Landscape painting authenticity identification method based on multi-scale feature fusion
The multi-scale feature fusion method enhances landscape painting authentication by capturing detailed features, addressing inefficiencies in existing subjective methods, achieving improved accuracy and efficiency in identifying authentic versus fake paintings.
Patent Information
- Application Number
- CN202411260363.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-10
- Publication Date
- 2025-01-03
AI Technical Summary
Existing methods for authenticating landscape paintings rely heavily on subjective human expertise and are inefficient and inaccurate, especially for complex paintings, leading to significant financial losses and emotional distress due to the prevalence of counterfeit works.
A landscape painting authenticity identification method based on multi-scale feature fusion, utilizing a network composed of a feature extraction module, feature fusion module, and authenticity identification module, trained with a specific objective function to enhance feature capture and identification accuracy.
The method achieves high identification efficiency and accuracy by capturing detailed features like brushstrokes and textures, improving robustness and generalization, outperforming existing technologies in authentication tasks.
Smart Images
Figure CN119251656A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of computer vision technology, and in particular relates to a method for identifying the authenticity of landscape paintings. Technical Background
[0002] As a key form of traditional Chinese painting, landscape painting boasts a long history and profound cultural connotations. Chinese landscape painting can be seen as a perfect fusion of ink and line art, focusing on rhythm, brushwork, and composition. Composition is primarily based on scattered perspective, with a particular emphasis on brushwork. Technically, it emphasizes the use of brushstrokes in a variety of ways, including lightness and heaviness, firmness and softness, squareness and roundness, and thickness. The textures of natural elements such as rocks, water, and trees in paintings are often more detailed and complex. Artists employ techniques such as dotting, outlining, and splashing ink to depict details such as the cracks in rocks, the ripples of water, and the grain of trees, lending the painting a vivid and textured quality.
[0003] As the landscape painting market prospers and its value continues to rise, counterfeit and substandard landscape paintings are becoming increasingly prevalent. Currently, a large number of forged landscape paintings exist on the market, causing significant financial losses and emotional distress to collectors and connoisseurs. Therefore, developing an accurate and reliable method for authenticating landscape paintings is of great significance.
[0004] Currently, methods for authenticating landscape paintings rely primarily on the experience and intuition of art connoisseurs, or through traditional manual observation and testing. These methods are subject to significant subjective influences and are easily misled by sophisticated forgers. Furthermore, traditional manual inspection methods are often time-consuming and labor-intensive, and struggle to provide comprehensive and accurate authentication of complex landscape paintings.
[0005] Based on the above situation, in the field of landscape painting style migration technology, the technical problem that needs to be urgently solved is to develop an objective and accurate identification method. By combining advanced technical means and scientific methods, the ability to identify the authenticity of landscape paintings can be improved, providing more reliable protection for the market and collection fields. Summary of the Invention
[0006] The technical problem to be solved by the present invention is to overcome the shortcomings of the above-mentioned existing technical problems and provide a landscape painting authenticity identification method based on multi-scale feature fusion with diversified information, high identification efficiency and fast identification speed.
[0007] The technical solution adopted to solve the above technical problems consists of the following steps:
[0008] (1) Establishing a dataset of landscape painting authenticity
[0009] The authentic and imitation landscape painting images are selected as the dataset, and the dataset is divided into training set and test set in the ratio of 7:3.
[0010] (2) Dataset preprocessing
[0011] Remove large areas of blank space in the landscape painting image, retain the main content of the landscape painting image, adjust it to 512×512 pixel image, save it in jpg format, and manually mark the authenticity of the landscape painting image.
[0012] (3) Constructing a landscape painting authenticity identification network
[0013] The landscape painting authenticity identification network is composed of a feature extraction module, a feature fusion module, and an authenticity identification module connected in series.
[0014] The feature extraction module is composed of a preprocessing module, a residual block group 1, a residual block group 2, and an attention module 1. The preprocessing module is connected in series with the residual block group 1, and the output end of the residual block group 1 is respectively connected to the input end of the residual block group 2 and the attention module 1. The attention module 1 is a CBAM attention mechanism; the preprocessing module is composed of a convolutional layer 1 and a pooling layer 1 connected in series in sequence; the residual block group 1 is composed of residual block 1-1 and residual block 2-1, residual block 2-2, residual block 1-2, residual block 2-3, residual block 2-4, and residual block 2-5 connected in series in sequence; the residual block group 2 is composed of residual block 1-3 and residual block 2-6, residual block 2-7, residual block 2-8, residual block 2-9, and residual block 2-10 connected in series in sequence.
[0015] The feature fusion module is composed of an upsampling layer and a splicing layer 1 connected in series.
[0016] The authenticity identification module is composed of an attention module 2, a pooling layer 2, an expansion layer, a fully connected layer, and a Softmax layer 1 connected in series.
[0017] (4) Training a landscape painting authenticity identification network
[0018] 1) Determine the objective function
[0019] The objective function includes the loss function L, the evaluation function ACC, and the evaluation function F1. The loss function L is determined by the following formula:
[0020] L=-(ylogp i +(1-y)log(1-p i ))
[0021]
[0022] Among them, i and c represent the authenticity category of the sample, that is, genuine and non-genuine, y represents the target class or non-target class, the target class y value is 0, the non-target class y value is 1, p i is the model’s predicted probability for the ith true or false category, zi Represents the model's prediction score for the i-th true or false category, z c Represents the model's prediction score for the c-th category.
[0023] Determine the evaluation function ACC as follows:
[0024]
[0025] Among them, TP represents the number of correct positive example predictions, FP represents the number of incorrect negative example predictions, TN represents the number of correct negative example predictions, and FN represents the number of incorrect positive example predictions. The sum of TP, FP, TN, and FN is the total number of samples. The sum of TP and TN is the number of correctly predicted samples. TP, FP, TN, and FN∈N, where N is a finite positive integer, and TP, FP, TN, and FN are not all 0 at the same time.
[0026] Determine the evaluation function F1 as follows:
[0027]
[0028] Where P is the precision rate, R is the recall rate, and TP, FP, TN, and FN are not all 0 at the same time.
[0029] 2) Training a landscape painting authenticity identification network
[0030] The training set is input into the authenticity identification network for training. The initial learning rate is set to 0.001, the number of training rounds is 100, the batch size of each training round is 8, and the iteration is carried out until the loss function converges. During the training of the authenticity identification network, the weights are continuously updated and the weight files are saved.
[0031] (5) Testing the Landscape Painting Authenticity Authentication Network
[0032] Input the test set into the trained landscape painting authenticity identification network for testing, load the saved weight file, and obtain the landscape painting authenticity identification results.
[0033] In step (3) of the present invention, a landscape painting authenticity identification network is constructed, wherein the residual block 1-1 is composed of a convolutional layer 2, a RELU activation function layer 1, a convolutional layer 3, a RELU activation function layer 2, a convolutional layer 4, a RELU activation function layer 3, and a convolutional layer 5, wherein the convolutional layer 2 is connected in series with the RELU activation function layer 1, the convolutional layer 3, the RELU activation function layer 2, the convolutional layer 4, and the RELU activation function layer 3, and the input end of the convolutional layer 5 is connected with the input end of the convolutional layer 2, and the output end is connected with the input end of the RELU activation function layer 3; the convolution kernel size of the convolutional layer 2 and the convolutional layer 5 is 1×1 with a step length of 2; the convolution kernel size of the convolutional layer 4 is 1×1 with a step length of 1; and the convolution kernel size of the convolutional layer 3 is 3×3 with a step length of 1.
[0034] The structures of the residual block 1-2 and the residual block 1-3 are the same as that of the residual block 1-1.
[0035] In step (3) of the present invention, the landscape painting authenticity identification network is constructed. The residual block 2-1 is composed of a convolution layer 6, a RELU activation function layer 4, a convolution layer 7, a RELU activation function layer 5, a convolution layer 8, and a RELU activation function layer 6. The convolution layer 6 is connected in series with the RELU activation function layer 4, the convolution layer 7, the RELU activation function layer 5, the convolution layer 8, and the RELU activation function layer 6. The input end of the convolution layer 6 is connected to the other input end of the RELU activation function layer 6. The convolution kernel size of the convolution layer 6 and the convolution layer 8 is 1×1 with a step size of 1. The convolution kernel size of the convolution layer 7 is 3×3 with a step size of 1. The structures of the residual blocks 2-2, 2-3, 2-4, 2-5, 2-6, 2-7, 2-8, 2-9, and 2-10 are the same as that of the residual block 2-1.
[0036] In step (3) of the present invention, the landscape painting authenticity identification network is constructed, and the attention module 2 is composed of a position sensing module, an attention aggregation module, an interaction layer 1, an interaction layer 2, a weight adjustment layer, and a feature enhancement module; the position sensing module is connected in series with the attention aggregation module, the interaction layer 1, the interaction layer 2, and the weight adjustment layer, and one output end of the feature enhancement module is connected to the interaction layer 1, and the other output end is connected to the interaction layer 2.
[0037] The position perception module of the present invention is connected to the splicing layer 2 by the output ends of the pooling layer 3 and the pooling layer 4 respectively.
[0038] The attention aggregation module of the present invention is composed of a convolution layer 9, a normalization layer, a pooling layer 5, and a Softmax layer 2 connected in series in sequence. The convolution kernel size of the convolution layer 9 is 1×1 and the step size is 1.
[0039] The feature enhancement module of the present invention consists of a dilated convolutional layer connected in series with pooling layer 6, convolutional layer 10, convolutional layer 11, and softmax layer 3. One output of the dilated convolutional layer is connected to the other input of interaction layer 1, and the output of softmax layer 3 is connected to the other input of interaction layer 2. The convolution kernel size of convolutional layers 10 and 11 is 1×1 with a stride of 1. The convolution kernel size of the dilated convolutional layer is 3×3 with a stride of 1 and an interval of 1.
[0040] Because the landscape painting authenticity identification network of the present invention adopts a feature extraction module, it solves the technical problem that the existing technology of landscape painting image feature extraction is insufficient in richness, and cannot fully capture the detailed information in the painting when processing complex landscape paintings, and can fully capture dimensional features; adopts a feature fusion module, which solves the technical problem of isolated feature information processing in the existing technology, and improves the robustness and generalization ability of the identification network; adopts an authenticity identification module, which solves the technical problem that the existing technology does not focus enough on detailed features such as brushstrokes, ink color changes and textures in landscape paintings, resulting in poor identification results, and improves the identification accuracy and efficiency. Compared with the existing technology, the present invention has the advantages of high feature extraction efficiency and high identification accuracy, and can be used for landscape painting authenticity identification. BRIEF DESCRIPTION OF THE DRAWINGS
[0041] Figure 1 This is a flowchart of Example 1 of the present invention.
[0042] Figure 2 It is a structural diagram of the landscape painting authenticity identification network.
[0043] Figure 3 yes Figure 2 Schematic diagram of the feature extraction module.
[0044] Figure 4 yes Figure 3 Schematic diagram of the structure of the residual block 1-1.
[0045] Figure 5 yes Figure 3 Schematic diagram of the structure of the residual block 2-1.
[0046] Figure 6 yes Figure 2 Schematic diagram of the structure of the feature fusion module.
[0047] Figure 7 yes Figure 2 Schematic diagram of the structure of the authenticity identification module.
[0048] Figure 8 yes Figure 7 Schematic diagram of the structure of attention module 2.
[0049] Figure 9 This is a computer simulation experiment effect diagram of the method in Example 1. DETAILED DESCRIPTION
[0050] The present invention will be further described below with reference to the accompanying drawings and examples, but the present invention is not limited to the following embodiments.
[0051] Example 1
[0052] exist Figure 1In this embodiment, the landscape painting authenticity identification method based on multi-scale feature fusion consists of the following steps:
[0053] (1) Establishing a dataset of landscape painting authenticity
[0054] The authentic and imitation landscape painting images are selected as the dataset, and the dataset is divided into training set and test set in the ratio of 7:3.
[0055] (2) Dataset preprocessing
[0056] Remove large areas of blank space in the landscape painting image, retain the main content of the landscape painting image, adjust it to 512×512 pixel image, save it in jpg format, and manually mark the authenticity of the landscape painting image.
[0057] (3) Constructing a landscape painting authenticity identification network
[0058] The following is a schematic diagram of the network structure for identifying the authenticity of landscape paintings in this embodiment. Figure 2 In the embodiment, the landscape painting authenticity identification network is composed of a feature extraction module, a feature fusion module, and an authenticity identification module connected in series.
[0059] Figure 3 Given Figure 2 Schematic diagram of the structure of the feature extraction module. Figure 3 In the embodiment, the feature extraction module is composed of a preprocessing module, a residual block group 1, a residual block group 2, and an attention module 1. The preprocessing module is connected in series with the residual block group 1. The output of the residual block group 1 is connected to the input of the residual block group 2 and the attention module 1 respectively. The attention module 1 is a CBAM attention mechanism. The preprocessing module of this embodiment is composed of a convolutional layer 1 and a pooling layer 1 connected in series. The residual block group 1 of this embodiment is composed of residual block 1-1 and residual block 2-1, residual block 2-2, residual block 1-2, residual block 2-3, residual block 2-4, and residual block 2-5 connected in series. The residual block group 2 of this embodiment is composed of residual block 1-3 and residual block 2-6, residual block 2-7, residual block 2-8, residual block 2-9, and residual block 2-10 connected in series.
[0060] Since the landscape painting authenticity identification network of the present invention adopts a feature extraction module, it solves the technical problem that the landscape painting image feature extraction in the prior art is insufficiently rich and cannot fully capture the detailed information in the painting when processing complex landscape paintings, and can fully capture multi-dimensional features.
[0061] Figure 4 Given Figure 3 Schematic diagram of the structure of residual block 1-1. Figure 4In the embodiment, the residual block 1-1 is composed of a convolution layer 2, a RELU activation function layer 1, a convolution layer 3, a RELU activation function layer 2, a convolution layer 4, a RELU activation function layer 3, and a convolution layer 5. The convolution layer 2 is connected in series with the RELU activation function layer 1, the convolution layer 3, the RELU activation function layer 2, the convolution layer 4, and the RELU activation function layer 3. The input end of the convolution layer 5 is connected to the input end of the convolution layer 2, and the output end is connected to the input end of the RELU activation function layer 3; the convolution kernel size of the convolution layer 2 and the convolution layer 5 is 1×1 with a step size of 2; the convolution kernel size of the convolution layer 4 is 1×1 with a step size of 1; the convolution kernel size of the convolution layer 3 is 3×3 with a step size of 1.
[0062] The structures of residual block 1-2 and residual block 1-3 in this embodiment are the same as that of residual block 1-1.
[0063] Figure 5 Given Figure 3 Schematic diagram of the structure of residual block 2-1. Figure 5 In the embodiment, the residual block 2-1 is composed of a convolution layer 6, a RELU activation function layer 4, a convolution layer 7, a RELU activation function layer 5, a convolution layer 8, and a RELU activation function layer 6. The convolution layer 6 is connected in series with the RELU activation function layer 4, the convolution layer 7, the RELU activation function layer 5, the convolution layer 8, and the RELU activation function layer 6. The input end of the convolution layer 6 is connected to the other input end of the RELU activation function layer 6. The convolution kernel size of the convolution layer 6 and the convolution layer 8 is 1×1 with a step size of 1; the convolution kernel size of the convolution layer 7 is 3×3 with a step size of 1. The structures of the residual blocks 2-2, 2-3, 2-4, 2-5, 2-6, 2-7, 2-8, 2-9, and 2-10 of this embodiment are the same as the structure of the residual block 2-1.
[0064] Figure 6 Given Figure 2 Schematic diagram of the structure of the feature fusion module in . Figure 6 In the embodiment, the feature fusion module is composed of an upsampling layer and a splicing layer 1 connected in series.
[0065] Since the landscape painting authenticity identification network of the present invention adopts a feature fusion module, it solves the problem of isolated feature information processing technology in the prior art and improves the robustness and generalization ability of the identification network.
[0066] Figure 7 Given Figure 2 Schematic diagram of the structure of the authenticity identification module. Figure 7 The authenticity identification module of this embodiment is composed of an attention module 2, a pooling layer 2, an expansion layer, a fully connected layer, and a Softmax layer 1 connected in series.
[0067] The landscape painting authenticity identification network of the present invention adopts an authenticity identification module, which solves the technical problem that the existing technology does not focus enough on detailed features such as brushstrokes, ink color changes and textures in landscape paintings, resulting in poor identification results, and improves the accuracy and efficiency of identification.
[0068] Figure 8 The structural diagram of attention module 2 is given. Figure 8 In the embodiment, the attention module 2 is composed of a position perception module, an attention aggregation module, an interaction layer 1, an interaction layer 2, a weight adjustment layer, and a feature enhancement module; the position perception module is connected in series with the attention aggregation module, the interaction layer 1, the interaction layer 2, and the weight adjustment layer, and one output end of the feature enhancement module is connected to the interaction layer 1, and the other output end is connected to the interaction layer 2.
[0069] In the position perception module of this embodiment, the output ends of the pooling layer 3 and the pooling layer 4 are connected to the splicing layer 2 respectively.
[0070] The attention aggregation module of this embodiment is composed of a convolution layer 9, a normalization layer, a pooling layer 5, and a Softmax layer 2 connected in series in sequence. The convolution kernel size of the convolution layer 9 is 1×1 and the step size is 1.
[0071] The feature enhancement module of this embodiment is composed of a dilated convolution layer and a pooling layer 6, a convolution layer 10, a convolution layer 11, and a Softmax layer 3 connected in series in sequence. One output end of the dilated convolution layer is connected to the other input end of the interaction layer 1, and the output end of the Softmax layer 3 is connected to the other input end of the interaction layer 2; the convolution kernel size of the convolution layer 10 and the convolution layer 11 is 1×1 with a step size of 1; the convolution kernel size of the dilated convolution layer is 3×3 with a step size of 1 and an interval of 1.
[0072] (4) Training a landscape painting authenticity identification network
[0073] 1) Determine the objective function
[0074] The objective function includes the loss function L, the evaluation function ACC, and the evaluation function F1. The loss function L is determined by the following formula:
[0075] L=-(ylogp i +(1-y)log(1-p i ))
[0076]
[0077] Among them, i and c represent the authenticity category of the sample, that is, genuine and non-genuine, y represents the target class or non-target class, the target class y value is 0, the non-target class y value is 1, p i is the model’s predicted probability for the ith true or false category, z i Represents the model's prediction score for the i-th true or false category, zc Represents the model's prediction score for the cth category;
[0078] Determine the evaluation function ACC as follows:
[0079]
[0080] Wherein, TP represents the number of correct positive example predictions, FP represents the number of incorrect negative example predictions, TN represents the number of correct negative example predictions, and FN represents the number of incorrect positive example predictions. The sum of TP, FP, TN, and FN is the total number of samples. The sum of TP and TN is the number of correctly predicted samples. TP, FP, TN, and FN∈N, where N is a finite positive integer. In this embodiment, N is 800, and TP, FP, TN, and FN are not all 0.
[0081] Determine the evaluation function F1 as follows:
[0082]
[0083] Where P is the precision rate, R is the recall rate, and TP, FP, TN, and FN are not all 0 at the same time;
[0084] 2) Training a landscape painting authenticity identification network
[0085] The training set is input into the authenticity identification network for training. The initial learning rate is set to 0.001, the number of training rounds is 100, the batch size of each round of training is 8, and it is iterated until the loss function converges. During the training process of the authenticity identification network, the weights are continuously updated and the weight files are saved.
[0086] (5) Testing the Landscape Painting Authenticity Authentication Network
[0087] Input the test set into the trained landscape painting authenticity identification network for testing, load the saved weight file, and obtain the landscape painting authenticity identification results.
[0088] In order to verify the beneficial effects of the present invention, a computer simulation experiment was carried out using the method of Example 1 of the present invention. The experimental results are shown in FIG. Figure 9 .exist Figure 9 In the picture, a is the identification image of the authentic landscape painting, b is the identification image of the fake landscape painting, c is the identification image of the authentic landscape painting, and d is the identification image of the fake landscape painting. Figure 9 It can be seen that the present invention can accurately identify the authenticity of landscape paintings.
[0089] The same test set was tested with each saved model, and the model accuracy was tested using the evaluation code. The method of Example 1 of the present invention was used to conduct comparative experiments with the deep residual network (hereinafter referred to as Resnet-50) and the multi-task feature fusion network (hereinafter referred to as MTFFNet). The evaluation function ACC and the evaluation function F1 were used as indicators for the evaluation method. The larger the values of the evaluation function ACC and the evaluation function F1, the better the method. The experimental results of the evaluation functions ACC and F1 are shown in Table 1.
[0090] Table 1 Evaluation results of evaluation functions ACC and F1 of Example 1 and existing methods
[0091] Experimental methods Evaluation function ACC (%) Evaluation function F1 (%) Resnet-50 85.4 86.6 MTFFNet 87.8 89.3 Example 1 94.7 93.8
[0092] As can be seen from Table 1, the evaluation function ACC value of the method in Example 1 is 9.3% higher than the evaluation function ACC value of the Resnet-50 method and 6.9% higher than the evaluation function ACC value of the MTFFNet method; the evaluation function F1 value of the method in Example 1 is 7.2% higher than the evaluation function F1 value of the Resnet-50 method and 4.5% higher than the evaluation function F1 value of the MTFFNet method.
Claims
1. A method for identifying the authenticity of landscape paintings based on multi-scale feature fusion, characterized in that It consists of the following steps: (1) Establishing a dataset of landscape painting authenticity The authentic and imitation landscape painting images are selected as the data set, and the data set is divided into training set and test set in the ratio of 7:3; (2) Dataset preprocessing Remove the large blank areas in the landscape painting image, retain the main content of the landscape painting image, adjust it to 512×512 pixel image, save it in jpg format and manually mark the landscape painting image for authenticity; (3) Constructing a landscape painting authenticity identification network The landscape painting authenticity identification network is composed of a feature extraction module, a feature fusion module, and an authenticity identification module in series. The feature extraction module is composed of a preprocessing module, a residual block group 1, a residual block group 2, and an attention module 1. The preprocessing module is connected in series with the residual block group 1, and the output end of the residual block group 1 is connected to the input end of the residual block group 2 and the attention module 1 respectively. The attention module 1 is a CBAM attention mechanism; the preprocessing module is composed of a convolution layer 1 and a pooling layer 1 connected in series in sequence; the residual block group 1 is composed of a residual block 1-1 and a residual block 2-1, a residual block 2-2, a residual block 1-2, a residual block 2-3, a residual block 2-4, and a residual block 2-5 connected in series in sequence; the residual block group 2 is composed of a residual block 1-3 and a residual block 2-6, a residual block 2-7, a residual block 2-8, a residual block 2-9, and a residual block 2-10 connected in series in sequence; The feature fusion module is composed of an upsampling layer and a splicing layer 1 in series; The authenticity identification module is composed of an attention module 2, a pooling layer 2, an expansion layer, a fully connected layer, and a Softmax layer 1 connected in series in sequence; (4) Training a landscape painting authenticity identification network 1) Determine the objective function The objective function includes the loss function L, the evaluation function ACC, and the evaluation function F1. The loss function L is determined as follows: L=-(ylogp i +(1-y)log(1-p i )) Among them, i and c represent the authenticity category of the sample, that is, authentic and non-authentic, y represents the target class or non-target class, the target class y value is 0, and the non-target class y value is 1, p i is the model’s predicted probability for the ith true or false category, z i represents the model's prediction score for the ith true or false category, z c Represents the model's prediction score for the cth category; Determine the evaluation function ACC as follows: Among them, TP represents the number of correct positive example predictions, FP represents the number of wrong negative example predictions, TN represents the number of correct negative example predictions, FN represents the number of wrong positive example predictions, the sum of TP and FP, TN, and FN is the total number of samples, the sum of TP and TN is the number of correctly predicted samples, TP, FP, TN, FN∈N, N is a finite positive integer, and TP, FP, TN, and FN are not 0 at the same time; Determine the evaluation function F1 as follows: Among them, P is the precision rate, R is the recall rate, and TP, FP, TN, and FN are not 0 at the same time; 2) Training a landscape painting authenticity identification network The training set is input into the authenticity identification network for training. The initial learning rate is set to 0.001, the number of training rounds is 100, the batch size of each round of training is 8, and it is iterated until the loss function converges. During the training of the authenticity identification network, the weights are continuously updated and the weight files are saved. (5) Testing the landscape painting authenticity identification network The test set is input into the trained landscape painting authenticity identification network for testing, and the saved weight file is loaded to obtain the landscape painting authenticity identification result.
2. The method for identifying the authenticity of landscape paintings based on multi-scale feature fusion according to claim 1 is characterized in that: In step (3) of constructing a landscape painting authenticity identification network, the residual block 1-1 is composed of a convolutional layer 2, a RELU activation function layer 1, a convolutional layer 3, a RELU activation function layer 2, a convolutional layer 4, a RELU activation function layer 3, and a convolutional layer 5, wherein the convolutional layer 2 is connected in series with the RELU activation function layer 1, the convolutional layer 3, the RELU activation function layer 2, the convolutional layer 4, and the RELU activation function layer 3, respectively; the input end of the convolutional layer 5 is connected with the input end of the convolutional layer 2, and the output end is connected with the input end of the RELU activation function layer 3; the convolutional kernel size of the convolutional layer 2 and the convolutional layer 5 is 1×1 with a step length of 2; the convolutional kernel size of the convolutional layer 4 is 1×1 with a step length of 1; the convolutional kernel size of the convolutional layer 3 is 3×3 with a step length of 1; The structures of the residual block 1-2 and the residual block 1-3 are the same as the structure of the residual block 1-1.
3. The method for identifying the authenticity of landscape paintings based on multi-scale feature fusion according to claim 1, characterized in that: In step (3) of constructing a landscape painting authenticity identification network, the residual block 2-1 is composed of a convolution layer 6, a RELU activation function layer 4, a convolution layer 7, a RELU activation function layer 5, a convolution layer 8, and a RELU activation function layer 6, wherein the convolution layer 6 is connected in series with the RELU activation function layer 4, the convolution layer 7, the RELU activation function layer 5, the convolution layer 8, and the RELU activation function layer 6, and the input end of the convolution layer 6 is connected with the other input end of the RELU activation function layer 6; the convolution kernel size of the convolution layer 6 and the convolution layer 8 is 1×1 with a step size of 1; the convolution kernel size of the convolution layer 7 is 3×3 with a step size of 1. The structures of the residual block 2-2, the residual block 2-3, the residual block 2-4, the residual block 2-5, the residual block 2-6, the residual block 2-7, the residual block 2-8, the residual block 2-9, and the residual block 2-10 are the same as the structure of the residual block 2-1.
4. The method for identifying the authenticity of landscape paintings based on multi-scale feature fusion according to claim 1, characterized in that: In step (3) of constructing a landscape painting authenticity identification network, the attention module 2 is composed of a position perception module, an attention aggregation module, an interaction layer 1, an interaction layer 2, a weight adjustment layer, and a feature enhancement module; the position perception module is connected in series with the attention aggregation module, the interaction layer 1, the interaction layer 2, and the weight adjustment layer, and one output end of the feature enhancement module is connected to the interaction layer 1, and the other output end is connected to the interaction layer 2.
5. The method for identifying the authenticity of landscape paintings based on multi-scale feature fusion according to claim 4 is characterized in that: The position perception module is connected to the splicing layer 2 via the output ends of the pooling layer 3 and the pooling layer 4 respectively.
6. The method for identifying the authenticity of landscape paintings based on multi-scale feature fusion according to claim 4, characterized in that: The attention aggregation module is composed of a convolution layer 9, a normalization layer, a pooling layer 5, and a Softmax layer 2 connected in series in sequence. The convolution kernel size of the convolution layer 9 is 1×1 and the step size is 1.
7. The method for identifying the authenticity of landscape paintings based on multi-scale feature fusion according to claim 4, characterized in that: The feature enhancement module is composed of a hole convolution layer and a pooling layer 6, a convolution layer 10, a convolution layer 11, and a Softmax layer 3 connected in series in sequence, one output end of the hole convolution layer is connected to the other input end of the interaction layer 1, and the output end of the Softmax layer 3 is connected to the other input end of the interaction layer 2; the convolution kernel size of the convolution layer 10 and the convolution layer 11 is 1×1 with a step size of 1; the convolution kernel size of the hole convolution layer is 3×3 with a step size of 1 and an interval of 1.