates the step-prior learning loss; Step S3-3: Every t iterations, the trained conditional frequency domain interactive target counting model is used to infer the training dataset described in step S1 in sequence to generate reliable prior labels and guide the learning of the model output. Step S3-4: Validate the conditional frequency domain interactive target counting model using samples from the validation dataset, and save the model parameters with the smallest error; Step S3-5: Return to step S3-2 until training is complete.

3. The conditional frequency domain interactive target counting method based on step prior learning according to claim 2, characterized in that: In step S3-1, the prior labels of the samples are initialized. The specific operation is as follows: For the i-th sample x i The prior label {Ind(x i Den(x) i Let Ind(x) i ) = 0, Den(x) i ) = 0, where Ind(x) i ) represents the current sample x i Does the label have a prior label? Den(x) i ) represents the current sample x i Density map.

4. The conditional frequency domain interactive target counting method based on step prior learning according to claim 3, characterized in that: In step S3-3, a reliable prior label is generated. The specific operation is as follows: Step S3-3-1: Transfer the i-th sample x i density map m(x) i Sum all elements and add them together with g. t (x i By comparing the number of targets, we can obtain the current sample x. i Prior label Ind(x) i ),Den(x i The specific calculation formula is as follows: In the formula, δ is the scaling factor; Step S3-3-2: Load the current sample x i Existing prior label {Ind(x i ) old Den(x) i ) old If Ind(x) i ) old =0, then save {Ind(x)} i Den(x) i )}, denoted as {Ind(x i ) old Den(x) i ) old }; If Ind(x i ) old =1,Ind(x) i ) = 1 and Then save {Ind(x)} i Den(x) i )}, denoted as {Ind(x i ) old Den(x) i ) old }; Step S3-3-3: When the iteration number is τ, load the saved historical prior label {Ind(x i ) old Den(x) i ) old The network's learning at the current stage is guided by the following loss function: In the formula, Φ(·,·) represents the SSIM loss function.

5. The conditional frequency domain interactive target counting method based on step prior learning according to claim 4, characterized in that: In step S1, data is collected using existing datasets and real-world scenarios. Unlabeled samples are manually labeled, and the labeled data is divided into training datasets, validation datasets, and test datasets.

6. A conditional frequency domain interactive target counting system based on step prior learning, characterized in that, The conditional frequency domain interactive target counting method based on step prior learning, as described in any one of claims 1-5, includes an input unit, a conditional frequency domain interactive target counting model unit, a step prior learning strategy unit, and an output unit, wherein: The input unit is used to input the target image and label it to establish a training dataset and a validation dataset, as well as to input the target image to be counted; The conditional frequency domain interactive target counting model unit is equipped with a conditional frequency domain interactive target counting model, which includes a feature extractor, a conditional frequency domain interaction module, and a regressor. The feature extractor is used to extract high-level semantic features of the input samples, the conditional frequency domain interaction module is used to enhance global and local features, and the regressor is used to regress the final density map. The conditional frequency domain interactive target counting model is trained using a training dataset according to a step-prior learning strategy. The best-performing frequency domain adaptive target counting model on the evaluation metrics is selected and saved using a validation dataset to obtain the final frequency domain adaptive target counting model. The target image to be counted is input into the final frequency domain adaptive target counting model to obtain the target count. The step-prior learning strategy unit is used to guide the learning of the conditional frequency domain interactive target counting model during the training process using the step-prior learning strategy. The output unit is used to output the target count.

7. A conditional frequency domain interactive target counting device based on step prior learning, comprising a processor and a computer program stored in memory and executable on the processor, characterized in that: When the processor executes the program, it implements the conditional frequency domain interactive target counting method based on step prior learning as described in any one of claims 1-5.

Images