Similar trademark change detection method based on selective state space model

Abstract

The invention discloses a similar trademark change detection method based on a selective state space model. The method comprises the following steps: training a selective state space model by using a detection data set to obtain a similar trademark change detection model; the selective state space model comprises an encoder, a connector, a decoder, a segmented image expansion module, a linear mapping calculation module and a convolutional neural network; wherein the encoder adopts a twin network architecture, and each branch in the architecture comprises a first encoder sub-block, a second encoder sub-block, a third encoder sub-block and a segmented image merging module which are used for processing input data in sequence; the connector comprises two visual state space calculation modules which are connected in series; the decoder comprises three decoder sub-blocks and a segmented image expansion module; and inputting a target trademark image and a similar trademark image into the detection model to obtain a change detection result. According to the method, the application capability of the model to long-range features is enhanced while the total quantity of model parameters is reduced, and the prediction precision of the model is greatly improved.

Description

Technical Field

[0001] The invention belongs to the technical field of computer software and relates to a similar trademark change detection method based on a selective state space model. Background Art

[0002] A trademark is a mark used to identify and distinguish the source of goods or services, and is an important intellectual property right of an individual or organization. According to Article 8 of my country's Trademark Law, "Any mark that can distinguish the goods of a natural person, legal person or other organization from the goods of others, including words, graphics, letters, numbers, three-dimensional marks, color combinations and sounds, etc., as well as the combination of the above elements, can be applied for registration as a trademark." Article 57, paragraph 2, stipulates that "using a trademark similar to its registered trademark on the same kind of goods without the permission of the trademark registrant, or using a trademark identical or similar to its registered trademark on similar goods, which is likely to cause confusion" is an infringement of the exclusive right to a registered trademark.

[0003] At a time when innovation is driving rapid economic and social development, the protection of intellectual property rights is constantly being strengthened, but the misappropriation and theft of trademarks still occur frequently. Infringement no longer only includes the direct misappropriation and theft of registered trademarks belonging to other rights holders, but more often it is based on the patterns and graphic designs of registered trademarks, using methods such as changing colors, adding or deleting graphics to achieve the purpose of visual confusion with the original trademark image. This type of infringement is not only more covert, but also causes more serious commercial loss and brand damage to the infringed party. At the same time, computer vision intelligent models with deep learning technology as the core have been applied in various application scenarios, and in the problem of trademark similarity change detection, it has been able to detect changes in trademark images with slight changes.

[0004] In the existing technical solutions, Alshowaish et al. implemented a trademark similarity detection model based on shape similarity by using convolutional neural networks, and realized feature extraction of trademark images in different storage formats through convolutional neural networks, and found the similarity between them based on the distance between a given image and each stored image. Two different pre-trained convolutional neural networks, ResNet-50 and VGG-16, were used as feature extractors. At the same time, in order to reduce the impact of high-dimensional features on recognition performance, the PCA algorithm was used to reduce the dimensionality of the extracted features to improve the recognition accuracy of similar trademarks. This method uses a similarity evaluation model based on a pre-trained visual model and a convolutional neural network to form a similarity evaluation model between similar trademarks, but this method has a weak ability to capture tiny features in the image, and cannot find the difference between two similar trademarks. It can only give an evaluation result measured by similarity distance. At the same time, the use of a PCA-based dimensionality reduction algorithm leads to a large loss of image features, and fails to achieve a balance between data dimension and evaluation accuracy.

[0005] Liu et al. proposed a neural network model that uses the semantics, visual information, and pronunciation information of the characters appearing in the trademark as data features to evaluate the similarity between similar trademarks. In terms of visual feature detection, this technical solution uses different trademark feature vectors to construct a similarity matrix to calculate the maximum similarity evaluation vector and the relative maximum similarity position vector, and then uses these two vectors as input to calculate the probability distribution of similarity classification through convolutional neural networks. The focus of this method is to form a new feature vector calculation mode by constructing a similarity matrix, but this method can only provide classification results for image similarity and dissimilarity, and cannot effectively provide the ability to capture changes and differences between images. Summary of the invention

[0006] In view of the problems existing in the prior art, the purpose of the present invention is to provide a method for detecting changes in similar trademarks based on a selective state space model. In the present invention, starting from the provisions in the Trademark Law and the problem of detecting pattern changes of similar trademarks in judicial practice, a technical method based on a selective state space model is proposed to focus on the problem of similarity determination of trademark patterns, so as to improve the ability to detect differences between images; the present invention is implemented by a machine learning algorithm based on a selective state space model, which has a stronger feature capture capability than existing convolutional neural networks and neural networks based on attention mechanisms.

[0007] The main purpose of this invention is to provide effective solutions and means for comparing changes between similar trademarks by applying computer vision and deep learning technology, improve the application capabilities of computers in this field, reduce the proportion of manual work in trademark applications, trademark infringement law enforcement, and intellectual property case trials, and thus assist in the protection of intellectual property rights.

[0008] This method uses a selective state space model to design and propose a trademark similarity change detection method, which can predict the center position, length and width of the difference between two similar trademarks. First, by adopting a U-shaped twin network model, the image features of two similar trademarks are fully utilized to provide a model architecture foundation for achieving differentiated prediction capabilities; and by using a processing module based on a selective state space model as the core to form an encoder and a decoder, it has a good effect on reducing the total amount of model parameters while strengthening the model's ability to apply long-range features, which can effectively improve the model's prediction accuracy.

[0009] The technical solution of the present invention is:

[0010] A method for detecting changes in similar trademarks based on a selective state space model comprises the following steps:

[0011] 1) obtaining a trademark similarity change detection dataset, wherein the trademark similarity change detection dataset includes multiple pairs of sample images, each pair of sample images includes a trademark image and its similar image;

[0012] 2) Using the trademark similarity change detection data set to train a selective state space model to obtain a similar trademark change detection model; the selective state space model includes an encoder, a connector, a decoder, a segmented image expansion module, a linear mapping calculation module and a convolutional neural network; wherein the encoder adopts a twin network architecture, including two branches with the same structure, each of which includes a first encoder sub-block, a second encoder sub-block, a third encoder sub-block and a segmented image merging module that sequentially processes input data; the connector includes two visual state space calculation modules connected in series; the decoder includes three decoder sub-blocks and a segmented image expansion module; one image in a pair of sample images is divided into multiple sub-images and input into one branch, and the other image in the pair of sample images is divided into multiple sub-images and input into another branch;

[0013] The encoder is used to calculate the input image from shallow features to deep features, and input the deep features to the connector and send the shallow features to the decoder; wherein the first encoder sub-block is used to extract the first shallow features from each input sub-image and input them to the second encoder sub-block and the third decoder sub-block respectively, the second encoder sub-block is used to extract the second shallow features from the first shallow features of each sub-image and input them to the third encoder sub-block and the second decoder sub-block respectively, the third encoder sub-block is used to extract the deep features from the second shallow features of each sub-image and input them to the segmentation image merging module and the first decoder sub-block respectively; the segmentation image merging module is used to merge the deep features of each sub-image to obtain the deep features of the input image and input them to the connector;

[0014] The connector is used to extract the input deep features and input the extracted features to the first decoder sub-block of the decoder; each decoder sub-block of the decoder decodes the input feature data to obtain image representations corresponding to different levels of each pair of sample images, and the image representations output by the third decoder sub-block are restored in turn by the segmentation image expansion module and the linear mapping calculation module, and then input into the convolutional neural network for prediction to obtain the change area corresponding to each pair of sample images;

[0015] Calculate the loss value of the predicted change area of ​​each pair of sample images and the annotation information of the pair of sample images according to the selected loss function, and optimize the selective state space model according to the loss value;

[0016] 3) Inputting a target trademark image and its similar trademark images into the similar trademark change detection model, and returning the change detection result of the similar trademark image relative to the target trademark.

[0017] Furthermore, the similar image corresponding to the trademark image in each pair of sample images is an image with the trademark image as the background and other icon data as the foreground.

[0018] Furthermore, the loss function L = L hm +λ wh L wh The loss value is calculated; where λ wh For custom hyperparameters, the focal loss function L hm Used to calculate the difference between the predicted change area center point and the labeled change area center point, the change area length and width loss function L wh Used to calculate the difference between the length and width of the predicted change area and the length and width of the marked change area.

[0019] Furthermore, the first encoder sub-block includes a linear embedded coding calculation module that processes the input data in sequence and two serially connected visual state space calculation modules, the linear embedded coding calculation module is used to perform linear calculations on each input sub-image to obtain its encoding vector and input it into the visual state space calculation module; the visual state space calculation module is used to learn the feature representation of the image; the corresponding visual state space calculation modules in the two encoders of the twin network share model parameters.

[0020] Furthermore, the second encoder sub-block has the same structure as the third encoder sub-block, both of which include a segmented image merging module that processes the input data in sequence and two serially connected visual state space calculation modules; wherein the segmented image merging module is used to complete the merging of segmented images and the expansion of feature channels.

[0021] Furthermore, the decoder sub-block includes a segmented image expansion module that processes the input data in sequence and two visual state space calculation modules connected in series; wherein the segmented image expansion module is used to upsample the input features, reconstruct the feature maps of adjacent dimensions into feature maps with higher resolution, and then upsample the feature maps reconstructed with higher resolution by a factor of 2, while reducing the dimension of the feature channels of the feature maps reconstructed with higher resolution to half.

[0022] Furthermore, the visual state space calculation module includes three calculation paths in total; in the first calculation path, the original input data is processed in sequence by the first layer normalization module, the first linear calculation module, the deep convolution module, the selective state space calculation module, and the second layer normalization module; in the second calculation path, the calculation result of the first layer normalization module is used as the input of the second linear calculation module, and the calculation result of the second linear calculation module and the calculation result of the second layer normalization module are respectively input into the matrix multiplication operation unit for matrix multiplication operation, and the result of the matrix multiplication operation is input into the third linear calculation module layer; in the third calculation path, the original input data and the calculation result of the third linear calculation module are subjected to matrix addition operation.

[0023] Furthermore, the trademark image and its similar image in each pair of sample images are H×W×1 images, wherein the first encoder sub-block is used to extract from each input sub-image The second encoder sub-block is used to extract the first shallow features of each sub-image from the first shallow features of each sub-image. The third encoder sub-block is used to extract from the second shallow features of each sub-image The segmented image merging module converts the dimension of the deep features output by the third encoder sub-block into First decoder sub-block output The image representation of the second decoder sub-block outputs The image representation of the third decoder sub-block outputs Image representation; where H is the vertical resolution value of the image, W is the horizontal resolution value of the image, and C is the number of feature channels of the image.

[0024] A server, characterized in that it comprises a memory and a processor, wherein the memory stores a computer program, the computer program is configured to be executed by the processor, and the computer program comprises instructions for executing the above method.

[0025] A computer-readable storage medium stores a computer program, wherein the computer program implements the above method when executed by a processor.

[0026] Compared with the prior art, the advantages of the present invention are as follows:

[0027] Traditional methods mostly use convolutional neural networks combined with classification algorithms to provide qualitative similarity comparison capabilities, but lack effective change detection and annotation capabilities; at the same time, related algorithms cannot efficiently capture long-range dependency features between images, resulting in low prediction accuracy and weak model application capabilities. In the present invention, by using a selective state space model as a basis and adopting a change detection method, the changes and differences between similar trademark images can be effectively captured, achieving high-precision change detection and difference annotation capabilities, and has strong application prospects in trademark image change detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 Flow chart of this method.

[0029] Figure 2 This is the overall structure diagram of the model.

[0030] Figure 3 This is the structure diagram of the visual state space calculation module. DETAILED DESCRIPTION

[0031] The present invention is further described in detail below in conjunction with the accompanying drawings. The examples given are only used to explain the present invention but not to limit the scope of the present invention.

[0032] The present invention provides a trademark similarity change detection method based on a selective state space model, which is used for differential detection between similar trademark images, and then identifying the image similarity features of the trademarks. By adopting a model architecture based on a selective state space model as a core module, the calculation and application capabilities for long-range features between trademark images are improved, the recognition accuracy of differential features is improved, and the time consumed in the prediction process is reduced. The technical solution in the example of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the example of the present invention. Figure 1 It is a flow chart of the technical method implemented by the present invention. Figure 2 It is a model structure diagram implemented by the present invention, Figure 3 This is the architecture diagram of the visual state space calculation module. This technical method includes the following steps:

[0033] 1. Process the trademark similarity change detection dataset and divide it into training set, validation set and test set.

[0034] 2. The encoder part of the trademark similarity change detection model is calculated to complete the calculation from shallow features to deep features for two similar trademark images;

[0035] 3. The connector part of the trademark similarity change detection model calculates and further obtains the deep features of the two images, and outputs the calculation results to the decoder part;

[0036] 4. The decoder part of the trademark similarity change detection model is calculated based on the deep feature data obtained in the previous step and the corresponding hierarchical feature data obtained by the jump connection to obtain the decoded feature data;

[0037] 5. Calculate the loss function of the trademark similarity change detection model and train the model to obtain the detection model formed by training;

[0038] 6. Input the trademark image to be detected and return the identification results of the changed parts in the similar trademark image.

[0039] Step 1 includes the following sub-steps:

[0040] Step 1.1: Based on the open source trademark dataset Logo-2K+, select 100 different trademark images as the trademark basic dataset.

[0041] Step 1.2: Use the open source logo dataset Icon645 as the variable icon dataset, select 200 icon data, use the OpenCV image fusion method to take the icon data as the foreground, and use the trademark image selected in step 1.1 as the background to form 20,000 similar trademark image data with certain differences. The 100 trademark images selected in step 1.1 and the corresponding 20,000 synthetic images are used to form 20,000 sample pairs as the dataset.

[0042] Step 1.3: Split the dataset into training set T according to the different purposes of training and testing train , validation set T val And the test set T test , the sample ratio of the three data sets is 6:2:2.

[0043] T cases =T train ∪T val ∪T test (1)

[0044] Step 2 includes the following sub-steps:

[0045] Step 2.1: Process the two images separately through the image segmentation processing module, and divide each image into 16 sub-images in equal proportion.

[0046] Step 2.2: The encoder uses a twin network architecture. The two images are input into the encoder with the same structure. The encoder is divided into three encoder sub-blocks with similar structures and a segmented image merging module.

[0047] In the first encoder sub-block, a linear embedding coding calculation module and two visual state space calculation modules are included, wherein the structure of the visual state space calculation module is as follows: Figure 3 As shown in the figure, the corresponding visual state space calculation modules in the two encoders of the twin network share model parameters. The visual state space calculation module contains three calculation paths and seven calculation submodules. In the middle calculation path, the first is the first layer normalization module, the second is the first linear calculation, the third is the deep convolution module, the fourth is the selective state space calculation module, and the fifth is the second layer normalization module. In the calculation path on the right, after the calculation of the first layer normalization module is completed, the calculation result is used as the input of the second linear calculation module, and the calculation result of the second linear calculation module is matrix multiplied with the input of the second layer normalization module in the middle calculation path, and the result of the matrix multiplication is input into the third linear calculation module layer. In the third calculation path, the original input and the calculation result of the third linear calculation module are matrix added.

[0048] The linear embedding coding calculation module obtains the coding vector of each small image formed by segmentation through linear calculation, and uses its calculation result as the input of the visual state space calculation module. The visual state space calculation module is used to learn the feature representation of the image layer by layer during the training process.

[0049] The selective state space calculation module is an algorithm model modeled on a linear time-invariant system as a prototype.

[0050]

[0051] The first equation h′(t)=Ah(t)+B(x) is the state equation, where h(t) is the hidden state and A∈C N×N It is a trainable parameter matrix that stores all historical information represented by the coefficient matrix and can be used to determine the influence of the hidden state in the past time on the hidden state used to update the next moment. The projection parameter vector B∈C N Determines the influence of the input x(t) on the hidden state.

[0052] The second equation y(t) = Ch(t) + Dx(t) is the output equation, describing the hidden state through the projection parameter matrix C∈C N Transformed into output, and input x(t) through the parameter D∈C 1 However, in order to simplify the equation, D = 0 is generally assumed.

[0053] After recursion, the formula becomes:

[0054]

[0055] After the calculations of the two visual state space calculation modules, the two output results are concatenated as tensors and then concatenated with the corresponding calculation results in the decoder using a jump connection.

[0056] Step 2.3: The second encoder sub-block of the encoder is a computing module consisting of a segmented image merging module and two visual state space computing modules. The segmented image merging module completes the merging of segmented images and the expansion of feature channels. By downsampling, the dimension of the tensor is reduced by half, and the dimension of the channel is doubled. The dimension is calculated to be Where H is the vertical resolution value of the original image, W is the horizontal resolution value of the original image, and C is the number of feature channels of the original image.

[0057] Step 2.4: The structure of the third encoder sub-block of the encoder is the same as that of the second encoder sub-block. After calculation by the segmented image merging module, a dimension of The tensor is the new image feature representation.

[0058] Step 2.5: The encoder's segmented image merging module converts the dimension of the tensor output by the third encoder sub-block into

[0059] Step 3 includes the following sub-steps:

[0060] Step 3.1: The connection part from the encoder to the decoder in the model is a calculation module composed of two visual state space calculation modules. The first visual state space calculation module receives the operation results calculated by the segmentation image merging modules in the two encoders, connects and merges the two encoding vectors, and then calculates them through the first visual state space calculation module.

[0061] Step 3.2: The second visual state space calculation module further calculates and obtains deep image features, and combines the subsequent jump connection process to merge the differentiated image features of the shallow and deep layers to enhance the fused spatial details.

[0062] Step 4 includes the following sub-steps:

[0063] Step 4.1, in the decoder, its specific structure is in correspondence with the encoder, including three decoder sub-blocks and a segmented image expansion module. The structures of the three decoder sub-blocks are the same, all consisting of a segmented image expansion module and two visual state space calculation modules, and receiving feature information transmitted by the corresponding components of the encoder through jump connections. The role of the segmented image expansion module is to upsample the extracted deep features, reconstruct the feature maps of adjacent dimensions into feature maps with higher resolution, and then use 2 times upsampling for the reconstructed feature maps with higher resolution, while reducing the dimension of the feature channel to half.

[0064] Step 4.2: The second decoder sub-block expands the image features into After that, the new image representation result is obtained through two calculations of the visual state space calculation module.

[0065] Step 4.3: The third decoder sub-block expands the image features into After that, the new image representation result is obtained through two calculations of the visual state space calculation module.

[0066] Step 4.4: By calculating the segmented image expansion module, the feature dimension of the feature map is restored to

[0067] Step 4.5: Set the dimension to The feature map of is used as the input of a linear mapping calculation module, and after calculation, the result feature map with a dimension of H×W×1 is output. The last dimension is compressed and reduced to H×W.

[0068] Step 4.6, pass the resulting feature map through a convolutional neural network with 2 convolution kernels, and output two matrices of dimension H×W, which are defined as the prediction mapping matrix of the center point of the change area Length and width prediction mapping of the changing area range

[0069] Step 5 includes the following sub-steps:

[0070] Step 5.1. The purpose of this method is to obtain the center position of the difference part in the two images through model calculation, and use the focus loss function to calculate the loss value.

[0071] The center point map of the image change area calculated based on the labeled data is defined as S hm ∈[0,1] 128×128×1 , that is, the value of the center point is 1 and the probability of the surrounding points decreases with the increase of distance in a Gaussian distribution, as shown in the following formula:

[0072]

[0073] in the formula is mapped to S hm The x coordinate of the marked center point, similarly is mapped to S hm The y coordinate of the marked center point. In the formula, x and y are the coordinate positions of the pixel points in the mapping image, and σ is the custom variance.

[0074] The focal loss function is defined as

[0075]

[0076] In this formula, N is the number of center points of the differenced part in the two images, and α and β are defined hyperparameters to penalize the gap between the predicted center point and the true value.

[0077] And the loss function of the length and width of the changing area:

[0078]

[0079] In the above formula is the length and width tuple of the pth data change region, and p is the number of the change region. These two loss functions are combined into the final loss function through the following linear function:

[0080] L=L hm +λ wh L wh (7)

[0081] where λ wh It is a custom hyperparameter used to balance the values ​​of various components in the loss function.

[0082] Step 5.2: Use the training set T for the model defined above train Complete the training and use T val Verify the model's capabilities during the training process and adjust the training parameters using T test Test the predictive ability of the model.

[0083] Step 6 includes the following sub-steps:

[0084] Step 6.1: Input two trademark images as trademark image data to be compared.

[0085] Step 6.2: The model outputs the center point coordinates, length, width and offset value of the change area.

[0086] Although the specific embodiments of the present invention are disclosed for the purpose of illustration, the purpose is to help understand the content of the present invention and implement it accordingly, those skilled in the art will understand that various substitutions, changes and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the content disclosed in the best embodiment, and the scope of the present invention is subject to the scope defined in the claims.

Claims

1. A method for detecting changes in similar trademarks based on a selective state space model, the steps of which include: 1) obtaining a trademark similarity change detection dataset, wherein the trademark similarity change detection dataset includes multiple pairs of sample images, each pair of sample images includes a trademark image and its similar image; 2) Using the trademark similarity change detection data set to train a selective state space model to obtain a similar trademark change detection model; the selective state space model includes an encoder, a connector, a decoder, a segmented image expansion module, a linear mapping calculation module and a convolutional neural network; wherein the encoder adopts a twin network architecture, including two branches with the same structure, each of which includes a first encoder sub-block, a second encoder sub-block, a third encoder sub-block and a segmented image merging module that sequentially processes input data; the connector includes two visual state space calculation modules connected in series; the decoder includes three decoder sub-blocks and a segmented image expansion module; one image in a pair of sample images is divided into multiple sub-images and input into one branch, and the other image in the pair of sample images is divided into multiple sub-images and input into another branch; The encoder is used to calculate the input image from shallow features to deep features, and input the deep features to the connector and send the shallow features to the decoder; wherein the first encoder sub-block is used to extract the first shallow features from each input sub-image and input them to the second encoder sub-block and the third decoder sub-block respectively, the second encoder sub-block is used to extract the second shallow features from the first shallow features of each sub-image and input them to the third encoder sub-block and the second decoder sub-block respectively, the third encoder sub-block is used to extract the deep features from the second shallow features of each sub-image and input them to the segmentation image merging module and the first decoder sub-block respectively; the segmentation image merging module is used to merge the deep features of each sub-image to obtain the deep features of the input image and input them to the connector; The connector is used to extract the input deep features and input the extracted features to the first decoder sub-block of the decoder; each decoder sub-block of the decoder decodes the input feature data to obtain image representations corresponding to different levels of each pair of sample images, and the image representations output by the third decoder sub-block are restored in turn by the segmentation image expansion module and the linear mapping calculation module, and then input into the convolutional neural network for prediction to obtain the change area corresponding to each pair of sample images; Calculate the loss value of the predicted change area of ​​each pair of sample images and the annotation information of the pair of sample images according to the selected loss function, and optimize the selective state space model according to the loss value; 3) Inputting a target trademark image and its similar trademark images into the similar trademark change detection model, and returning the change detection result of the similar trademark image relative to the target trademark.

2. The method according to claim 1, characterized in that The similar image corresponding to the trademark image in each pair of sample images is an image with the trademark image as the background and other icon data as the foreground.

3. The method according to claim 1, characterized in that Using the loss function L = L hm +λ wh L wh The loss value is calculated; where λ wh For custom hyperparameters, the focal loss function L hm Used to calculate the difference between the predicted change area center point and the labeled change area center point, the change area length and width loss function L wh Used to calculate the difference between the length and width of the predicted change area and the length and width of the marked change area.

4. The method according to claim 1, characterized in that: The first encoder sub-block includes a linear embedded coding calculation module that processes the input data in sequence and two serially connected visual state space calculation modules. The linear embedded coding calculation module is used to perform linear calculations on each input sub-image to obtain its encoding vector and input it into the visual state space calculation module; the visual state space calculation module is used to learn the feature representation of the image; the corresponding visual state space calculation modules in the two encoders of the twin network share model parameters.

5. The method according to claim 1, characterized in that The second encoder sub-block has the same structure as the third encoder sub-block, both of which include a segmented image merging module that processes input data in sequence and two serially connected visual state space calculation modules; wherein the segmented image merging module is used to complete the merging of segmented images and the expansion of feature channels.

6. The method according to claim 1, characterized in that The decoder sub-block includes a segmented image expansion module that processes input data in sequence and two serially connected visual state space calculation modules; wherein the segmented image expansion module is used to upsample the input features, reconstruct the feature maps of adjacent dimensions into feature maps with higher resolution, and then upsample the feature maps reconstructed with higher resolution by a factor of 2, while reducing the dimension of the feature channels of the feature maps reconstructed with higher resolution to half.

7. The method according to any one of claims 1 to 6, characterized in that: The visual state space calculation module includes three calculation paths in total; in the first calculation path, the original input data is processed by the first layer normalization module, the first linear calculation module, the deep convolution module, the selective state space calculation module, and the second layer normalization module in sequence; In the second calculation path, the calculation result of the first layer normalization module is used as the input of the second linear calculation module, and the calculation result of the second linear calculation module and the calculation result of the second layer normalization module are respectively input into the matrix multiplication operation unit for matrix multiplication operation, and the result of the matrix multiplication operation is input into the third linear calculation module layer; in the third calculation path, the original input data and the calculation result of the third linear calculation module are subjected to matrix addition operation.

8. The method according to claim 1, characterized in that The trademark image and its similar image in each pair of sample images are H×W×1 images, where the first encoder sub-block is used to extract The second encoder sub-block is used to extract the first shallow features of each sub-image from the first shallow features of each sub-image. The third encoder sub-block is used to extract from the second shallow features of each sub-image The segmented image merging module converts the dimension of the deep features output by the third encoder sub-block into First decoder sub-block output The image representation of the second decoder sub-block outputs The image representation of the third decoder sub-block outputs Image representation; where H is the vertical resolution value of the image, W is the horizontal resolution value of the image, C is the number of feature channels of the image.

9. A server, characterized in that: The invention comprises a memory and a processor, wherein the memory stores a computer program, the computer program is configured to be executed by the processor, and the computer program comprises instructions for executing the method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 8 is implemented.

Images