Trademark image retrieval method and system based on deep learning

Abstract

The invention relates to the technical field of trademark image retrieval, and discloses a trademark image retrieval method and system based on deep learning, and the method comprises the steps: obtaining a trademark image; scoring the trademark image to obtain image complexity; according to the trademark image and the image complexity, wavelet transformation is carried out to obtain a feature vector, dynamic adjustment is carried out, and feature representation is obtained; according to the feature representation, inputting the feature representation into a preset semantic association model to obtain candidate association information; according to the candidate association information, calculating the similarity between the image semantic vector and the brand node by using a knowledge graph reasoning technology, and obtaining an association result based on a similarity threshold value; dynamically generating a text label of the trademark image according to an association result, and mapping the text label and an entity in the brand knowledge base by adopting an entity linking technology to obtain a mapping result; and inputting the mapping result into a preset semantic link network to obtain a trademark retrieval result. According to the method, the accuracy of trademark image retrieval can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of icon image retrieval, and particularly to a trademark image retrieval method and system based on deep learning. Background Art

[0002] Currently, with the rapid development of the brand economy, trademark image retrieval systems play an increasingly important role in the fields of intellectual property protection, brand management, and market monitoring. Existing trademark databases have shown the characteristics of being massive and diverse. There are significant differences in the complexity (including graphic hierarchy, color composition, element density, etc.) and design styles (such as minimalism, abstract art, realistic style, etc.) of trademark designs in different industries. This diversity makes it difficult for traditional retrieval systems to establish an effective association model between image features and brand information. Especially in cross - industry retrieval scenarios, the system often fails to accurately identify trademark images with similar visual features but belonging to different categories.

[0003] In an existing technology, a trademark retrieval method based on deep hashing extracts image features through a convolutional neural network, then compresses the feature dimension using binary hash codes to achieve fast retrieval, and optimizes the hash function through adversarial training to minimize the Hamming distance of hash codes for similar trademarks.

[0004] Existing technologies generally have the problem of disconnection between image feature representation and brand information semantics. Key information such as trademark detail textures and stylized elements is lost during the feature compression process, resulting in the inability to effectively distinguish trademarks with similar main body structures but significant local design differences, and causing low accuracy in trademark image retrieval. Summary of the Invention

[0005] The present invention provides a trademark image retrieval method and system based on deep learning to solve the problem of low accuracy in trademark image retrieval.

[0006] In a first aspect, to solve the above - mentioned technical problem, the present invention provides a trademark image retrieval method and system based on deep learning, including: Obtain a trademark image; According to the trademark image, perform quantization scoring to obtain the image complexity; According to the trademark image and the image complexity, perform wavelet transform to obtain a feature vector, and dynamically adjust according to the feature vector to obtain a feature representation; Input the feature representation into a preset semantic association model to obtain candidate association information; According to the candidate association information, use knowledge graph reasoning technology to calculate the similarity between the image semantic vector and the brand node, and filter out high - confidence association results by setting a similarity threshold; According to the associated result, dynamically generate a text label for the trademark image, and use entity linking technology to map the text label to an entity in the brand knowledge base to obtain a mapping result; Input the mapping result into a preset semantic linking network to obtain a trademark retrieval result.

[0007] In an alternative embodiment, the quantifying and scoring the trademark image to obtain an image complexity includes: According to the trademark image, calculate the image texture complexity score through a gray-level co-occurrence matrix; According to the trademark image, calculate the color richness score using a color histogram method; According to the trademark image, calculate the shape diversity score using a contour analysis method; Based on the texture complexity score, the color richness score, and the shape diversity score, weighted to obtain the image complexity.

[0008] In an alternative embodiment, the performing wavelet transform on the trademark image and the image complexity to obtain a feature vector, and dynamically adjusting according to the feature vector to obtain a feature representation includes: According to the image complexity, determine whether it exceeds a preset complexity threshold. If it exceeds the complexity threshold, determine it as a complex trademark image. If it does not exceed the complexity threshold, skip the following steps; Perform wavelet transform on the complex trademark image, obtain image information in different frequency bands, and perform feature extraction on the image information in different frequency bands to obtain a feature vector; By calculating the background interference degree and visual redundancy degree of the complex trademark image, dynamically adjust the weight of the feature vector to obtain a weight adjustment result; According to the weight adjustment result, perform feature fusion to obtain a fine-grained feature representation.

[0009] In an alternative embodiment, the inputting the feature representation into a preset semantic association model to obtain candidate association information includes: According to the feature representation, input it into a preset semantic association model, and map the features to a high-dimensional semantic space through a fully connected layer to generate an image semantic vector; According to a pre-trained brand dictionary, obtain a set of brand semantic vectors; Calculate the cosine similarity between the image semantic vector and the brand semantic vector; According to the cosine similarity and a preset similarity threshold, filter out brand semantic vectors higher than the similarity threshold to obtain a candidate brand set; Perform clustering analysis on the brand semantic vectors in the candidate brand set, divide the brand association categories, and obtain candidate association information according to the brand association categories.

[0010] In an alternative embodiment, the method of calculating the similarity between the image semantic vector and the brand node using the knowledge graph reasoning technology according to the candidate association information, and screening out the high-confidence association results by setting a similarity threshold includes: According to the candidate association information, use the knowledge graph reasoning technology to obtain brand node information from the candidate brand set; Calculate the cosine similarity between the image semantic vector and the brand node information to generate a similarity matrix; According to the preset similarity threshold, screen the brand nodes in the similarity matrix that are higher than the threshold to determine the high-confidence brand association results; According to the high-confidence brand association results, if there are multiple high-confidence brand nodes, use a clustering algorithm to group the brand nodes; According to the grouped results of the brand nodes after clustering, generate a mapping relationship table between the brand association category and the image semantic vector; According to the mapping relationship table, output the association results between the image semantic vector and the brand node; Store the association results in the knowledge graph, update the association relationship between the brand node and the image semantic vector, and obtain the high-confidence association results.

[0011] In an alternative embodiment, the method of dynamically generating a text label for the trademark image according to the association result and using the entity linking technology to map the text label to an entity in the brand knowledge base to obtain a mapping result includes: Dynamically generate a text label for the trademark image according to the association result; Match the text label with the pre-established brand knowledge base, and use the entity linking technology to determine the brand entity corresponding to the text label; According to the brand entity, establish a mapping relationship between the image feature and the brand entity; According to the mapping relationship, construct a semantic link network between the image and the brand, and use the cross-modal retrieval technology to map the text label to an entity in the brand knowledge base according to the semantic link network to obtain a mapping result.

[0012] In an alternative embodiment, the method of inputting the mapping result into a preset semantic link network to obtain a trademark retrieval result includes: Arrange the brand information candidate set in descending order according to the mapping result to generate a preliminary retrieval result; According to the preliminary retrieval results, data cleaning is performed to delete null values and outliers, and the processed retrieval results are obtained; According to the processed retrieval results and combined with the user's historical retrieval records, personalized adjustment is performed through a collaborative filtering algorithm to obtain the final retrieval results.

[0013] In a second aspect, the present invention provides a trademark image retrieval system based on deep learning, including: A trademark acquisition module for acquiring trademark images; A complexity calculation module for performing quantization scoring according to the trademark image to obtain the image complexity; A feature extraction module for performing wavelet transform on the trademark image and the image complexity to obtain a feature vector, and dynamically adjusting according to the feature vector to obtain a feature representation; An associated information acquisition module for inputting the feature representation into a preset semantic association model to obtain candidate associated information; A similarity matching module for calculating the similarity between the image semantic vector and the brand node according to the candidate associated information by using knowledge graph reasoning technology, and screening out high-confidence associated results by setting a similarity threshold; An entity mapping module for dynamically generating a text label of the trademark image according to the associated result, and using entity linking technology to map the text label to an entity in the brand knowledge base to obtain a mapping result; A trademark retrieval module for inputting the mapping result into a preset semantic link network to obtain a trademark retrieval result.

[0014] In a third aspect, the present invention further provides an electronic device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the method for trademark image retrieval based on deep learning described in any one of the above.

[0015] In a fourth aspect, the present invention further provides a computer-readable storage medium. The computer-readable storage medium includes a stored computer program. When the computer program runs, it controls the device where the computer-readable storage medium is located to execute the method for trademark image retrieval based on deep learning described in any one of the above.

[0016] Compared with the prior art, the present invention has the following beneficial effects: The present invention discloses a trademark image retrieval method based on deep learning, including obtaining trademark images; performing quantization scoring according to the trademark images to obtain image complexity; performing wavelet transform according to the trademark images and the image complexity to obtain feature vectors, and dynamically adjusting according to the feature vectors to obtain feature representations; inputting the feature representations into a preset semantic association model to obtain candidate association information; according to the candidate association information, using knowledge graph reasoning technology to calculate the similarity between the image semantic vector and the brand node, and screening out high-confidence association results by setting a similarity threshold; dynamically generating text labels of the trademark images according to the association results, and using entity linking technology to map the text labels to entities in the brand knowledge base to obtain mapping results; inputting the mapping results into a preset semantic link network to obtain trademark retrieval results.

[0017] The trademark image retrieval method based on deep learning proposed by the present invention significantly improves the accuracy and efficiency of trademark image retrieval through multi-level information processing and intelligent reasoning technology. First, obtaining the complexity of the image through quantization scoring helps to make refined distinctions for different types of trademarks, avoiding the misdetection problems caused by ignoring details due to complexity in traditional methods. Then, using wavelet transform to extract the feature vectors of the image and performing dynamic adjustment not only enhances the feature expression ability of the trademark image but also enables the model to adapt to the visual differences of different trademarks. Further, through the preset semantic association model and knowledge graph reasoning technology, candidate association information can be accurately obtained, and by calculating the similarity between the image semantic vector and the brand node, the most relevant trademark images can be screened out. This combination method based on semantic vectors and knowledge graphs effectively reduces the noise interference in trademark image retrieval, thereby improving the relevance of the retrieval results. In addition, using entity linking technology to accurately map the trademark images to entities in the brand knowledge base not only ensures the accuracy of the retrieval results but also makes the final retrieval results have stronger interpretability and practicality. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a schematic flowchart of a trademark image retrieval method based on deep learning provided by the first embodiment of the present invention; Figure 2 is a schematic structural diagram of a trademark image retrieval system based on deep learning provided by the second embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0019] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0020] Referring to Figure 1 , the first embodiment of the present invention provides a trademark image retrieval method based on deep learning, including the following steps: S11, obtaining a trademark image; S12, performing quantization scoring according to the trademark image to obtain the image complexity; S13, performing wavelet transform on the trademark image and the image complexity to obtain a feature vector, and dynamically adjusting according to the feature vector to obtain a feature representation; S14, inputting the feature representation into a preset semantic association model to obtain candidate association information; S15, using the knowledge graph reasoning technology according to the candidate association information to calculate the similarity between the image semantic vector and the brand node, and screening out high-confidence association results by setting a similarity threshold; S16, dynamically generating a text label for the trademark image according to the association result, and using entity linking technology to map the text label to an entity in the brand knowledge base to obtain a mapping result; S17, inputting the mapping result into a preset semantic link network to obtain a trademark retrieval result.

[0021] The present invention discloses a trademark image retrieval method based on deep learning, including obtaining a trademark image; performing quantization scoring according to the trademark image to obtain the image complexity; performing wavelet transform on the trademark image and the image complexity to obtain a feature vector, and dynamically adjusting according to the feature vector to obtain a feature representation; inputting the feature representation into a preset semantic association model to obtain candidate association information; using the knowledge graph reasoning technology according to the candidate association information to calculate the similarity between the image semantic vector and the brand node, and screening out high-confidence association results by setting a similarity threshold; dynamically generating a text label for the trademark image according to the association result, and using entity linking technology to map the text label to an entity in the brand knowledge base to obtain a mapping result; inputting the mapping result into a preset semantic link network to obtain a trademark retrieval result.

[0022] The trademark image retrieval method based on deep learning proposed by the present invention significantly improves the accuracy and efficiency of trademark image retrieval through multi-level information processing and intelligent reasoning techniques. First, the complexity of the image is obtained through quantitative scoring, which helps to refine the distinction of different types of trademarks and avoids the misdetection problem caused by ignoring details due to complexity in traditional methods. Then, wavelet transform is used to extract the feature vectors of the image and perform dynamic adjustment, which not only enhances the feature expression ability of the trademark image but also enables the model to adapt to the visual differences of different trademarks. Further, through the preset semantic association model and knowledge graph reasoning technology, candidate association information can be accurately obtained, and by calculating the similarity between the image semantic vector and the brand node, the most relevant trademark image can be screened out. This combined method based on semantic vectors and knowledge graphs effectively reduces the noise interference in trademark image retrieval, thereby improving the relevance of the retrieval results. In addition, the entity linking technology is used to accurately map the trademark image to the entity in the brand knowledge base, which not only ensures the accuracy of the retrieval results but also makes the final retrieval results more interpretable and practical.

[0023] In step S11, it is necessary to obtain the trademark image.

[0024] It should be noted that the acquisition of trademark images can usually be carried out through multiple channels. First, the image data of registered trademarks can be obtained through public trademark databases, such as the website of the National Trademark Office or the platform of the International Trademark Organization. Second, web crawler technology can also be used to collect product images with trademarks from the Internet, or relevant image data can be obtained from e-commerce platforms, brand official websites and other channels. In addition, the data set can be enriched through manual annotation or the trademark image library provided by partners. Through these methods, the diversity and representativeness of trademark images can be ensured, providing high-quality data support for the subsequent deep learning model training and retrieval.

[0025] In step S12, according to the trademark image, quantitative scoring is performed to obtain the image complexity.

[0026] In one implementation, the quantitative scoring according to the trademark image to obtain the image complexity includes: According to the trademark image, calculate the image texture complexity score through the gray-level co-occurrence matrix; According to the trademark image, calculate the color richness score by using the color histogram method; According to the trademark image, calculate the shape diversity score by using the contour analysis method; According to the texture complexity score, the color richness score and the shape diversity score, the image complexity is obtained by weighting.

[0027] It should be noted that the gray-level co-occurrence matrix calculates the texture complexity score of the image by analyzing the spatial relationship of the gray levels of the trademark image, so as to extract the texture features of the image. By calculating the co-occurrence frequency of different gray-level pairs in the image, the texture detail complexity of the image can be quantified. This method can effectively reflect the texture patterns existing in the image, thus providing a basis for the complexity evaluation of the image.

[0028] It should be noted that the color richness score is evaluated by analyzing the color histogram of the trademark image. The color histogram can reflect the saturation and distribution uniformity of the image color by statistically analyzing the frequency distribution of different colors in the image. A trademark with rich colors contains more color levels, while a trademark with a single color shows a lower color richness. This score helps to measure the color complexity of the trademark image visually.

[0029] It should be noted that the shape diversity score is calculated by the contour analysis method, mainly analyzing the shape change of the objects in the trademark image. Contour analysis involves the detection of the image edge, shape geometric features and structure. By extracting the edge information of the image, the complexity and diversity of the shape are quantified. The level of shape diversity reflects the creativity and complexity of the trademark design, which helps to distinguish the morphological features between different trademarks.

[0030] It should be noted that the system first calculates its texture complexity score, with a full score of 100 points. Color analysis shows that 15 different hues are used, and the color richness score is 90 points. Shape analysis finds 5 main geometric elements, and the shape diversity score is 75 points. Combining these dimensions, the system gives an overall complexity score of 83 points. If the preset threshold is 80 points, the logo is determined to be a high-complexity image, triggering the detailed hierarchical feature extraction process. The significance of this process lies in more precisely describing complex images and laying a foundation for subsequent image retrieval and comparison. First, the extraction algorithm based on texture features uses the local binary pattern (LBP) to capture local texture information. This method can effectively describe the subtle texture changes in the image, such as the brushstroke effect or texture filling in the logo. The color feature extraction uses the color moment algorithm, which not only statistically analyzes the color distribution but also considers the spatial information.

[0031] In step S13, according to the trademark image and the image complexity, wavelet transform is performed to obtain a feature vector, and dynamic adjustment is performed according to the feature vector to obtain a feature representation.

[0032] In one implementation manner, the performing wavelet transform according to the trademark image and the image complexity to obtain a feature vector, and performing dynamic adjustment according to the feature vector to obtain a feature representation includes: According to the image complexity, determine whether it exceeds a preset complexity threshold. If it exceeds the complexity threshold, determine it as a complex trademark image; if it does not exceed the complexity threshold, skip the following steps; Perform wavelet transform on the complex trademark image, obtain image information in different frequency bands, and extract features from the image information in different frequency bands to obtain feature vectors; By calculating the background interference degree and visual redundancy degree of the complex trademark image, dynamically adjust the weights of the feature vectors to obtain a weight adjustment result; According to the weight adjustment result, perform feature fusion to obtain a fine-grained feature representation.

[0033] It should be noted that wavelet transform is a powerful image processing tool that can decompose complex trademark images into information of different frequency bands. This multi-scale decomposition can effectively separate the details and overall structure in the image. For example, for a luxury brand logo containing fine textures, colorful hues, and complex shapes, the low-frequency band contains the overall outline and main color blocks, the middle-frequency band contains color transitions and secondary graphic elements, while the high-frequency band contains fine lines and texture details. When extracting texture features in the low-frequency band, the gray-level co-occurrence matrix is used to capture the texture information of the image. For the above-mentioned luxury brand trademark, the gray-level co-occurrence matrix will show high contrast and homogeneity. The color feature extraction in the middle-frequency band can adopt the color moment algorithm, which considers not only the color distribution but also the spatial information. The shape features in the high-frequency band can be extracted through edge detection and contour analysis, which will identify the geometric elements in the logo, such as circles, triangles, or irregular curves. The calculation of background interference degree and visual redundancy is crucial for improving the accuracy of feature extraction. For example, if the logo background is a solid color, the background interference degree will be low, and at this time, the weight of the features in the background area can be reduced. On the contrary, if the background contains complex patterns, its weight needs to be increased to prevent misjudgment. The calculation of visual redundancy can help identify repeated elements, such as the repeated use of the brand name, so as to avoid overemphasizing these repeated information. Feature fusion is a key step in combining different features into a comprehensive image representation. For the luxury brand logo, a high-dimensional feature vector will be obtained, where the texture features account for 40%, the color features account for 35%, and the shape features account for 25%. This weight distribution reflects the importance of texture and color in the logo design. Classification algorithms, such as support vector machines (SVM), can classify the logo into specific categories, such as "luxury", "fashion", or "high-end" according to these features. This classification helps to quickly locate logos with similar styles in a large-scale trademark database. Finally, clustering algorithms such as K-means can group similar logos. Through this series of steps, complex trademark images can be comprehensively and accurately described and analyzed, providing strong support for trademark registration, infringement detection, and design innovation. The advantage of this method is that it can adapt to images of different complexities and provide a fine-grained feature representation, thus achieving more accurate image comparison and retrieval.

[0034] In step S14, according to the feature representation, it is input into a preset semantic association model to obtain candidate association information.

[0035] In one implementation manner, the inputting according to the feature representation into a preset semantic association model to obtain candidate association information includes: According to the feature representation, it is input into a preset semantic association model, and the features are mapped to a high-dimensional semantic space through a fully connected layer to generate an image semantic vector; Obtain a set of brand semantic vectors according to a pre-trained brand dictionary; Calculate the cosine similarity between the image semantic vector and the brand semantic vector; According to the cosine similarity and a preset similarity threshold, filter out brand semantic vectors higher than the similarity threshold to obtain a candidate brand set; Perform clustering analysis on the brand semantic vectors in the candidate brand set, divide brand association categories, and obtain candidate association information according to the brand association categories.

[0036] It should be noted that the construction of the semantic association model is usually based on natural language processing and image understanding technologies in deep learning. This model transforms the trademark image into a vector representation with high-level semantic meaning by fusing the visual features of the image and the semantic information of the brand. During the construction process, a convolutional neural network is used to map these features to a high-dimensional semantic space through a fully connected layer or other deep network structures, thereby generating the semantic vector of the image. At the same time, a set of brand semantic vectors is obtained through a pre-trained brand dictionary, and these semantic vectors represent the concepts and characteristics of the brand.

[0037] It should be noted that the role of the semantic association model is to match the visual features of the trademark image with the semantic information of the brand, ensuring that the retrieval not only depends on the appearance of the image but also can understand the deep semantic relationship between the image and the brand. By calculating the cosine similarity between the image semantic vector and the brand semantic vector, the model can accurately evaluate the association degree between the trademark image and different brands. Finally, through filtering and clustering analysis, the model can classify the trademark and the most relevant brands into the same category, thereby providing a high-precision candidate brand set and accurate association information.

[0038] It should be noted that the core role of knowledge graph reasoning technology in brand association analysis is reflected in the following aspects: First, by constructing the mapping relationship between brand nodes and image semantic vectors, the system can accurately evaluate the association degree between images and brands based on calculation methods such as cosine similarity. For example, by calculating the similarity between a sports shoe image and the "Nike" brand node, the system can intuitively conclude that the image has a high association with "Nike" and a low association with "Gucci", thus providing a clear basis for subsequent brand retrieval and analysis. Second, setting a similarity threshold (such as 0.7) ensures high-confidence brand nodes while effectively filtering out low-correlation brands, thereby improving the precision rate. This setting can be dynamically adjusted according to different application scenarios to achieve the best retrieval effect. Third, the application of clustering algorithms enables multiple high-confidence brand nodes to be grouped according to similarity, revealing potential association patterns between brands. For example, the system can divide sports brands into two categories, "outdoor sports" and "indoor fitness", through the K-means algorithm, providing a more refined perspective for brand analysis and helping to understand the multi-faceted nature of image content.

[0039] It should be noted that generating a mapping relationship table between brand association categories and image semantic vectors makes the association between images and brand categories more intuitive and multi-dimensional. For an image showing a mountain bike, the system may have a high-confidence association with multiple "outdoor sports" brands and further classify these brands, thus providing support for the market positioning of the brands. Finally, the continuous update of the knowledge graph and the accumulation of association relationships enable the system to gradually optimize the accuracy and comprehensiveness of brand association analysis as the amount of data increases, improving the precision in future analysis. For example, with the processing of more skiing images, the system can discover and establish associations between "skiing" and concepts such as "thermal clothing" and "safety equipment", providing deeper insights for brand analysis.

[0040] In step S15, according to the candidate association information, using knowledge graph reasoning technology, calculate the similarity between the image semantic vector and the brand node, and filter out high-confidence association results by setting a similarity threshold.

[0041] In one implementation, the step of calculating the similarity between the image semantic vector and the brand node according to the candidate association information, using knowledge graph reasoning technology, and filtering out high-confidence association results by setting a similarity threshold includes: According to the candidate association information, using knowledge graph reasoning technology, obtain brand node information from the candidate brand set; Calculate the cosine similarity between the image semantic vector and the brand node information to generate a similarity matrix; According to a preset similarity threshold, filter the brand nodes in the similarity matrix that are higher than the threshold to determine the brand association results with high confidence; According to the brand association results with high confidence, if there are multiple brand nodes with high confidence, use a clustering algorithm to group the brand nodes; According to the grouping results of the brand nodes after clustering, generate a mapping relationship table between brand association categories and image semantic vectors; According to the mapping relationship table, output the association results between the image semantic vectors and the brand nodes; Store the association results in the knowledge graph, update the association relationship between the brand nodes and the image semantic vectors, and obtain the association results with high confidence.

[0042] It should be noted that the combination of image processing technology and cross-modal retrieval methods plays an important role in trademark image analysis and brand association. First, by using image processing technology to extract features such as the color, texture, and shape of trademark images, the image can be converted into specific feature vectors, which lays a foundation for subsequent text generation and brand recognition. For example, for a trademark that is red and round, the system can extract two key features, "red" and "round", and generate relevant text labels based on these features, such as "simple", providing textual support for the semantic understanding of the image. Second, entity linking technology matches the generated text labels with actual brands in the brand knowledge base, further strengthening the association between image features and brand entities. For example, the system can establish connections between labels such as "red" and "round" and the trademarks of well-known brands to ensure an effective mapping from abstract image features to specific brands.

[0043] It should be noted that establishing a semantic link network is a key link in this technical framework. By connecting images and brand nodes with each other, rich brand association information can be constructed. Through the optimization of graph neural networks, the system can deeply learn the potential relationship between images and brands, enhancing the accuracy and depth of brand recognition. For example, the system can discover that certain color combinations are more likely to appear in specific industries, providing valuable insights for brand strategies. Cross-modal retrieval technology enables users to search for relevant brands by inputting images or retrieve similar trademark images by brand names. This two-way retrieval method greatly improves the efficiency of brand management, competitor analysis, and market research.

[0044] In step S16, according to the association results, dynamically generate text labels for trademark images, and use entity linking technology to map the text labels to entities in the brand knowledge base to obtain mapping results.

[0045] In one implementation, based on the association result, dynamically generate a text label for the trademark image, and use entity linking technology to map the text label to an entity in the brand knowledge base to obtain a mapping result, including: Dynamically generate a text label for the trademark image according to the association result; Match the text label with a pre-established brand knowledge base, and use entity linking technology to determine the brand entity corresponding to the text label; Establish a mapping relationship between the image features and the brand entity according to the brand entity; According to the mapping relationship, construct a semantic link network between the image and the brand. Through cross-modal retrieval technology, according to the semantic link network, map the text label to an entity in the brand knowledge base to obtain a mapping result.

[0046] It should be noted that "dynamically generating a text label for the trademark image and using entity linking technology to map the text label to an entity in the brand knowledge base" in step S16 realizes a close association between the image and the brand entity. First, by analyzing the features of the image, relevant text labels are generated, and these labels reflect the visual content of the trademark image, such as color, shape, pattern and other information. Then, through entity linking technology, the system compares these labels with the pre-constructed brand knowledge base to determine the corresponding brand entity. For example, for a trademark containing a red circular pattern, the system may map it to a well-known brand with a red circular logo through the labels "red" and "circular".

[0047] It is worth noting that the system establishes a mapping relationship between the image features and the brand entity, and connects the image node and the brand node through a semantic link network to form richer brand association information. This network structure not only improves the association accuracy between the image and the brand, but also provides strong support for subsequent cross-modal retrieval. Using this semantic link network, users can accurately retrieve brand information related to it by inputting a trademark image or a text label, thereby optimizing brand management and market analysis.

[0048] In step S17, input the mapping result into a preset semantic link network to obtain a trademark retrieval result.

[0049] In one implementation, the inputting the mapping result into a preset semantic link network to obtain a trademark retrieval result includes: Arrange the brand information candidate set in descending order according to the mapping result to generate a preliminary retrieval result; According to the preliminary retrieval result, perform data cleaning, delete null values and outliers to obtain a processed retrieval result; According to the processed retrieval results, combined with the user's historical retrieval records, personalized adjustment is carried out through the collaborative filtering algorithm to obtain the final retrieval results.

[0050] It should be noted that the trademark retrieval process in step S17 improves the accuracy of the retrieval results and the personalized experience through multi-stage processing. First, the mapping results are input into a preset semantic link network, and the system sorts the brand information candidate set in descending order based on the mapping relationship to generate preliminary retrieval results. This process ensures a high degree of correlation between the retrieved brands and the trademark images. Then, the system performs data cleaning on the preliminary retrieval results, removing null values and outliers, eliminating invalid information that may interfere with the retrieval results, and ensuring the reliability and accuracy of the retrieval results.

[0051] It is worth noting that in the further processing process, the system combines the user's historical retrieval records and performs personalized adjustment through the collaborative filtering algorithm. This algorithm dynamically optimizes the retrieval results according to the user's historical behavior and preferences, provides more accurate brand recommendations, and improves the user experience. This personalized adjustment not only improves the relevance of the retrieval results but also enhances the system's adaptability to different user needs. Finally, through this series of steps, users can obtain trademark retrieval results that better meet their needs, thereby improving the efficiency of brand management, market analysis, and other aspects.

[0052] In summary, the present invention discloses a trademark image retrieval method based on deep learning, including obtaining a trademark image; performing quantization scoring according to the trademark image to obtain the image complexity; performing wavelet transform according to the trademark image and the image complexity to obtain a feature vector, and dynamically adjusting according to the feature vector to obtain a feature representation; inputting the feature representation into a preset semantic association model to obtain candidate association information; according to the candidate association information, using knowledge graph reasoning technology to calculate the similarity between the image semantic vector and the brand node, and screening out high-confidence association results by setting a similarity threshold; dynamically generating a text label for the trademark image according to the association results, and using entity linking technology to map the text label to an entity in the brand knowledge base to obtain a mapping result; inputting the mapping result into a preset semantic link network to obtain a trademark retrieval result.

[0053] The trademark image retrieval method based on deep learning proposed by the present invention significantly improves the accuracy and efficiency of trademark image retrieval through multi-level information processing and intelligent reasoning techniques. First, the complexity of the image is obtained by quantitative scoring, which helps to refine the differentiation of different types of trademarks and avoids the misdetection problem caused by ignoring details due to complexity in traditional methods. Then, wavelet transform is used to extract the feature vectors of the image and perform dynamic adjustment, which not only enhances the feature expression ability of the trademark image but also enables the model to adapt to the visual differences of different trademarks. Further, through the preset semantic association model and knowledge graph reasoning technology, candidate association information can be accurately obtained, and by calculating the similarity between the image semantic vector and the brand node, the most relevant trademark images are screened out. This combined method based on semantic vectors and knowledge graphs effectively reduces the noise interference in trademark image retrieval, thereby improving the relevance of the retrieval results. In addition, the entity linking technology is used to accurately map the trademark image to the entities in the brand knowledge base, which not only ensures the accuracy of the retrieval results but also makes the final retrieval results more interpretable and practical.

[0054] Referring to Figure 2 , the second embodiment of the present invention provides a trademark image retrieval system based on deep learning, including: A trademark acquisition module for acquiring trademark images; A complexity calculation module for performing quantitative scoring on the basis of the trademark image to obtain the image complexity; A feature extraction module for performing wavelet transform on the basis of the trademark image and the image complexity to obtain feature vectors, and performing dynamic adjustment on the basis of the feature vectors to obtain feature representations; An association information acquisition module for inputting the feature representation into a preset semantic association model to obtain candidate association information; A similarity matching module for calculating the similarity between the image semantic vector and the brand node by using the knowledge graph reasoning technology on the basis of the candidate association information, and screening out high-confidence association results by setting a similarity threshold; An entity mapping module for dynamically generating text labels of the trademark image on the basis of the association results, and using the entity linking technology to map the text labels to the entities in the brand knowledge base to obtain mapping results; A trademark retrieval module for inputting the mapping results into a preset semantic link network to obtain trademark retrieval results.

[0055] It should be noted that the trademark image retrieval system based on deep learning provided by the embodiments of the present invention is used to execute all the process steps of the trademark image retrieval method based on deep learning in the above embodiments, and the working principles and beneficial effects of the two correspond one by one, so they will not be elaborated here.

[0056] An embodiment of the present invention further provides an electronic device. The electronic device includes: a processor, a memory, and a computer program stored in the memory and executable on the processor, such as a trademark acquisition program. When the processor executes the computer program, the steps in the above-mentioned embodiments of each deep learning-based trademark image retrieval method are implemented, such as Figure 1 step S11 shown. Alternatively, when the processor executes the computer program, the functions of each module / unit in the above-mentioned system embodiments are implemented, such as the trademark acquisition module.

[0057] Exemplarily, the computer program may be divided into one or more modules / units, and the one or more modules / units are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing specific functions, and the instruction segments are used to describe the execution process of the computer program in the electronic device.

[0058] The electronic device may be a computing device such as a desktop computer, a notebook, a palm computer, and a smart tablet. The electronic device may include, but is not limited to, a processor and a memory. Those skilled in the art can understand that the above components are only examples of the electronic device and do not constitute a limitation on the electronic device. It may include more or fewer components than the above, or combine certain components, or different components. For example, the electronic device may further include input / output devices, network access devices, a bus, etc.

[0059] The so-called processor may be a central processing unit (CPU), or may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. The processor is the control center of the electronic device and connects various parts of the entire electronic device through various interfaces and lines.

[0060] The memory can be used to store the computer programs and / or modules. By running or executing the computer programs and / or modules stored in the memory and invoking the data stored in the memory, the processor realizes various functions of the electronic device. The memory mainly includes a program storage area and a data storage area. Among them, the program storage area can store an operating system, application programs required for at least one function (such as a sound playback function, an image playback function, etc.); the data storage area can store data created according to the use of the mobile phone (such as audio data, phone book, etc.). In addition, the memory can include high-speed random access memory and can also include non-volatile memory, such as a hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one magnetic disk storage device, flash memory device, or other volatile solid-state storage devices.

[0061] Among them, if the modules / units integrated in the electronic device are implemented in the form of software function units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, to implement all or part of the processes in the above-mentioned embodiment methods of the present invention, it can also be completed by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by the processor, the steps of the above-mentioned various method embodiments can be realized. Among them, the computer program includes computer program code, and the computer program code can be in the form of source code, object code, executable file, or some intermediate form, etc. The computer-readable medium can include: any entity or system, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disc, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc. It should be noted that the content included in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the computer-readable medium does not include electrical carrier signals and telecommunication signals.

[0062] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. In addition, in the attached drawings of the system embodiments provided by the present invention, the connection relationship between modules indicates that they have a communication connection, which can be specifically implemented as one or more communication buses or signal lines. Those of ordinary skill in the art can understand and implement it without creative efforts.

[0063] The specific embodiments described above have further elaborated on the purpose, technical solutions, and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not used to limit the protection scope of the present invention. In particular, it is pointed out that for those skilled in the art, any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A trademark image retrieval method based on deep learning, characterized in that: Executed by a computer, including: Obtain trademark images; Performing quantitative scoring based on the trademark image to obtain image complexity; According to the trademark image and the image complexity, wavelet transform is performed to obtain a feature vector, and dynamic adjustment is performed according to the feature vector to obtain a feature representation; According to the feature representation, input into a preset semantic association model to obtain candidate association information; Based on the candidate association information, the similarity between the image semantic vector and the brand node is calculated using knowledge graph reasoning technology, and high-confidence association results are screened out by setting a similarity threshold; According to the association result, dynamically generate a text label of the trademark image, and use entity linking technology to map the text label with an entity in the brand knowledge base to obtain a mapping result; The mapping result is input into a preset semantic link network to obtain a trademark search result.

2. The trademark image retrieval method based on deep learning according to claim 1, characterized in that: The step of performing quantitative scoring according to the trademark image to obtain image complexity includes: According to the trademark image, calculating the image texture complexity score by using a gray level co-occurrence matrix; Calculating a color richness score using a color histogram method based on the trademark image; Calculating a shape diversity score using a contour analysis method based on the trademark image; The image complexity is obtained by weighting the texture complexity score, the color richness score and the shape diversity score.

3. The trademark image retrieval method based on deep learning according to claim 1, characterized in that: The step of performing wavelet transform according to the trademark image and the image complexity to obtain a feature vector, and dynamically adjusting according to the feature vector to obtain a feature representation includes: According to the complexity of the image, determine whether it exceeds a preset complexity threshold, if it exceeds the complexity threshold, determine it as a complex trademark image, if it does not exceed the complexity threshold, skip the following steps; Performing wavelet transformation on the complex trademark image to obtain image information of different frequency bands, and performing feature extraction on the image information of different frequency bands to obtain feature vectors; By calculating the background interference and visual redundancy of the complex trademark image, the weight of the feature vector is dynamically adjusted to obtain a weight adjustment result; According to the weight adjustment result, feature fusion is performed to obtain a fine-grained feature representation.

4. The trademark image retrieval method based on deep learning according to claim 1, characterized in that: The step of inputting the feature representation into a preset semantic association model to obtain candidate association information includes: According to the feature representation, the feature is input into a preset semantic association model, and the feature is mapped to a high-dimensional semantic space through a fully connected layer to generate an image semantic vector; According to the pre-trained brand dictionary, obtain a set of brand semantic vectors; Calculating the cosine similarity between the image semantic vector and the brand semantic vector; According to the cosine similarity and a preset similarity threshold, brand semantic vectors having a value higher than the similarity threshold are screened out to obtain a candidate brand set; Cluster analysis is performed on the brand semantic vectors in the candidate brand set to divide brand association categories, and candidate association information is obtained based on the brand association categories.

5. The trademark image retrieval method based on deep learning according to claim 1, characterized in that: The method of calculating the similarity between the image semantic vector and the brand node based on the candidate association information and screening out association results with high confidence by setting a similarity threshold comprises: According to the candidate association information, using knowledge graph reasoning technology, brand node information is obtained from the candidate brand set; Calculating the cosine similarity between the image semantic vector and the brand node information to generate a similarity matrix; According to a preset similarity threshold, brand nodes in the similarity matrix with a degree higher than the threshold are screened to determine high-confidence brand association results; According to the high-confidence brand association result, if there are multiple high-confidence brand nodes, a clustering algorithm is used to group the brand nodes; According to the clustered brand node grouping results, a mapping relationship table between brand association categories and image semantic vectors is generated; According to the mapping relationship table, output the association result between the image semantic vector and the brand node; The association results are stored in the knowledge graph, and the association relationship between the brand node and the image semantic vector is updated to obtain a high-confidence association result.

6. The trademark image retrieval method based on deep learning according to claim 1, characterized in that: The text label of the trademark image is dynamically generated according to the association result, and the text label is mapped with the entity in the brand knowledge base by using entity linking technology to obtain a mapping result, including: dynamically generating a text label for the trademark image according to the association result; Matching the text label with a pre-established brand knowledge base, and using entity linking technology to determine the brand entity corresponding to the text label; According to the brand entity, a mapping relationship between the image feature and the brand entity is established; According to the mapping relationship, a semantic link network between the image and the brand is constructed, and through cross-modal retrieval technology, according to the semantic link network, the text label is mapped with the entity in the brand knowledge base to obtain a mapping result.

7. The trademark image retrieval method based on deep learning according to claim 1, characterized in that: The step of inputting the mapping result into a preset semantic link network to obtain a trademark search result includes: Arrange the brand information candidate set in descending order according to the mapping result to generate a preliminary search result; According to the preliminary search results, data cleaning is performed to delete null values ​​and abnormal values ​​to obtain processed search results; According to the processed search results, combined with the user's historical search records, personalized adjustments are made through collaborative filtering algorithms to obtain the final search results.

8. A trademark image retrieval system based on deep learning, characterized in that: include: A trademark acquisition module, used to acquire a trademark image; A complexity calculation module, used to perform quantitative scoring based on the trademark image to obtain image complexity; A feature extraction module, used to perform wavelet transformation according to the trademark image and the image complexity to obtain a feature vector, and to perform dynamic adjustment according to the feature vector to obtain a feature representation; A correlation information acquisition module, used to input the feature representation into a preset semantic correlation model to obtain candidate correlation information; A similarity matching module is used to calculate the similarity between the image semantic vector and the brand node based on the candidate association information and to filter out association results with high confidence by setting a similarity threshold; An entity mapping module, used to dynamically generate a text label of the trademark image according to the association result, and use entity linking technology to map the text label with an entity in the brand knowledge base to obtain a mapping result; The trademark search module is used to input the mapping result into a preset semantic link network to obtain a trademark search result.

9. An electronic device, characterized in that: The invention comprises a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, the trademark image retrieval method based on deep learning as described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located is controlled to execute the trademark image retrieval method based on deep learning as described in any one of claims 1 to 7.

Images