Trademark image retrieval method and retrieval device

Abstract

The present invention provides a retrieval method and retrieval device for trademark images, the method comprising: obtaining trademark data and preprocessing; constructing a hash network model and training it; using the trained hash network model to perform variable segment coding on the preprocessed trademark data to establish a variable segment coding library corresponding to the trademark data; using the trained hash network model to perform variable segment coding on the trademark image to be detected, coarsely matching the variable segment coding of the trademark image to be detected and the variable segment coding library, and generating a picture set I according to the coarse matching result; calculating the cosine similarity between the trademark image to be detected and each image in the picture set I, and extracting the image with the highest similarity to the trademark image to be detected according to the cosine similarity. The characteristic of the present invention is to perform trademark image recognition by constructing a hash network model, which can assist trademark registration auditors to effectively and quickly find trademarks similar to the trademark to be reviewed in the trademark library, reduce their workload, and improve work efficiency.

Description

Technical Field

[0001] The present invention relates to the field of image recognition technology, and specifically to a trademark image retrieval method, a trademark image retrieval device and a non-temporary computer-readable storage medium. Background Art

[0002] A trademark is a mark of a product, a mark that distinguishes the brand or service of one operator from the products or services of other operators. Trademarks are protected by law, and the registrant has exclusive rights. Trademarks are the product of the commodity economy, and in the modern global economic circle with close economic ties, trademarks play a unique role. After trademark registration, the registrant has exclusive rights to the trademark, which is protected by law and can better maintain his own brand, not only protecting the long-term development of the brand he operates, but also effectively preventing counterfeiters and protecting his own economic interests by legal means.

[0003] With the increasing number of trademarks and the increasing frequency of trademark infringement lawsuits, trademark infringement retrieval has become one of the research hotspots in the field of information science. How trademark registration reviewers can efficiently and quickly retrieve suspected infringing images from a large number of trademark images and how trademark registration personnel can effectively avoid infringement risks are a series of problems that need to be solved. Summary of the invention

[0004] In order to solve the above technical problem, the first embodiment of the present invention provides a method for retrieving trademark images.

[0005] A second embodiment of the present invention provides a trademark image retrieval device.

[0006] A third aspect of the present invention provides a non-transitory computer-readable storage medium.

[0007] The technical solution adopted by the present invention is as follows:

[0008] The first aspect of the present invention proposes a method for retrieving trademark images, including the following steps: obtaining trademark data and preprocessing it, wherein the trademark data includes at least one trademark image; constructing a hash network model and training it; using the trained hash network model to perform variable segment encoding on the preprocessed trademark data to establish a variable segment encoding library corresponding to the trademark data; obtaining a trademark image to be detected, using the trained hash network model to perform variable segment encoding on the trademark image to be detected, coarsely matching the variable segment encoding of the trademark image to be detected and the variable segment encoding library, and generating a picture set I based on the coarse matching result; calculating the cosine similarity between the trademark image to be detected and each image in the picture set I, and extracting the image with the highest similarity to the trademark image to be detected based on the cosine similarity.

[0009] The above-mentioned trademark image retrieval method of the present invention also has the following additional technical features:

[0010] According to one embodiment of the present invention, trademark data is obtained and preprocessed, specifically including: geometric enhancement and color transformation enhancement of the trademark image; learning the distribution of data through a GAN (Generative Adversarial Network) model, and randomly generating pictures consistent with the distribution of the trademark data for sample expansion; cropping, color channel conversion and image pixel normalization of the trademark image for format conversion.

[0011] According to one embodiment of the present invention, a hash network model is constructed and trained, specifically including: extracting high-dimensional deep features of a training set through a feature extraction network; establishing a graph convolutional autoencoder for supervising hash network learning, and calculating a similarity matrix A based on the high-dimensional deep features as input to train the graph convolutional autoencoder; establishing a hash network for feature dimensionality reduction, training the hash network using the high-dimensional deep features, and using the trained graph convolutional autoencoder for supervision; reducing the dimensionality of high-dimensional features through the hash network, performing Gaussian fitting on the reduced features, and then performing variable segment encoding.

[0012] According to an embodiment of the present invention, variable segment coding is performed specifically by the following formula:

[0013] B[i][j]=000,F Gauss [i][j]∈(-∞,u[i]-2σ[i]]

[0014] B[i][j]=001,F Gauss [i][j]∈(u[i]-2σ[i],u[i]]

[0015] B[i][j]=011,F Gauss [i][j]∈(u[i],u[i]+2σ[i]]

[0016] B[i][j]=111,F Gauss [i][j]∈(u[i]+2σ[i],+∞)

[0017] Where i = 1, 2, ..., N, j = 1, 2, ..., M, N is the number of images, M is the dimension of the low-dimensional feature, and F Gauss [i][j] are the features of the low-dimensional features after Gaussian fitting, and u[i] is the image x i Corresponding to the mean of low-dimensional features, σ[i] is the image x i Corresponding to the variance of the low-dimensional feature, B[i][j] is the hash code after variable segment coding.

[0018] According to one embodiment of the present invention, a coarse match is performed on the variable segment code of the trademark image to be detected and the variable segment code library, and a picture set I is generated based on the coarse matching result, specifically including: calculating the Hamming distance between the variable segment code of the trademark image to be detected and the trademark image code in the variable segment code library, and sorting the calculation results; obtaining the trademark data corresponding to a Hamming distance less than a preset distance, and forming a picture set I.

[0019] The second aspect of the present invention proposes a retrieval device for trademark images, including: an acquisition module, the acquisition module is used to acquire and preprocess trademark data, the trademark data includes at least one trademark image; a training module, the training module is used to construct a hash network model and perform training; an establishment module, the establishment module is used to use the trained hash network model to perform variable segment encoding on the preprocessed trademark data to establish a variable segment encoding library corresponding to the trademark data; a coarse matching module, the coarse matching module is used to acquire a trademark image to be detected, use the trained hash network model to perform variable segment encoding on the trademark image to be detected, perform coarse matching on the variable segment encoding of the trademark image to be detected and the variable segment encoding library, and generate a picture set I according to the coarse matching result; a fine matching module, calculates the cosine similarity between the trademark image to be detected and each image in the picture set I, and extracts the image with the highest similarity to the trademark image to be detected according to the cosine similarity.

[0020] The above-mentioned trademark image retrieval device of the present invention also has the following additional technical features:

[0021] According to one embodiment of the present invention, the training module is specifically used to: extract high-dimensional deep features of the training set through a feature extraction network; establish a graph convolutional autoencoder for supervising hash network learning, and calculate the similarity matrix A as input to train the graph convolutional autoencoder based on the high-dimensional deep features; establish a hash network for feature dimensionality reduction, use the high-dimensional deep features to train the hash network, and use the trained graph convolutional autoencoder for supervision; reduce the dimensionality of high-dimensional features through the hash network, perform Gaussian fitting on the reduced features, and then perform variable segment encoding.

[0022] According to an embodiment of the present invention, the training module specifically performs variable segment coding through the following formula:

[0023] B[i][j]=000,F Gauss [i][j]∈(-∞,u[i]-2σ[i]]

[0024] B[i][j]=001,F Gauss [i][j]∈(u[i]-2σ[i],u[i]]

[0025] B[i][j]=011,F Gauss [i][j]∈(u[i],u[i]+2σ[i]]

[0026] B[i][j]=111,F Gauss [i][j]∈(u[i]+2σ[i],+∞)

[0027] Where i = 1, 2, ..., N, j = 1, 2, ..., M, N is the number of images, M is the dimension of the low-dimensional feature, and F Gauss [i][j] are the features of the low-dimensional features after Gaussian fitting, and u[i] is the image x i Corresponding to the mean of low-dimensional features, σ[i] is the image x i Corresponding to the variance of the low-dimensional feature, B[i][j] is the hash code after variable segment coding.

[0028] The third aspect of the present invention provides a non-temporary computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the trademark image retrieval method described in the first aspect of the present invention.

[0029] Beneficial effects of the present invention:

[0030] The present invention constructs a hash network model to perform trademark image recognition, which can not only assist trademark registration reviewers to effectively and quickly find trademarks in the trademark library that are similar to the trademark to be reviewed, thereby reducing their workload and improving work efficiency, but also provide information guidance for trademark registration applicants to avoid the risk of trademark infringement before applying for a trademark, thereby increasing the success rate of trademark applications, which is of great significance to improving the efficiency of trademark application and review.

[0031] When training the hash network model, the present invention adopts variable segment coding after nonlinear fitting of features, which has a certain smoothing effect and can avoid overfitting, thereby improving the recognition accuracy of the network. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 is a flowchart of a method for retrieving a trademark image according to an embodiment of the present invention;

[0033] Figure 2 is a flowchart of a method for retrieving a trademark image according to another embodiment of the present invention.

[0034] Figure 3 is a diagram showing the principle of training a hash network according to an embodiment of the present invention;

[0035] Figure 4 FIG. 4 is a block diagram of a device for searching trademark images according to an embodiment of the present invention. DETAILED DESCRIPTION

[0036] 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.

[0037] Figure 1 FIG. 1 is a flowchart of a method for retrieving a trademark image according to an embodiment of the present invention. Figure 1 As shown, the method comprises the following steps:

[0038] S1, obtaining trademark data and performing preprocessing, where the trademark data includes at least one trademark image.

[0039] Furthermore, according to an embodiment of the present invention, trademark data is obtained and preprocessed, specifically including: geometric enhancement and color transformation enhancement of the trademark image; learning the distribution of data through the GAN model, and randomly generating pictures consistent with the distribution of the trademark data for sample expansion; cropping, color channel conversion and image pixel normalization of the trademark image for format conversion so that the image meets the input requirements of the feature extraction network.

[0040] S2, build a hash network model and train it.

[0041] Further, according to one embodiment of the present invention, Figure 2 The hash network model is constructed and trained as shown, including:

[0042] S21, extract high-dimensional deep features of the training set through the feature extraction network.

[0043] S22, establish a graph convolutional autoencoder for supervised hash network learning, and calculate the similarity matrix A based on high-dimensional deep features as input to train the graph convolutional autoencoder.

[0044] S23, establish a hash network for feature dimensionality reduction, use high-dimensional deep features to train the hash network, and use the trained graph convolutional autoencoder for supervision.

[0045] S24, the high-dimensional features are reduced in dimension through the hash network, and the reduced features are subjected to Gaussian fitting and then variable segment encoding.

[0046] Specifically, Figure 3As shown in Figure 1, a graph convolutional autoencoder for supervised learning is built, and the high-dimensional features extracted by the feature extraction are used to train the autoencoder to obtain the autoencoder. Considering that the global similarity relationship between all images in the database and the topological structure in the graph network have a "certain degree" of correlation, a graph convolutional autoencoder is used as a supervisory network to extract the high-dimensional deep features F extracted by the feature extraction network. high ={f(ω high , x i )|i=1,2,...N} and the similarity matrix A calculated based on the high-dimensional deep features are used as input to train the graph convolutional autoencoder, where f is the feature extraction network, ω high is the network weight, x i is the data of the training set, F high The above-trained graph convolutional autoencoder is used to train the hash network. The trained hash network reduces the dimension of the high-dimensional deep features to obtain the reduced-dimensional feature F low ={h(ω low , x i )i=1,2,...N},where h is the hash network, ω high is the network weight, x i is the data of the training set, F low is the feature after dimensionality reduction.

[0047] S3, using the trained hash network model to perform variable segment encoding on the preprocessed trademark data to establish a variable segment encoding library corresponding to the trademark data.

[0048] Furthermore, in one implementation of the present invention, variable segment coding is performed specifically by the following formula:

[0049] B[i][j]=000,F Gauss [i][j]∈(-∞,u[i]-2σ[i]]

[0050] B[i][j]=001,F Gauss [i][j]∈(u[i]-2σ[i],u[i]]

[0051] B[i][j]=011,F Gauss [i][j]∈(u[i],u[i]+2σ[i]]

[0052] B[i][j]=111,F Gauss [i][j]∈(u[i]+2σ[i],+∞)

[0053] Where i = 1, 2, ..., N, j = 1, 2, ..., M, N is the number of images, M is the dimension of the low-dimensional feature, and F Gauss[i][j] are the features of the low-dimensional features after Gaussian fitting, and u[i] is the image x i Corresponding to the mean of low-dimensional features, σ[i] is the image x i Corresponding to the variance of the low-dimensional feature, B[i][j] is the hash code after variable segment coding.

[0054] Specifically, after Gaussian fitting of the reduced-dimensional features, four-segment hash coding is used, and three-bit binary is used to encode each segment. Different nonlinear fitting methods can also be used or the coding bits can be expanded according to different needs. In this way, there is a certain smoothing effect, overfitting can be avoided, and the recognition accuracy of the network can be improved.

[0055] S4, obtaining a trademark image to be detected, using the trained hash network model to perform variable segment encoding on the trademark image to be detected, performing coarse matching between the variable segment encoding of the trademark image to be detected and the variable segment encoding library, and generating a picture set I according to the coarse matching result.

[0056] Further, according to an embodiment of the present invention, a coarse match is performed between the variable segment code of the trademark image to be detected and the variable segment code library, and a picture set I is generated based on the coarse matching result, specifically including: calculating the Hamming distance between the variable segment code of the trademark image to be detected and the trademark image code in the variable segment code library, and sorting the calculation results; obtaining the trademark data corresponding to a Hamming distance less than a preset distance, and forming a picture set I.

[0057] S5, calculating the cosine similarity between the trademark image to be detected and each image in the image set I, and extracting the image with the highest similarity to the trademark image to be detected according to the cosine similarity.

[0058] Specifically, the cosine similarity between the low-dimensional features of the image to be retrieved and the images in the picture set I is calculated, and the image corresponding to the maximum cosine similarity is extracted, which is the top-k image with the highest similarity.

[0059] In summary, the trademark image retrieval method according to the embodiment of the present invention can perform trademark image recognition by constructing a hash network model, which can not only assist trademark registration review personnel to effectively and quickly find trademarks similar to the trademark to be reviewed in the trademark library, reduce their workload and improve work efficiency, but also provide information guidance for trademark registration applicants, avoid the risk of trademark infringement before trademark application, thereby increasing the success rate of trademark applications, which is of great significance to improving the efficiency of trademark application and review. When training the hash network model, variable segment coding after nonlinear fitting of features is used, which has a certain smoothing effect, can avoid overfitting, and improve the recognition accuracy of the network.

[0060] Corresponding to the above-mentioned trademark image retrieval method, the present invention also proposes a trademark image retrieval device. Since the device embodiment of the present invention corresponds to the above-mentioned method embodiment, details not disclosed in the device embodiment can be referred to the above-mentioned method embodiment and will not be repeated in the present invention.

[0061] Figure 4 FIG. 1 is a block diagram of a retrieval device for a trademark image according to an embodiment of the present invention. Figure 4 As shown, the device includes: an acquisition module 1, a training module 2, a building module 3, a coarse matching module 4 and a fine matching module 5.

[0062] Among them, the acquisition module 1 is used to acquire and preprocess trademark data, and the trademark data includes at least one trademark image; the training module 2 is used to construct a hash network model and perform training; the establishment module 3 is used to use the trained hash network model to perform variable segment encoding on the preprocessed trademark data to establish a variable segment encoding library corresponding to the trademark data; the coarse matching module 4 is used to acquire the trademark image to be detected, use the trained hash network model to perform variable segment encoding on the trademark image to be detected, perform coarse matching between the variable segment encoding of the trademark image to be detected and the variable segment encoding library, and generate a picture set I according to the coarse matching result; the fine matching module 5 is used to calculate the cosine similarity between the trademark image to be detected and each image in the picture set I, and extract the image with the highest similarity to the trademark image to be detected according to the cosine similarity.

[0063] According to one embodiment of the present invention, the training module 2 is specifically used to: extract high-dimensional deep features of the training set through a feature extraction network; establish a graph convolutional autoencoder for supervising hash network learning, and calculate the similarity matrix A as input to train the graph convolutional autoencoder based on the high-dimensional deep features; establish a hash network for feature dimensionality reduction, train the hash network using high-dimensional deep features, and use the trained graph convolutional autoencoder for supervision; reduce the dimensionality of high-dimensional features through the hash network, perform Gaussian fitting on the reduced features, and then perform variable segment encoding.

[0064] According to one embodiment of the present invention, the training module 2 specifically performs variable segment coding through the following formula:

[0065] B[i][j]=000,F Gauss [i][j]∈(-∞,u[i]-2σ[i]]

[0066] B[i][j]=001,F Gauss [i][j]∈(u[i]-2σ[i],u[i]]

[0067] B[i][j]=011,F Gauss [i][j]∈(u[i],u[i]+2σ[i]]

[0068] B[i][j]=111,F Gauss [i][j]∈(u[i]+2σ[i],+∞)

[0069] Where i = 1, 2, ..., N, j = 1, 2, ..., M, N is the number of images, M is the dimension of the low-dimensional feature, and F Gauss [i][j] are the features of the low-dimensional features after Gaussian fitting, and u[i] is the image x i Corresponding to the mean of low-dimensional features, σ[i] is the image x i Corresponding to the variance of the low-dimensional feature, B[i][j] is the hash code after variable segment coding.

[0070] According to the retrieval device of the trademark image of the embodiment of the present invention, by constructing a hash network model for trademark image recognition, it can not only assist trademark registration review personnel to effectively and quickly find trademarks similar to the trademark to be reviewed in the trademark library, reduce their workload and improve work efficiency, but also provide information guidance for trademark registration applicants to avoid the risk of trademark infringement before trademark application, thereby increasing the success rate of trademark applications, which is of great significance to improving the efficiency of trademark application and review. When training the hash network model, variable segment coding after nonlinear fitting of features is used, which has a certain smoothing effect, can avoid overfitting, and improve the recognition accuracy of the network.

[0071] The embodiment of the present invention further provides a non-transitory computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the method for retrieving trademark images described in the above embodiment is implemented.

[0072] According to the non-transitory computer-readable storage medium of an embodiment of the present invention, when the computer program stored thereon is executed by the processor, trademark data is obtained and preprocessed, a hash network model is constructed and trained, and then the preprocessed trademark data is variable-segment encoded using the trained hash network model to establish a variable-segment encoding library corresponding to the trademark data, and then the trademark image to be detected is obtained, and the trademark image to be detected is variable-segment encoded using the trained hash network model, and the variable-segment encoding of the trademark image to be detected and the variable-segment encoding library are roughly matched, and a picture set I is generated according to the rough matching result, and the cosine similarity between the trademark image to be detected and each image in the picture set I is calculated, and the image with the highest similarity to the trademark image to be detected is extracted according to the cosine similarity, so that the trademark registration review personnel can effectively and quickly find trademarks similar to the trademark to be reviewed in the trademark library, reduce their workload and improve work efficiency, and provide information guidance for trademark registration applicants, avoid the risk of trademark infringement before trademark application, thereby improving the success rate of trademark applications, which is of great significance to improving the efficiency of trademark application and review. When training the hash network model, variable segment coding after nonlinear fitting of features is used, which has a certain smoothing effect and can avoid overfitting, thereby improving the recognition accuracy of the network.

[0073] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0074] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of the present invention, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0075] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, segment or portion of code comprising one or more executable instructions for implementing the steps of a custom logical function or process, and the scope of the preferred embodiments of the present invention includes alternative implementations in which functions may not be performed in the order shown or discussed, including performing functions in a substantially simultaneous manner or in reverse order depending on the functions involved, which should be understood by technicians in the technical field to which the embodiments of the present invention belong.

[0076] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute the instructions), or in combination with these instruction execution systems, devices or apparatuses. For the purpose of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in combination with these instruction execution systems, devices or apparatuses. More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk box (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium and then editing, interpreting or processing in other suitable ways if necessary, and then stored in a computer memory.

[0077] It should be understood that each part of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system.

[0078] A person skilled in the art may understand that all or part of the steps in the method for implementing the above-mentioned embodiment may be completed by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, which, when executed, includes one or a combination of the steps of the method embodiment.

[0079] In addition, each functional unit in each embodiment of the present invention may be integrated into a processing module, or each unit may exist physically separately, or two or more units may be integrated into one module. The above-mentioned integrated module may be implemented in the form of hardware or in the form of a software functional module. If the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it may also be stored in a computer-readable storage medium.

[0080] The storage medium mentioned above may be a read-only memory, a magnetic disk or an optical disk, etc. Although the embodiments of the present invention have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limiting the present invention. A person of ordinary skill in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A trademark image retrieval method, characterized in that: The following steps are involved: Acquire and pre-process trademark data, wherein the trademark data includes at least one trademark image; Build a hash network model and train it; The trained hash network model is used to perform variable segment encoding on the pre-processed trademark data to establish a variable segment encoding library corresponding to the trademark data; Acquire a trademark image to be detected, use the trained hash network model to perform variable segment encoding on the trademark image to be detected, perform rough matching on the variable segment encoding of the trademark image to be detected and the variable segment encoding library, and generate a picture set I according to the rough matching result; Calculating the cosine similarity between the trademark image to be detected and each image in the picture set I, and extracting the image with the highest similarity to the trademark image to be detected according to the cosine similarity; Specifically, variable segment coding is performed through the following formula: , , , , in, , is the number of images, is the dimension of the low-dimensional feature, is the feature after Gaussian fitting of low-dimensional features, For images The mean value of the corresponding low-dimensional feature, For images The variance corresponding to the low-dimensional feature, It is the hash code after variable segment encoding; Build a hash network model and train it, including: Extract high-dimensional deep features of the training set through the feature extraction network; Build a graph convolutional autoencoder for supervised hash network learning and calculate the similarity matrix based on the high-dimensional deep features As input to train the graph convolutional autoencoder; Establishing a hash network for feature dimensionality reduction, using the high-dimensional deep features to train the hash network, and using the trained graph convolutional autoencoder for supervision; The high-dimensional features are reduced in dimension through the hash network, and the reduced-dimensional features are subjected to Gaussian fitting and then variable segment encoding.

2. The trademark image retrieval method according to claim 1, characterized in that: Obtain trademark data and perform preprocessing, including: Perform geometric enhancement and color transformation enhancement on trademark images; The GAN model is used to learn the distribution of data and randomly generate images that are consistent with the trademark data distribution for sample expansion; The trademark image is cropped, color channel converted, and image pixel normalized for format conversion.

3. The trademark image retrieval method according to claim 1, characterized in that: The variable segment code of the trademark image to be detected is roughly matched with the variable segment code library, and a picture set I is generated according to the rough matching result, specifically including: Calculating the Hamming distance between the variable segment code of the trademark image to be detected and the trademark image code in the variable segment code library, and sorting the calculation results; The trademark data corresponding to the Hamming distance being less than the preset distance is obtained to form an image set I.

4. A trademark image retrieval device, characterized in that: include: An acquisition module, the acquisition module is used to acquire trademark data and perform preprocessing, the trademark data includes at least one trademark image; A training module, wherein the training module is used to construct a hash network model and perform training; An establishment module, the establishment module is used to use the trained hash network model to perform variable segment encoding on the pre-processed trademark data to establish a variable segment encoding library corresponding to the trademark data; A coarse matching module, the coarse matching module is used to obtain a trademark image to be detected, use the trained hash network model to perform variable segment coding on the trademark image to be detected, perform coarse matching on the variable segment coding of the trademark image to be detected and the variable segment coding library, and generate a picture set I according to the coarse matching result; A fine matching module, which is used to calculate the cosine similarity between the trademark image to be detected and each image in the picture set I, and extract the image with the highest similarity to the trademark image to be detected according to the cosine similarity; Specifically, variable segment encoding is performed through the following formula: , , , , in, , is the number of images, is the dimension of the low-dimensional feature, is the feature after Gaussian fitting of low-dimensional features, For images The mean value of the corresponding low-dimensional feature, For images The variance corresponding to the low-dimensional feature, It is the hash code after variable segment encoding; The training module is specifically used for: Extract high-dimensional deep features of the training set through the feature extraction network; Build a graph convolutional autoencoder for supervised hash network learning and calculate the similarity matrix based on the high-dimensional deep features As input to train the graph convolutional autoencoder; Establishing a hash network for feature dimensionality reduction, using the high-dimensional deep features to train the hash network, and using the trained graph convolutional autoencoder for supervision; The high-dimensional features are reduced in dimension through the hash network, and the reduced-dimensional features are subjected to Gaussian fitting and then variable segment encoding.

5. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method for retrieving a trademark image as described in any one of claims 1 to 3 is implemented.

Images