Conditional frequency domain interaction target counting method and device based on stepping prior learning

Abstract

The invention discloses a conditional frequency domain interaction target counting method and device based on stepping prior learning. The method comprises the following steps: firstly, collecting a target image, marking and establishing a training data set and a verification data set; a conditional frequency domain interaction target counting model is designed, the representation capacity of the model is enhanced, and the conditional frequency domain interaction target counting model comprises a feature extractor, a conditional frequency domain interaction module and a regression device; training the conditional frequency domain interaction target counting model by using a training data set, and guiding the learning of the conditional frequency domain interaction target counting model by using a stepping prior learning strategy in the training process; screening and storing the frequency domain self-adaptive target counting model with the best performance on the evaluation indexes by adopting the verification data set; and obtaining a final frequency domain adaptive target counting model. Compared with the existing counting model, the target counting effect is more excellent for the extraction of the features of the enhanced target counting model.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular to a conditional frequency domain interactive target counting method and device based on step-by-step prior learning. Background Art

[0002] The target aggregation phenomenon is widely present in many scenarios, including business analysis, public safety management, traffic management, plant phenotyping, pest analysis, etc. For example, during certain holidays or promotions, the number of customers in a shopping mall may increase dramatically, forming a high-density aggregation. This may lead to a decline in the shopping experience, increase safety hazards (such as stampedes), and also put pressure on the operation of merchants. During the migratory season of migratory birds every year, certain specific locations will become habitats for a large number of birds. The detection of these birds is conducive to the assessment of species diversity and ecological balance. In the face of the target aggregation phenomenon mentioned above, accurate dense target counting is crucial for assessing risks and formulating effective management measures.

[0003] The dense object counting task aims to automatically identify and count the number of densely distributed objects in images or videos through computer vision technology. In recent years, a lot of research has been carried out around this task, and fruitful results have been achieved. One of the current research focuses is to use Bayesian loss for regression based on point annotation, which has achieved excellent performance. However, this loss uses points as regression objects to generate density maps, but there is no clear density map target, which will cause distribution drift problems in the learning process, and thus cannot effectively deal with irregular targets and occluded targets. Summary of the invention

[0004] Purpose of the invention: In order to overcome the deficiencies in the prior art, the present invention provides a conditional frequency domain interactive target counting method and device based on step prior learning, which constructs a robust target counting model by utilizing the step prior learning strategy to improve the robustness and adaptability of the model in multiple fields.

[0005] Technical solution: To achieve the above purpose, the technical solution adopted by the present invention is:

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

[0007] Step S1: Collect target images and annotate them to establish training data sets and verification data sets.

[0008] Step S2: Design a conditional frequency domain interactive target counting model to enhance the representation ability of the model. 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.

[0009] Step S3: Use the training data set to train the conditional frequency-domain interactive target counting model described in step S2. During the training process, use the step-by-step prior learning strategy to guide the learning of the conditional frequency-domain interactive target counting model.

[0010] Step S4: Use the validation data set to select the frequency domain adaptive object counting model with the best performance in terms of evaluation indicators and save it.

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

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

[0013] Preferably, the specific method for constructing the conditional frequency domain interactive target counting model in step 2 is:

[0014] Step S2-1: Use the pre-trained network model to extract the feature y(x) of the output sample x, denoted as y(x)=φ(x,θ y ), where φ(·,·) is the feature extractor, θ y is the network parameter.

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

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

[0017] Step S2-4: The output z(x) of step S2-2 is passed through a 1×1 convolutional layer and an adaptive average pooling layer to obtain a conditional mapping Further multiply by the learnable convolution kernel Get the conditional convolution kernel

[0018] Step S2-5: The output z(x) of step S2-2 is passed through a 1×1 convolutional layer and an adaptive average pooling layer to obtain a conditional mapping Further multiply by the learnable convolution kernel Get the conditional convolution kernel

[0019] Step S2-6: Use the conditional convolution kernel K0 in step S2-4 to convolve the real feature R(x) in step S2-3, and then obtain the conditional real feature after a series of convolution layers.

[0020] Step S2-7: Use the conditional convolution kernel K1 in step S2-5 to convolve the real feature I(x) in step S2-3, and then obtain the conditional real feature after a series of convolution layers.

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

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

[0023] Preferably: the method for training the step-conditional frequency domain interactive target counting model using the training data set in step S3 is:

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

[0025] Step S3-2: Send the samples in the training data set into the conditional frequency domain interactive target counting model in batches and calculate the total loss Then the model is optimized using the Adam optimizer.

[0026]

[0027] Where N represents the number of samples in the current batch. Indicates whether the jth annotation point in the real annotation is reliable. represents the deviation of the predicted density map at the jth labeled point, where the deviation is calculated by Bayesian loss, γ represents the weight factor, represents the step-wise 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 data set described in step S1 in turn to generate reliable prior labels to guide the learning of the model output.

[0029] Step S3-4: Use the samples in the validation data set to validate the conditional frequency-domain interactive target counting model, and save the model parameters with the smallest error.

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

[0031] Preferably, the priori label of the sample is initialized in step S3-1, and the specific operation is:

[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 Whether the label of has a priori label, Den(x i ) represents the current sample x i Density map of .

[0033] Preferably, in step S3-3, a reliable priori label is generated, and the specific operation is as follows:

[0034] Step S3-3-1: Take the i-th sample x i The density map m(x i ) add all the elements together, and t (x i ) to obtain the current sample x i The prior label {Ind(x i ), Den(x i )}, the specific calculation formula is as follows:

[0035]

[0036] Where δ is the scale factor.

[0037] Step S3-3-2: Load the current sample x i The existing prior label {Ind(x i ) old ,Den(x i ) old}; if {Ind(x i ) old = 0, then save {Ind(xi ), Den(x i )}, recorded 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 )}, recorded as {Ind(x i ) old ,Den(x i ) old}.

[0038] Step S3-3-3: When the number of iterations is τ, load the saved historical prior labels {Ind(x i ) old ,Den(x i ) old}, guide the learning of the current network through the following loss function:

[0039]

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

[0041] Preferably, in step S1, existing data sets and real scene acquisition data are used to manually annotate unlabeled samples, and the annotated data are divided into a training data set, a verification data set and a test data set.

[0042] A conditional frequency domain interactive target counting system based on step prior learning, comprising 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 a target image and annotate it to establish a training data set and a verification data set, and to input a target image to be counted.

[0044] The conditional frequency domain interactive target counting model unit is provided with a conditional frequency domain interactive target counting model, and the conditional frequency domain interactive target counting model includes a feature extractor, a conditional frequency domain interactive module and a regressor, wherein the feature extractor is used to extract high-level semantic features of the input sample, the conditional frequency domain interactive module is used to strengthen global and local features, and the regressor is used to regress the final density map. The training data set is used to guide the training of the conditional frequency domain interactive target counting model according to the step prior learning strategy, and the verification data set is used to screen the frequency domain adaptive target counting model with the best performance in the evaluation index and save it 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-by-step a priori learning strategy unit is used to guide the learning of the conditional frequency-domain interactive target counting model using the step-by-step a priori learning strategy during the training process.

[0046] The output unit is used for outputting 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, wherein the processor implements the conditional frequency domain interactive target counting method based on step prior learning when executing the program.

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

[0049] 1. The present invention proposes a conditional frequency domain interaction module, which uses the characteristics of the input samples as conditional mapping to guide the generation of learnable convolution kernels and enhance the model's extraction of global and local features.

[0050] 2. The present 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, reduces the risk of density representation deviation caused by the regression of an uncertain density map, and achieves robust density regression. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a flow chart of the conditional frequency domain interactive target counting method based on step-by-step prior learning provided by the present invention.

[0052] Figure 2 It is a schematic diagram of the overall structure of the conditional frequency domain interactive target counting model of the present invention. DETAILED DESCRIPTION

[0053] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these examples are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0054] A conditional frequency domain interactive target counting method based on step-by-step prior learning is firstly established, and the unlabeled samples are manually labeled using the existing data set and the real scene collection data; then, the conditional frequency domain interactive target counting model is designed to enhance the representation ability of the model; further, the conditional frequency domain interactive target counting model is trained using the training data set, and during the training process, the step-by-step prior learning strategy is used to guide the learning of the conditional frequency domain interactive target counting model; furthermore, the frequency domain adaptive target counting model with the best performance in the evaluation index is selected using the verification data set and saved; after completing the model training, the final frequency domain adaptive target counting model is obtained, such as Figure 1 As shown, the specific steps include:

[0055] Step S1: Collect target images and annotate them to establish a training sample library to obtain a training data set and a verification data set.

[0056] The existing data sets and real scene collection data are used to manually annotate the unlabeled samples, and the labeled data are divided into a training data set, a verification data set, and a test data set.

[0057] Step S2: Design a conditional frequency domain interactive target counting model to enhance the representation ability of the model, 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 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 in Figure 2, the specific construction method of the conditional frequency domain interactive target counting model is:

[0059] Step S2-1: Use the pre-trained network model to extract the feature y(x) of the output sample x, denoted as y(x)=φ(x,θ y ), where φ(·,·) is the feature extractor, θ y is the network parameter.

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

[0061] Step S2-3: The feature z(x) in step S3-2 is converted to the frequency domain through the Fourier transform function, and the real feature R(x) and the imaginary feature I(x) corresponding to the feature z(x) are obtained respectively.

[0062] Step S2-4: The output z(x) of step S2-2 is passed through a 1×1 convolutional layer and an adaptive average pooling layer to obtain a conditional mapping Further multiply by the learnable convolution kernel Get the conditional convolution kernel

[0063] Step S2-5: The output z(x) of step S2-2 is passed through a 1×1 convolutional layer and an adaptive average pooling layer to obtain a conditional mapping Further multiply by the learnable convolution kernel Get the conditional convolution kernel

[0064] Step S2-6: Use the conditional convolution kernel K0 in step S2-4 to convolve the real feature R(x) of step S3-3, and then obtain the conditional real feature after a series of convolution layers.

[0065] Step S2-7: Use the conditional convolution kernel K1 in step S2-5 to convolve the real feature I(x) in step S3-3, and then obtain the conditional real feature after a series of convolution layers.

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

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

[0068] Step S3: Use the training data set to train the conditional frequency-domain interactive target counting model described in step S2. During the training process, use the step-by-step prior learning strategy to guide the learning of the conditional frequency-domain interactive target counting model.

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

[0070] Initialize the prior label of the sample. The specific operation is:

[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 Whether the label of has a priori label, Den(x i ) represents the current sample x i Density map of .

[0072] Step S3-2: Send the samples in the training data set into the conditional frequency domain interactive target counting model in batches and calculate the total loss Then the model is optimized using the Adam optimizer.

[0073]

[0074] Where N represents the number of samples in the current batch. Indicates whether the jth annotation point in the real annotation is reliable. represents the deviation of the predicted density map at the jth labeled point, where the deviation is calculated by Bayesian loss, γ represents the weight factor, represents the step-wise 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 data set described in step S1 in turn to generate reliable prior labels to guide the learning of the model output.

[0076] Generate reliable prior labels. The specific operations are:

[0077] Step S3-3-1: Take the i-th sample x i The density map m(x i ) add all the elements together, and t (x i ) to obtain the current sample x i The prior label {Ind(x i ), Den(x i )}, the specific calculation formula is as follows:

[0078]

[0079] Where δ is the scale factor.

[0080] Step S3-3-2: Load the current sample x i The 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 )}, recorded 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 )}, recorded as {Ind(x i ) old ,Den(x i ) old}.

[0081] Step S3-3-3: When the number of iterations is τ, load the saved historical prior labels {Ind(x i ) old ,Den(x i ) old}, guide the learning of the current network through the following loss function:

[0082]

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

[0084] Step S3-4: Use the samples in the validation data set to validate the conditional frequency-domain interactive target counting model, and save the model parameters with the smallest error.

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

[0086] Step S4: Use the validation data set to select the frequency domain adaptive object counting model with the best performance in terms of evaluation indicators and save it.

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

[0088] Step S6: Use the test data set to perform performance test on the model.

[0089] When counting, the target image to be counted is collected, and the target image to be counted is 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 provided, including 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 a target image and annotate it to establish a training data set and a verification data set, and to input a target image to be counted.

[0092] The conditional frequency domain interactive target counting model unit is provided with a conditional frequency domain interactive target counting model, and the conditional frequency domain interactive target counting model includes a feature extractor, a conditional frequency domain interactive module and a regressor, wherein the feature extractor is used to extract high-level semantic features of the input sample, the conditional frequency domain interactive module is used to strengthen global and local features, and the regressor is used to regress the final density map. The training data set is used to guide the training of the conditional frequency domain interactive target counting model according to the step prior learning strategy, and the verification data set is used to screen the frequency domain adaptive target counting model with the best performance in the evaluation index and save it 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-by-step a priori learning strategy unit is used to guide the learning of the conditional frequency-domain interactive target counting model using the step-by-step a priori learning strategy during the training process.

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

[0095] In another embodiment, a conditional frequency-domain interactive target counting device based on step-by-step prior learning is provided, comprising 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] The present invention enhances the extraction of target counting model features, and the target counting effect is better than that of the existing counting model.

[0097] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A conditional frequency domain interactive target counting method based on step-by-step prior learning, characterized in that: The steps include: Step S1: Collect target images and annotate them to establish training data sets and verification data sets; Step S2: designing a conditional frequency domain interactive target counting model to enhance the representation capability of the model, wherein the conditional frequency domain interactive target counting model comprises a feature extractor, a conditional frequency domain interactive module and a regressor, wherein the feature extractor is used to extract high-level semantic features of 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; Step S3: using the training data set to train the conditional frequency domain interactive target counting model in step S2, and during the training process, using the step-by-step prior learning strategy to guide the learning of the conditional frequency domain interactive target counting model; Step S4: Use the validation data set to select the frequency domain adaptive target counting model with the best performance in terms of evaluation indicators and save it; Step S5: Return to step S3 until the model training is completed to obtain the final frequency domain adaptive target counting model; Step S6: Acquire the target image to be counted, and input the target image 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-by-step prior learning according to claim 1 is characterized in that: The specific construction method of the conditional frequency domain interactive target counting model in step 2 is: Step S2-1: Use the pre-trained network model to extract the feature y(x) of the output sample x, denoted as y(x)=φ(x,θy), where φ(·,·) is the feature extractor, θ y is the network parameter; Step S2-2: Use a series of convolutional layers to transform the features obtained in step S2-1 to obtain the transformed features in is a series of convolution operations, θ z are the parameters of the convolutional layer; Step S2-3: Convert the transformed feature z(x) in step S2-2 to the frequency domain through the Fourier transform function, and obtain the real feature R(x) and the imaginary feature I(x) corresponding to the transformed feature z(x); Step S2-4: The output z(x) of step S2-2 is passed through a 1×1 convolutional layer and an adaptive average pooling layer to obtain a conditional mapping Further multiply by the learnable convolution kernel Get the conditional convolution kernel Step S2-5: The output z(x) of step S2-2 is passed through a 1×1 convolutional layer and an adaptive average pooling layer to obtain a conditional mapping Further multiply by the learnable convolution kernel Get the conditional convolution kernel Step S2-6: Use the conditional convolution kernel K in step S2-4 o Convolve the real part feature R(x) of step S2-3, and then obtain the conditional real part feature after a series of convolution layers Step S2-7: Use the conditional convolution kernel K1 in step S2-5 to convolve the real feature I(x) in step S2-3, and then obtain the conditional real feature after a series of convolution layers. Step S2-8: The conditional real part feature obtained in step S2-6 and the conditional real part feature obtained in step S2-7 Perform inverse Fourier transform and add the transformed feature z(x) in step S2-2 to obtain the feature Step S2-9: Output the features in step S2-8 After passing through the regressor, the density prediction map m(x) is obtained.

3. The conditional frequency domain interactive target counting method based on step-by-step prior learning according to claim 2 is characterized in that: The method for training the step-conditional frequency-domain interactive target counting model using the training data set 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: Send the samples in the training data set into the conditional frequency domain interactive target counting model in batches and calculate the total loss Then use the Adam optimizer to optimize the model; Where N represents the number of samples in the current batch. Indicates whether the jth annotation point in the real annotation is reliable. represents the deviation of the predicted density map at the jth labeled point, where the deviation is calculated by Bayesian loss, γ represents the weight factor, represents the step-wise prior learning loss; Step S3-3: every cloud iteration, the trained conditional frequency domain interactive target counting model is used to sequentially infer the training data set described in step S1 to generate reliable prior labels to guide the learning of model output; Step S3-4: Use the samples in the validation data set to validate the conditional frequency domain interactive target counting model, and save the model parameters with the smallest error; Step S3-5: Return to step S3-2 until the training is completed.

4. The conditional frequency domain interactive target counting method based on step-by-step prior learning according to claim 3 is characterized in that: In step S3-1, the prior labels of the samples are initialized. The specific operations are: 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( xi ) represents the current sample x i Whether the label of has a priori label, Den(x i ) represents the current sample x i Density map of .

5. The conditional frequency domain interactive target counting method based on step-by-step prior learning according to claim 4 is characterized in that: In step S3-3, a reliable priori label is generated, and the specific operations are as follows: Step S3-3-1: Take the i-th sample x i The density map m(x i ) add all the elements together, and t (x i ) to obtain the current sample x i The prior label {Ind(x i ), Den(x i )}, the specific calculation formula is as follows: Where δ is the scale factor; Step S3-3-2: Load the current sample x i The 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 )}, recorded 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 )}, recorded as {Ind(x i ) old ,Den(x i ) old }; Step S3-3-3: When the number of iterations is τ, load the saved historical prior labels {Ind(x i ) old ,Den(x i ) old }, guide the learning of the current network through the following loss function: Where Φ(·, ·) represents the SSIM loss function.

6. The conditional frequency domain interactive target counting method based on step-by-step prior learning according to claim 5 is characterized in that: In step S1, existing data sets and real scene collection data are used to manually annotate unlabeled samples, and the labeled data are divided into a training data set, a verification data set, and a test data set.

7. A conditional frequency domain interactive target counting system based on step-by-step prior learning, characterized in that: Input unit, conditional frequency domain interactive target counting model unit, step prior learning strategy unit, output unit, where: The input unit is used to input the target image and annotate it to establish a training data set and a verification data set, and input the target image to be counted; The conditional frequency domain interactive target counting model unit is provided with a conditional frequency domain interactive target counting model, and the conditional frequency domain interactive target counting model includes a feature extractor, a conditional frequency domain interactive module and a regressor, wherein the feature extractor is used to extract high-level semantic features of input samples, the conditional frequency domain interactive module is used to strengthen global and local features, and the regressor is used to regress the final density map; the training data set is used to guide the training of the conditional frequency domain interactive target counting model according to the step-by-step prior learning strategy, and the verification data set is used to screen the frequency domain adaptive target counting model with the best performance on the evaluation index and save it 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-by-step prior learning strategy unit is used to guide the learning of the conditional frequency domain interactive target counting model using the step-by-step prior learning strategy during the training process; The output unit is used for outputting the target count.

8. A conditional frequency domain interactive target counting device based on step-by-step prior learning, 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, the conditional frequency domain interactive target counting method based on step-by-step prior learning described in any one of claims 1-6 is implemented.

Images