Similarity estimation method based on multiple similarity mechanisms

By combining the similarity evaluation method of static and dynamic features, using measurement learning and multi-head attention mechanism, the problem of difficulty in constructing multi-angle evaluation indicators in the existing technology is solved, the target recognition and tracking accuracy of infrared tracking technology is improved, and the Test cost.

CN119025930APending Publication Date: 2024-11-26NO 8511 RES INST OF CASIC
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411006993.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-25
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

It is difficult for existing infrared tracking technology to construct effective similarity evaluation indicators from multiple angles in complex scenarios, resulting in insufficient target recognition and tracking accuracy, especially when there are many historical test samples.

Method used

Using a similarity estimation method based on multiple similarity mechanisms, through the combination of static and dynamic features, metric learning and multi-head attention mechanism, static feature consistency and dynamic feature similarity are calculated, and multiple evaluation indicators are fused to improve similarity Accuracy of the assessment.

Benefits of technology

It has achieved the measurement of the similarity between the current test samples and the historical test samples from multiple angles, improved the accuracy of static feature consistency matching accuracy and the accuracy of dynamic feature similarity evaluation, reduced repeatable tests, and reduced test costs.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119025930A_ABST
    Figure CN119025930A_ABST
Patent Text Reader

Abstract

The invention discloses a similarity estimation method based on multiple similarity mechanisms. The method comprises the following steps: firstly, constructing a feature space of a current test sample and a historical test sample in an infrared tracking process; secondly, establishing a measurement network based on static feature consistency; obtaining dynamic characteristics of the screened historical test sample and the current test sample; calculating a dynamic time bending distance similarity value; calculating a shape similarity value; calculating a cosine dynamic similarity value; and finally fusing a plurality of dynamic similarity evaluation results. And obtaining a similarity evaluation result between the current test sample and different historical data. According to the method, from the dynamic angle and the static angle, part of inconsistent historical data is eliminated through the static features, and the dynamic feature similarity of the test sample and the historical data is measured from multiple angles, so that the similarity matching efficiency is effectively improved, and the accuracy of similarity estimation of the infrared tracking test sample and the historical sample is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the field of infrared tracking, in particular to a similarity estimation method based on multiple similarity mechanisms. Background Art

[0002] As an important part of modern perception and monitoring systems, infrared tracking technology is widely used in military, security and other fields. Its core is to accurately track and locate the target by capturing the infrared radiation emitted by the target object. With the continuous development of technology, infrared tracking systems are facing more and more complex scenes and diverse challenges, so the requirements for target recognition and tracking accuracy are also getting higher and higher. In the research and development of infrared tracking technology, the similarity evaluation method between current test samples and historical test samples plays a vital role. By comparing and analyzing the similarity between current test samples and historical test samples, researchers can evaluate the performance of tracking algorithms, optimize parameter settings, and improve tracking accuracy. At the same time, similarity evaluation methods can also be used to verify the effectiveness of new algorithms, providing strong support for the further development of infrared tracking technology. However, most existing evaluation methods only measure similarity from a single indicator. How to construct different evaluation indicators from multiple perspectives and effectively integrate the results of multiple evaluation indicators still needs to be explored, especially when there are many historical test samples. Summary of the invention

[0003] The present invention proposes a similarity estimation method based on multiple similarity mechanisms, in order to measure the similarity between the current test sample and the historical test sample from different angles, thereby effectively improving the accuracy of similarity evaluation.

[0004] The technical solution to realize the present invention is: a similarity estimation method based on multiple similarity mechanisms, the steps are as follows:

[0005] Step 1: Construct the feature space of current test samples and historical test samples in the infrared tracking process.

[0006] Step 2: Establish a measurement network based on static feature consistency and calculate the static feature vector X of the current test sample. now and the static feature vector X of multiple historical test samples his Processing is performed to calculate the static feature similarity sim between the current test sample and each historical test sample, sim={sim1,sim2,...,sim j ,...,sim M}, and filter out historical test samples that meet the set similarity threshold λ, and obtain the filtered Z historical test samples, where sim j Represents the static feature vector X of the current test sample nowThe similarity between the static feature vector of the jth historical test sample and the historical test sample.

[0007] Step 3: Obtain the dynamic characteristics of the screened historical test samples and the current test samples.

[0008] Step 4: Calculate the dynamic feature vector TK of the current test sample now Each dynamic feature in the qth historical test sample dynamic feature vector TK his,q After summing up the similarity of each dynamic feature in TK now and TK his,q The dynamic time warping distance similarity value DTW (TK now ,TK his,q ).

[0009] Step 5: Calculate TK using Euclidean distance now Each dynamic feature in the qth historical test sample dynamic feature vector TK his,q The grey relational degree of each dynamic feature in TK now and TK his,q The shape similarity value S(TK now ,TK his,q ).

[0010] Step 6: Calculate TK using the cosine formula now and the dynamic feature vector TK of the qth historical test sample his,q The cosine dynamic similarity value A(TK now ,TK his,q ).

[0011] Step 7: Use formula (5) to get TK now and TK his,q The similarity calculation result θ q (TK now ,TK his,q ):

[0012] θ q (TK now ,TK his,q )=αDTW(TK now ,TK his,q )+βS(TK now ,TK his,q )+γA(TK now ,TK his,q )(5)

[0013] Among them, α, β, and γ are three weighted weights; and α+β+γ=1 is satisfied.

[0014] Step 8: For the Z similarity calculation results, the expression is as follows:

[0015] θ1(TK now ,TK his,1 ),θ2(TK now ,TK his,2 ),...,θ q (TK now ,TK his,q ),...,θ Z (TK now ,TK his,Z )

[0016] Sort and return the similarity evaluation result between the current test sample and the previous T historical test samples.

[0017] Compared with the prior art, the present invention has the following significant advantages:

[0018] (1) The present invention measures the similarity between the current test sample and the historical test sample from the perspectives of static and dynamic features. First, some historical test samples are screened by static feature consistency matching, thereby reducing the time required for dynamic feature similarity calculation. Then, a variety of dynamic feature similarity evaluation indicators are used to comprehensively measure the similarity between the current test sample and the historical test sample, thereby achieving a more accurate measurement.

[0019] (2) The present invention constructs a static feature consistency measurement index between the current test sample and the historical test sample through metric learning. This method can alleviate the problem that the traditional Euclidean distance or cosine similarity measurement method is difficult to accurately capture the similarity relationship between data due to the complex nonlinear distribution and diversity of the data, thereby effectively improving the static feature consistency matching accuracy.

[0020] (3) The present invention establishes a dynamic feature similarity evaluation method between the current test sample and the historical test sample through a multi-head attention mechanism, which can dynamically focus on different parts in the dynamic feature sequence and capture long-distance dependencies in the dynamic features, thereby improving the accuracy of dynamic feature similarity.

[0021] (4) The present invention proposes a similarity measurement method for current test samples and historical test samples. The method can improve the utilization rate of historical similar test samples and corresponding historical test data according to the similarity measurement results, reduce repetitive tests, and reduce test costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 The present invention provides a flowchart of a similarity estimation method based on multiple similarity mechanisms.

[0023] Figure 2 This is a flowchart of the static feature consistency matching based on metric learning proposed by the present invention.

[0024] Figure 3 This is the structural diagram of the dynamic similarity evaluation network based on the multi-head attention mechanism proposed in this invention. DETAILED DESCRIPTION

[0025] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0026] The following will further introduce the specific implementation method, as well as the technical difficulties and inventive points of this invention in combination with this design example.

[0027] In this embodiment, a similarity estimation method based on multiple similarity mechanisms comprehensively considers factors such as the performance of infrared tracking equipment, target characteristics, environment, and possible interference during infrared tracking, and integrates multiple similarity indicators to estimate the similarity between the test sample and the historical sample from both static and dynamic perspectives. The steps are as follows (e.g. Figure 1 ):

[0028] Step 1: Construct the feature space of the current test sample and the historical test sample in the infrared tracking process:

[0029] Step 1.1: Obtain the static feature vector X of the current test sample during infrared tracking now ={x now,1 ,x now,2 ,...,x now,i ,...,x now,N} and the static feature vectors X of multiple historical test samples his ={X his,1 ,X his,2 ,...,X his,j ,...,X his,M}, where x now,i Represents X now The i-th static feature in X his,j represents the static feature vector of the jth historical test sample among multiple historical test samples, and X his,j ={x his_j,1 ,x his_j,2 ,...,x his_j,i ,...,x his_j,N}, x his_j,i Represents X his,j The i-th static feature in , M represents the total number of historical test samples, and N represents the total number of static features.

[0030] Step 1.2: Obtain the dynamic feature vector Y of the current test sample during infrared tracking now ={y now,1 ,y now,2 ,...,y now,h ,...,y now,K} and the dynamic feature vector Y of multiple historical test samples his = {Y his,1 ,Y his,2 ,...,Y his,j ,...,Y his,M}, where y now,h Represents Y now The h-th dynamic feature in Y his,j represents the dynamic feature vector of the jth historical test sample, and Y his,j ={y his_j,1 ,y his_j,2 ,...,y his_j,h ,...,y his_j,K},y his_j,h Represents Y his,j is the hth dynamic feature in , and K represents the total number of dynamic features.

[0031] Step 2: Establish a measurement network based on static feature consistency and calculate the static feature vector X of the current test sample. now and the static feature vector X of multiple historical test samples his Processing is performed to calculate the static feature similarity sim = {sim1, sim2, ..., sim j ,...,sim M}, and filter out historical test samples that meet the set similarity threshold λ, and obtain the filtered Z historical test samples, where sim j Represents the static feature vector X of the current test sample now The similarity between the static feature vector of the jth historical test sample.

[0032] Step 2.1: Figure 2 As shown, a measurement network based on static feature consistency is established, including: two branches;

[0033] Among them, the first branch includes in sequence: the first embedding layer, a one-dimensional convolution layer with a step size of s and a convolution kernel of k×1, the first BN normalization layer, the first ReLu nonlinear activation layer, a maximum pooling layer with a convolution kernel of k×1, and the first fully connected layer.

[0034] The second branch includes: a second embedding layer, a one-dimensional convolution layer with a step size of s and a convolution kernel of k×1, a second BN normalization layer, a second ReLu nonlinear activation layer, a maximum pooling layer with a convolution kernel of k×1, and a second fully connected layer.

[0035] The current test sample and each historical test sample are input into the metric network for processing in turn, and the static feature vector F of the current test sample is obtained accordingly. now ={f now,1 ,f now,2 ,...,f now,i ,...,f now,N} and the static feature vector F of multiple historical test samples his ={F his,1 ,F his,2 ,...,F his,j ,...,F his,M}, where f now,i represents the i-th static feature of the current test sample, F his,j represents the static feature vector of the jth historical test sample among multiple historical test samples, and F his,j ={f his_j,1 ,f his_j,2 ,...,f his_j,p ,...,f his_j,N}, f his_j,p Indicates F his,j The pth static feature in .

[0036] Step 2.2: Based on F now , X now Construct a contrast loss function for the static vector of the current test sample based on F his , X his A contrast loss function for static vectors of historical test samples is constructed, so that the metric network is trained using the gradient descent method, and the contrast loss function is calculated to update the network parameters until the contrast loss function converges, thereby obtaining a trained metric network model.

[0037] Step 2.3: Calculate F using the cosine function now and F his,j Static feature similarity sim j (F now ,F his,j ), so we get F now The static feature similarity of the static feature vectors of the M historical test samples is obtained, and according to the static similarity threshold λ, the historical test samples with high static similarity to the current test sample are selected to obtain the static feature vectors G of the selected Z historical test samples. his = {G his,1 ,Ghis,2 ,...,G his,q ,...,G his,Z}, where G his,q represents the static feature vector of the qth historical test sample after screening, and G his,q = {G his_q,1 ,G his_q,2 ,...,G his_q,r ,...,G his_q,N}, G his_q,r Represents G his,q The rth static feature of , Z represents the total number of historical test samples after screening, Z≤M;

[0038] Step 2.4: Based on the static feature vector G of the Z historical test samples after screening his = {G his,1 ,G his,2 ,...,G his,q ,...,G his,Z}, and obtain the dynamic feature vector E of the Z historical test samples after screening his ={E his,1 ,E his,2 ,...,E his,q ,...,E his,Z}, where E his,q Represents the dynamic feature vector of the qth historical test sample after screening.

[0039] Step 3: Obtain the dynamic characteristics of the filtered historical test samples and the current test samples:

[0040] Step 3.1: Figure 3 As shown, a dynamic feature similarity calculation model is constructed, including: a pre-trained word segmenter, a pre-trained language representation model, a fully connected layer, and an activation function.

[0041] Step 3.2: The dynamic feature vector Y of the current test sample now = {Y now,1 ,Y now,2 ,...,Y now,h ,...,Y now,K} is input into the dynamic feature similarity calculation model, and the dynamic feature vector TK of the current test sample is obtained using formula (1): now ={TK now,1 ,TK now,2 ,...,TK now,h ,...,TK now,K}, where TK now,h Indicates TK now The h-th dynamic feature of .

[0042] TKnow =FC(RELU(FC(BERT(Tokenizer(Y now ))))) (1)

[0043] In formula (1), FC represents the fully connected layer, RELU represents the nonlinear activation function, BERT represents the pre-trained language representation model, and Tokenizer represents the word segmenter.

[0044] Step 3.3: The dynamic feature vectors E of the selected historical test samples are his ={E his,1 ,E his,2 ,...,E his,q ,...,E his,Z} is input into the dynamic feature similarity calculation model, and the dynamic feature vector TK of the selected multiple historical test samples is obtained using formula (2): his ={TK his,1 ,TK his,2 ,...,TK his,q ,...,TK his,Z}, where TK his,q Indicates TK his The qth dynamic eigenvector of , and TK his,q ={TK his_q,1 ,TK his_q,2 ,...,TK his_q,h ,...,TK his_q,K}, TK his_q,h Indicates TK his,q The h-th dynamic feature of

[0045] TK his,q =FC(RELU(FC(BERT(Tokenizer(E his,q ))))) (2)

[0046] Step 4: Calculate the dynamic feature vector TK of the current test sample now Each dynamic feature in the qth historical test sample dynamic feature vector TK his,q After summing up the similarity of each dynamic feature in TK now and TK his,q The dynamic time warping distance similarity value DTW (TK now ,TK his,q );

[0047]

[0048] Step 5: Calculate TK using Euclidean distance now Each dynamic feature in the qth historical test sample dynamic feature vector TKhis,q The grey relational degree of each dynamic feature in TK now and TK his,q The shape similarity value S(TK now ,TK his,q );

[0049] Step 6: Calculate TK using the cosine formula now and the dynamic feature vector TK of the qth historical test sample his,q The cosine dynamic similarity value A(TK now ,TK his,q ):

[0050]

[0051] T stands for transpose.

[0052] Step 7: Use formula (5) to get TK now and TK his,q The similarity calculation result θ q (TK now ,TK his,q ):

[0053] θ q (TK now ,TK his,q )=αDTW(TK now ,TK his,q )+βS(TK now ,TK his,q )+γA(TK now ,TK his,q )(5)

[0054] In formula (5), α, β, and γ are three weighted weights, and they satisfy α+β+γ=1.

[0055] Step 8: Calculate the Z similarity results θ1(TK now ,TK his,1 ),θ2(TK now ,TK his,2 ),...,θ q (TK now ,TK his,q ),...,θ Z (TK now ,TK his,Z ) to sort and return the similarity evaluation result between the current test sample and the previous T historical test samples.

Claims

1. A similarity estimation method based on multiple similarity mechanisms, characterized in that: Here are the steps: Step 1: Construct the feature space of the current test sample and the historical test samples in the infrared tracking process; Step 2: Establish a measurement network based on static feature consistency and calculate the static feature vector X of the current test sample. now and the static feature vector X of multiple historical test samples his Processing is performed to calculate the static feature similarity sim between the current test sample and each historical test sample, sim={sim1,sim2,...,sim j ,...,sim M }, and filter out historical test samples that meet the set similarity threshold λ, and obtain the filtered Z historical test samples, where sim j Represents the static feature vector X of the current test sample now The similarity between the static feature vector of the jth historical test sample and the historical test sample; Step 3: Obtain dynamic characteristics of the filtered historical test samples and the current test samples; Step 4: Calculate the dynamic feature vector TK of the current test sample now Each dynamic feature in the qth historical test sample dynamic feature vector TK his,q After summing the similarity of each dynamic feature in TK now and TK his,q The dynamic time warping distance similarity value DTW (TK now ,TK his,q ); Step 5: Calculate TK using Euclidean distance now Each dynamic feature in the qth historical test sample dynamic feature vector TK his,q The grey relational degree of each dynamic feature in TK is obtained now and TK his,q The shape similarity value S(TK now ,TK his,q ); Step 6: Calculate TK using the cosine formula now and the dynamic feature vector TK of the qth historical test sample his,q The cosine dynamic similarity value A(TK now ,TK his,q ); Step 7: Use formula (5) to get TK now and TK his,q The similarity calculation result θ q (TK now ,TK his,q ): θ q (K now ,TK his,q )=αDTW(TK now ,TK his,q )+βS(TK now ,TK his,q )+γA(TK now ,TK his,q )(5) Among them, α, β, γ are three weighted weights; and α+β+γ=1; Step 8: Calculate the Z similarity results. The expression is as follows: θ1(TK now ,TK his,1 ),θ2(TK now ,TK his,2 ),...,θ q (TK now ,TK his,q ),...,θ Z (TK now ,TK his,Z ) and returns the similarity evaluation result between the current test sample and the previous T historical test samples.

2. The similarity estimation method based on multiple similarity mechanisms according to claim 1, characterized in that: In step 1, the feature space of the current test sample and the historical test samples in the infrared tracking process is constructed as follows: Step 1.1: Obtain the static feature vector X of the current test sample during infrared tracking now ={x now,1 ,x now,2 ,...,x now,i ,...,x now,N } and the static feature vector X of multiple historical test samples his ={X his,1 ,X his,2 ,...,X his,j ,...,X his,M }, where x now,i Represents X now The i-th static feature in X his,j represents the static feature vector of the jth historical test sample, and X his,j ={x his_j,1 ,x his_j,2 ,...,x his_j,i ,...,x his_j,N }, x his_j,i Represents X his,j The i-th static feature in , M represents the total number of historical test samples, and N represents the total number of current static features; Step 1.2: Obtain the dynamic feature vector Y of the current test sample during infrared tracking now ={y now,1 ,y now,2 ,...,y now,h ,...,y now,K } and the dynamic feature vector Y of multiple historical test samples his ={Y his,1 ,Y his,2 ,...,Y his,n ,...,Y his,M }, where y now,h Represents Y now The h-th dynamic feature in Y his,n represents the dynamic feature vector of the nth historical test sample, and Y his,n ={y his_n,1 ,y his_n,2 ,...,y his_n,h ,...,y his_n,K },y his_n,h Represents Y his,n The hth dynamic feature in , K represents the total number of dynamic features.

3. The similarity estimation method based on multiple similarity mechanisms according to claim 1, characterized in that: Step 2 is as follows: Step 2.1: Establish a measurement network based on static feature consistency, and input the current test sample and each historical test sample into the measurement network in turn for processing to obtain the static feature vector F of the current test sample. now ={f now,1 ,f now,2 ,...,f now,i ,...,f now,N } and the static feature vector F of multiple historical test samples his ={F his,1 ,F his,2 ,...,F his,j ,...,F his,M }, where f now,i Represents vector F now The i-th static feature, F his,j Represents vector F his The static feature vector of the jth historical test sample in F his,j ={f his_j,1 ,f his_j,2 ,...,f his_j,p ,...,f his_j,N }, f his_j,p Indicates F his,j The p-th static feature in ; Step 2.2, based on F now 、X now Construct a contrast loss function for the static vector of the current test sample based on F his 、X his Constructing a contrastive loss function for static vectors of historical test samples, thereby training the metric network using a gradient descent method, and calculating the contrastive loss function to update network parameters until the contrastive loss function converges, thereby obtaining a trained metric network model; Step 2.3: Calculate F using the cosine function now and F his,j Static feature similarity sim j (F now ,F his,j ), thus obtaining F now The static feature similarity of the static feature vectors of the M historical test samples is obtained, and according to the static similarity threshold λ, the historical test samples with high static similarity to the current test sample are screened to obtain the static feature vectors G of the screened Z historical test samples. his ={G his,1 ,G his,2 ,...,G his,q ,...,G his,Z }, where G his,q represents the static feature vector of the qth historical test sample after screening, and G his,q ={G his_q,1 ,G his_q,2 ,...,G his_q,r ,...,G his_q,N }, G his_q,r Represents G his,q The rth static feature of , Z represents the total number of historical test samples after screening, Z≤M; Step 2.4: Based on the static feature vector G of the Z historical test samples after screening his ={G his,1 ,G his,2 ,...,G his,q ,...,G his,Z }, and obtain the dynamic feature vector E of the Z historical test samples after screening his ={E his,1 ,E his,2 ,...,E his,q ,...,E his,Z }, where E his,q Represents the dynamic feature vector of the qth historical test sample after screening.

4. The similarity estimation method based on multiple similarity mechanisms according to claim 3, characterized in that: In step 2.1, the measurement network based on static feature consistency includes: two branches; The first branch includes: the first embedding layer, a one-dimensional convolution layer with a stride of s and a convolution kernel of k×1, the first BN normalization layer, the first ReLu nonlinear activation layer, a maximum pooling layer with a convolution kernel of k×1, and the first fully connected layer. The second branch sequentially includes: the second embedding layer, a one-dimensional convolution layer with a stride of s and a convolution kernel of k×1, the second BN normalization layer, the second ReLu nonlinear activation layer, a maximum pooling layer with a convolution kernel of k×1, and the second fully connected layer.

5. The similarity estimation method based on multiple similarity mechanisms according to claim 1, characterized in that: In step 3, the dynamic characteristics of the filtered historical test samples and the current test samples are obtained as follows: Construct a dynamic feature similarity calculation model and perform now ={Y now,1 ,Y now,2 ,...,Y now,h ,...,Y now,K } and E his ={E his,1 ,E his,2 ,...,E his,q ,...,E his,Z } to obtain the dynamic feature vector TK of the current test sample now ={TK now,1 ,TK now,2 ,...,TK now,h ,...,TK now,K } and the dynamic feature vector TK of the Z historical test samples after screening his ={TK his,1 ,TK his,2 ,...,TK his,q ,...,TK his,Z }, where TK his,q Indicates TK his The qth dynamic eigenvector of ; TK now,h Indicates TK now The hth dynamic feature in .

6. The similarity estimation method based on multiple similarity mechanisms according to claim 5, characterized in that: The dynamic feature similarity calculation model in step 3 includes a pre-trained word segmenter, a pre-trained language representation model, a fully connected layer, and an activation function; The dynamic feature vector Y of the current test sample now ={Y now,1 ,Y now,2 ,...,Y now,h ,...,Y now,K } is input into the dynamic feature similarity calculation model, and the dynamic feature vector TK of the current test sample is obtained using formula (1): now ={TK now,1 ,TK now,2 ,...,TK now,h ,...,TK now,K }, where TK now,h Indicates TK now The h-th dynamic feature in ; BP now =FC(RELU(FC(BERT(Tokenizer(Y now ))))) (1) In formula (1), FC represents the fully connected layer, RELU represents the nonlinear activation function, BERT represents the pre-trained language representation model, and Tokenizer represents the word segmenter; The dynamic feature vector E of the Z historical test samples obtained after screening his ={E his,1 ,E his,2 ,...,E his,q ,...,E his,Z } is input into the dynamic feature similarity calculation model, and the dynamic feature vector TK of the filtered Z historical test samples is obtained using formula (2): his : TK his ={TK his,1 ,TK his,2 ,...,TK his,q ,...,TK his,Z }, Among them, TK his,q Indicates TK his The qth dynamic eigenvector of ; BP his,q =FC(RELU(FC(BERT(Tokenizer(E his,q ))))) (2)。

Citation Information

Cited By

  • Pest and disease monitoring, analyzing and processing system based on agriculture and forestry economic management

    CN119625542A