Graph training set determination method and device, electronic equipment, product and storage medium

Abstract

The invention provides a graph training set determination method and device, electronic equipment, a product and a storage medium, and relates to the technical field of image processing. The method comprises the following steps of: superposing associated elements corresponding to initial graphs in an initial graph set onto the corresponding initial graphs, and determining an extended graph set; determining a target graph set based on the initial graph set and the extended graph set; trigger words, visual semantics and metaphor semantics corresponding to the target graphs in the target graph set are determined as graph tags of the target graphs, and a graph tag set is obtained; and based on the target graph set and the graph label set, determining a graph training set. According to the method, the defect that common features are difficult to extract during model training due to inconsistent visual features of abstract graphs caused by a training set determination method used in the prior art can be solved, and the consistency of the visual features of the abstract graphs can be ensured so as to obtain a high-quality graph training set.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a method, device, electronic device, product and storage medium for determining a graphic training set. Background Art

[0002] In the field of AI-generated content, producing high-quality training sets is the basis for ensuring the effectiveness and accuracy of the model. In terms of training data collection, existing technologies mainly rely on obtaining graphic data from public resources and expanding these data through data enhancement techniques (such as cropping, rotation, flipping, and brightness adjustment) to form a rich training set; however, this training set determination method has significant problems that cannot be ignored. Expanding data through data enhancement technology often leads to inconsistent visual features of abstract graphics, such as differences in line thickness, color proportion, and feature elements, making it difficult for the model to extract common features during training, resulting in a gap between the images generated after the model training is completed and the expected effect. Therefore, there is an urgent need for an image training set determination method that can obtain high-quality training sets to ensure the consistency of abstract graphics in visual features, thereby improving the model's ability to extract common features. Summary of the invention

[0003] The present invention provides a method, device, electronic device, product and storage medium for determining a graphic training set, so as to solve the defect that the training set determination method used in the prior art leads to inconsistent visual features of abstract graphics, thereby making it difficult to extract common features during model training.

[0004] The present invention provides a method for determining a graphic training set, comprising: Overlaying the associated elements corresponding to each initial graphic in the initial graphic set onto the corresponding initial graphic to determine an extended graphic set; Determine a target graphics set based on the initial graphics set and the extended graphics set; Determine the trigger words, visual semantics and metaphorical semantics corresponding to each target graphic in the target graphic set as graphic labels of each target graphic to obtain a graphic label set; A graphic training set is determined based on the target graphic set and the graphic label set.

[0005] According to a method for determining a graphic training set provided by the present invention, the associated elements corresponding to any of the initial graphics are determined based on the following method: Extracting features of the initial graphic to obtain a first visual feature vector; Determine first similarity results between the first visual feature vector and feature vectors of each scene element in a preset scene element data set; The scene element corresponding to the first similarity result having the greatest similarity among the first similarity results is determined as the associated element corresponding to the initial graphic.

[0006] According to a method for determining a graphic training set provided by the present invention, the method of superimposing the associated elements corresponding to each initial graphic in the initial graphic set onto the corresponding initial graphic to determine the extended graphic set includes: Based on a preset layout rule, determining position information and proportion information of each of the associated elements on the corresponding initial graphic; Based on the position information and proportion information of each of the associated elements on the corresponding initial graphic, each of the associated elements is superimposed on the corresponding initial graphic to determine an extended graphic set.

[0007] According to a method for determining a graphic training set provided by the present invention, the visual semantics corresponding to any of the target graphics is determined based on the following method: Extracting the features of the target graphic to obtain a second visual feature vector; Determine a second similarity result between the second visual feature vector and a feature vector of each preset graphic in a preset graphic information set; the preset graphic information set includes a plurality of preset graphics and label information corresponding to each preset graphic; Determine a preset graphic corresponding to a second similarity result having the greatest similarity among the second similarity results as a similar graphic; The label information corresponding to the similar graphics is determined as the visual semantics of the target graphics.

[0008] According to a method for determining a graphic training set provided by the present invention, the metaphorical semantics and trigger words corresponding to any of the target graphics are determined based on the following method: Based on the visual semantics corresponding to the target graphic, a target text matching the visual semantics is retrieved from a preset text information set; Counting the frequency of each word in the target text, and taking the word with the highest frequency as the metaphorical semantics corresponding to the target graphic; Based on preset word extraction rules, the visual semantics corresponding to the target graphic and the metaphorical semantics corresponding to the target graphic are extracted to determine the trigger word corresponding to the target graphic.

[0009] According to a method for determining a graphic training set provided by the present invention, the graphic training set includes the target graphic set and the graphic label set. After determining the graphic training set, the method further includes: The graphic training set is iteratively updated at least once until a tuning training set that meets preset requirements is obtained; wherein any iterative update in the at least one iterative update includes: Determine the associated elements corresponding to each target graphic in the target graphic set, and superimpose the associated elements corresponding to each target graphic into the corresponding target graphic to obtain an extended target graphic set; Determine the expanded target graphics set and the target graphics set as the iteratively updated tuning target graphics set; Determine the trigger words, visual semantics and metaphorical semantics corresponding to each tuning target graphic in the tuning target graphic set as graphic labels of each tuning target graphic to obtain a tuning graphic label set; A tuning training set is determined based on the tuning target graph set and the tuning graph label set.

[0010] The present invention also provides a device for determining a graphic training set, comprising: An associated element superposition module, which is used to superimpose the associated elements corresponding to each initial graphic in the initial graphic set onto the corresponding initial graphic to determine an extended graphic set; A target graphics set determination module, configured to determine a target graphics set based on the initial graphics set and the extended graphics set; A graphic label determination module, which is used to determine the trigger words, visual semantics and metaphorical semantics corresponding to each target graphic in the target graphic set as the graphic label of each target graphic to obtain a graphic label set; The training set determination module is used to determine a graphic training set based on the target graphic set and the graphic label set.

[0011] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, any of the above-mentioned methods for determining a graphic training set is implemented.

[0012] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the method for determining a graphic training set as described in any one of the above is implemented.

[0013] The present invention also provides a computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the method for determining a graphic training set as described above is implemented.

[0014] The method, device, electronic device, product and storage medium for determining a graphic training set provided by the present invention superimpose the associated elements corresponding to each initial graphic in the initial graphic set onto the corresponding initial graphic to determine an extended graphic set; determine a target graphic set based on the initial graphic set and the extended graphic set; determine the trigger words, visual semantics and metaphorical semantics corresponding to each target graphic in the target graphic set as graphic labels of each target graphic to obtain a graphic label set; determine a graphic training set based on the target graphic set and the graphic label set; the present invention forms an extended graphic set by superimposing each initial graphic in the initial graphic set with its corresponding associated elements, thereby enriching the diversity of training data and Ensure that the extended graphics are consistent with the corresponding initial graphics in terms of visual features and semantic expression; secondly, for each target graphic in the target graphics set, extract the trigger words, visual semantics and metaphorical semantics and determine the graphic label for it. This process is achieved through semantic analysis and feature extraction algorithms to ensure that the graphic labels accurately reflect the multiple meanings of the target graphics, thereby enhancing the model's understanding and generation capabilities; construct a graphic training set based on the target graphics set and the graphic label set, which greatly improves the model's learning ability and generation effect for abstract graphics, and effectively solves the defect of the existing technology that the model training is difficult due to the inconsistency of the visual features of abstract graphics, thereby enhancing the model's generalization and generation capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0016] Figure 1 It is a flow chart of the method for determining a graphic training set provided by the present invention.

[0017] Figure 2 It is a schematic diagram of an initial graphic set of the method for determining a graphic training set provided by the present invention.

[0018] Figure 3 It is a schematic diagram of an extended graphic set of the method for determining a graphic training set provided by the present invention.

[0019] Figure 4 It is a schematic diagram of superimposed associated elements of the method for determining a graphic training set provided by the present invention.

[0020] Figure 5 It is a structural schematic diagram of the device for determining a graphic training set provided by the present invention.

[0021] Figure 6It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are 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.

[0023] During the in-depth study of the preparation of the graphic training set, the applicant gradually realized that the existing training set determination method has significant limitations, especially when dealing with abstract graphics, the inconsistency of visual features is particularly prominent. After in-depth analysis, the applicant found that the existing technology usually collects graphic data from public resources to form a training set, and uses data enhancement methods (such as cropping, flipping, rotating, adjusting brightness and contrast) to expand the collected graphics. This practice leads to large differences in the visual features of abstract graphics, such as line thickness, color proportion and diversity of feature elements, which makes it difficult for the model to extract common features during the training process, and there is a large gap between the final generated image and the expected effect.

[0024] In further research, the applicant found that although the existing technology attempts to solve this problem through data enhancement and standardization, there are still significant deficiencies. For example, although the data enhancement method can increase the sample size, it may cause the important semantics of abstract graphics to be blurred or lost. Especially when dealing with highly abstract graphics, traditional data enhancement methods cannot effectively improve the model's understanding ability, but may make the model's learning of graphic features more complicated. In addition, existing annotation methods often rely on the method of "algorithmic automatic annotation combined with manual modification", and they are also unable to extract the intuitive visual meaning and metaphorical meaning of abstract graphics, resulting in a lack of systematicity and consistency in the annotation of the training set.

[0025] In view of the above problems, the present invention proposes the following embodiments.

[0026] Figure 1 is a flow chart of the method for determining a graphic training set provided by the present invention, such as Figure 1 As shown, the method includes the following: Step 110 , superimposing the associated elements corresponding to each initial graphic in the initial graphic set onto the corresponding initial graphic to determine an extended graphic set.

[0027] It should be noted that each initial graphic in the initial graphic set refers to the most basic image, which is used as the basis for subsequent processing. The initial graphic is usually a visual object with independent meaning. The extended graphic set is a collection of graphics formed by superimposing the initial graphic and its corresponding associated elements. Each graphic in this set not only retains the basic features of the initial graphic, but also incorporates the semantic information of the associated elements, making the expression of the graphic more diverse.

[0028] Here, the associated elements refer to additional graphics that are related to the initial graphics to some extent, and can provide more contextual information for the initial graphics, thereby enriching the expression of the graphics. For example, if the initial graphics is a "car", its associated elements may be "road", "traffic lights" or "pedestrians". These associated elements can enrich the semantic level of the initial graphics and make them more representative in specific scenes.

[0029] For example, superimposing related elements onto the initial graphic can inject more contextual information into the graphic, thereby improving the visual and semantic expressiveness of the graphic. For example, a simple "car" graphic can better convey the meaning of a "traffic scene" after superimposing a "traffic light". In addition, the extended graphic set provides more diverse and complex inputs for subsequent model training, which helps the model extract more representative features and patterns during the learning process, thereby improving the accuracy and generalization ability of the model.

[0030] For the specific method of determining the associated elements, please refer to the following embodiments, which will not be elaborated here.

[0031] Step 120: determining a target graphics set based on the initial graphics set and the extended graphics set.

[0032] Here, the target graph set is a new graph set formed by integrating the initial graph set and the extended graph set, which can effectively support the training and testing process of the model.

[0033] like Figure 2 and Figure 3 As shown, Figure 2 shows the initial set of graphics, Figure 3 Shows an extended graphics set.

[0034] The advantage of doing this is that the target graphics set brings together the advantages of the initial graphics and the extended graphics, ensuring the richness and diversity of the graphics content, and providing a wider range of samples for subsequent model training; in addition, the diversity of the target graphics set and the high-quality graphics input can significantly improve the learning effect of the deep learning model in specific tasks, thereby improving the accuracy and robustness of the model; through the optimized target graphics set, the model can better capture the common features between graphics during the training process, promote the accuracy of feature extraction, and make the training results closer to expectations.

[0035] Step 130 , determining the trigger words, visual semantics and metaphorical semantics corresponding to each target graphic in the target graphic set as graphic labels of each target graphic to obtain a graphic label set.

[0036] Here, a trigger word is a keyword or phrase for a specific target graphic, which is designed to activate the model's understanding of the graphic. The trigger word is usually unique and does not directly overlap with known words to reduce the model's dependence on existing graphic styles. Visual semantics refers to the semantic information extracted by analyzing the appearance characteristics of the target graphic, which includes a combination of elements such as color, shape, and structure, reflecting the intuitive characteristics of the graphic. Metaphorical semantics refers to the semantics derived by analyzing the deep meaning and symbolic meaning conveyed by the graphic. This semantics usually goes beyond the surface characteristics of the graphic and explores in depth the meaning represented by the graphic in a specific cultural, emotional or social context. Metaphorical semantics is closely related to the role, composition, and application scenarios of the graphic.

[0037] For example, metaphorical semantics can reveal the function and purpose of a graphic in a specific context. For example, a graphic depicting a dove of peace is more than just an image of a bird. Metaphorical semantics may reflect deeper social values ​​such as "peace", "hope" and "unity".

[0038] For example, visual semantics refers to extracting semantic information related to a target graphic by analyzing its appearance features. For example, if the graphic is in the shape of a gear, then its visual semantics is "gear".

[0039] In one embodiment, the graphic label of each target graphic is composed of trigger words, visual semantics and metaphorical semantics; further, the graphic labels of all target graphics are integrated to obtain a graphic label set. Through the combination of trigger words, visual semantics and metaphorical semantics, the graphics can be described from multiple dimensions. This comprehensive information expression not only improves the richness of the labels, but also more accurately reflects the meaning of the graphics; the comprehensive use of trigger words, visual semantics and metaphorical semantics helps to eliminate the ambiguity that may arise from different graphics and ensure that the model can understand the exact meaning of each graphic.

[0040] The specific process of determining trigger words, visual semantics, and metaphorical semantics can be referred to the following embodiments, which will not be elaborated here.

[0041] Step 140: Determine a graphic training set based on the target graphic set and the graphic label set.

[0042] It should be noted that by integrating the target graphics set and the graphics label set, a high-quality graphics training set can be constructed, covering diverse graphics and rich label information; this training set will be used to train the model so that it can learn how to classify and identify based on graphics features.

[0043] The method for determining a graphic training set provided by an embodiment of the present invention forms an extended graphic set by superimposing each initial graphic in the initial graphic set with its corresponding associated elements, thereby enriching the diversity of training data and ensuring that the extended graphic is consistent with the corresponding initial graphic in terms of visual features and semantic expression; secondly, for each target graphic in the target graphic set, trigger words, visual semantics and metaphorical semantics are extracted and graphic labels are determined for it. This process is achieved through semantic analysis and feature extraction algorithms, ensuring that the graphic labels accurately reflect the multiple meanings of the target graphics, thereby enhancing the model's understanding and generation capabilities; constructing a graphic training set based on the target graphic set and the graphic label set greatly improves the model's learning ability and generation effect for abstract graphics, effectively solving the defect of the prior art that model training is difficult due to inconsistent visual features of abstract graphics, thereby enhancing the model's generalization and generation capabilities.

[0044] Based on any of the above embodiments, in this method, the associated element corresponding to any of the initial graphics is determined based on the following method: Extracting features of the initial graphic to obtain a first visual feature vector; Determine first similarity results between the first visual feature vector and feature vectors of each scene element in a preset scene element data set; The scene element corresponding to the first similarity result having the greatest similarity among the first similarity results is determined as the associated element corresponding to the initial graphic.

[0045] Here, the first visual feature vector refers to feature data extracted from the initial graphic, expressed in vector form, and capable of quantifying the appearance features of the graphic. The first similarity result refers to the similarity metric obtained during the calculation process, which is used to determine the degree of association between the initial graphic and the scene element.

[0046] It should be noted that the preset scene element dataset is a collection of various elements related to specific themes or concepts. These elements can be combined with the initial graphics in graphic design, data visualization or model training. The dataset aims to provide context and rich visual information for the graphics, helping to generate more attractive and expressive images. Specifically, the preset scene element dataset can significantly enhance the visual effect of the graphics by providing a variety of related elements. For example, if the initial graphic is a car pattern, the preset scene elements may include background city streets, traffic lights or other vehicles, which together construct a vivid scene.

[0047] In one embodiment, visual features are extracted from the initial graphic. These features are represented in the form of vectors and usually include information such as shape, color, and texture. These features can effectively characterize the appearance and structure of the graphic. Further, the extracted first visual feature vector is compared with the feature vector of each scene element in the preset scene element data set to calculate the similarity between them. The similarity calculation can be performed in a variety of ways, including but not limited to cosine similarity, Euclidean distance, etc., with the aim of finding the distance or angle of the feature vectors in a high-dimensional space. Further, by comparing the similarity results of all scene elements, the element with the greatest similarity is found. This element is the associated element of the initial graphic, which means that it is most similar to the initial graphic in visual features and has a higher correlation.

[0048] Optionally, in another embodiment, by comparing the similarity results of all scene elements, several elements with the highest similarity ranking are found as associated elements of the initial graphic, and further, these associated elements are simultaneously superimposed on the initial graphic to determine the extended graphic.

[0049] The method for determining a graphic training set provided by an embodiment of the present invention can provide richer contextual information for the initial graphic by determining associated elements, thereby enhancing the model's ability to understand and recognize the graphic during training; secondly, with the help of similarity analysis, the model can focus on elements with similar features, improve the expressiveness of abstract graphics, and ensure the semantic consistency and diversity of the generated graphics; in addition, through this precise method for determining associated elements, the quality of training data can be improved, thereby improving the generalization ability of the final model; therefore, this process plays a key role in the construction of the training set and lays a solid foundation for subsequent model training.

[0050] Based on any of the above embodiments, in the method, superimposing the associated elements corresponding to each initial graphic in the initial graphic set onto the corresponding initial graphic to determine the extended graphic set includes: Based on a preset layout rule, determining position information and proportion information of each of the associated elements on the corresponding initial graphic; Based on the position information and proportion information of each of the associated elements on the corresponding initial graphic, each of the associated elements is superimposed on the corresponding initial graphic to determine an extended graphic set.

[0051] It should be noted that the preset layout rules refer to a series of design principles followed when designing extended graphics, including the symmetry of elements, visual balance, color matching, etc. These rules can ensure that the generated graphics are visually beautiful and can effectively convey information. Position information refers to the specific coordinate position of the associated elements in the initial graphics, which determines the spatial relationship between the elements and the initial graphics and affects the overall visual effect. Proportional information refers to the description of the size of the associated elements, which is reasonably arranged relative to the size of the initial graphics to ensure the visibility of the elements and the integrity of the initial graphics.

[0052] Exemplarily, the position information, proportion information, area information, etc. of the associated elements in the initial graphic may be determined based on the preset layout rules.

[0053] like Figure 4 As shown, Figure 4 Shows the process of overlaying associated elements onto the initial graphic. Figure 4 The example shows the process of forming an extended graphic by superimposing traffic lights, wipers, lights, grilles and other related elements on the basis of the initial "car" graphic.

[0054] In one embodiment, the overlay operation refers to the process of placing associated elements onto the initial graphic according to determined position information and scale information to form an extended graphic, which is intended to increase the complexity and expressiveness of the graphic, thereby obtaining rich graphic data.

[0055] The method for determining a graphic training set provided in an embodiment of the present invention clarifies the position information and proportion information of related elements on the initial graphic through preset layout rules. This process ensures the reasonable matching and visual balance of each element and avoids visual clutter and incoordination; further, these related elements are accurately superimposed on the initial graphic, which not only enriches the content of the graphic and makes it more vivid and eye-catching, but also improves the expressiveness of the graphic and can better convey the design intent and theme; this method effectively overcomes the common problems of graphic simplification and insufficient expression in traditional methods, promotes the diversity and complexity of the training set, and provides a more comprehensive and consistent data basis for subsequent model training.

[0056] Based on any of the above embodiments, in the method, the visual semantics corresponding to any of the target graphics is determined based on the following method: Extracting the features of the target graphic to obtain a second visual feature vector; Determine a second similarity result between the second visual feature vector and a feature vector of each preset graphic in a preset graphic information set; the preset graphic information set includes a plurality of preset graphics and label information corresponding to each preset graphic; Determine a preset graphic corresponding to a second similarity result having the greatest similarity among the second similarity results as a similar graphic; The label information corresponding to the similar graphics is determined as the visual semantics of the target graphics.

[0057] Here, the second visual feature vector is a feature representation extracted from the target graphic, and contains important visual information related to the graphic. The second similarity result is a quantified result of the similarity between the target graphic and the preset graphic in the preset graphic information set, and is used to judge the similarity between the two.

[0058] It should be noted that the preset graphic information set is a data set containing multiple predefined graphics and their corresponding label information. It is designed to provide the model with rich reference data to support the recognition, classification and semantic understanding of graphics. Specifically, the preset graphic information set contains multiple preset graphics and the label information corresponding to each preset graphic. Each preset graphic corresponds to a label information that describes the meaning or characteristics of the graphic. These labels can be text descriptions, keywords or other forms of identification to help the model understand the concepts conveyed by the graphics. For example, for a preset graphic representing a gear, the label information may be "gear".

[0059] In one embodiment, among all the second similarity results, the result with the greatest similarity is found, that is, the similar graphic is determined. This enables a reference object closest to the target graphic to be selected from the preset graphics, so as to ensure that the visual semantics of the target graphic is more accurate and practical.

[0060] The method for determining a graphic training set provided in an embodiment of the present invention extracts the features of a target graphic and generates a second visual feature vector, which can quantify and express the visual information of the graphic; further, by calculating the similarity between the second visual feature vector and the feature vectors of each preset graphic in the preset graphic information set, it is possible to find a sample closest to the target graphic, thereby improving the accuracy of recognition. Finally, by using the label information corresponding to similar graphics as the visual semantics of the target graphic, not only is the semantic understanding of the target graphic enhanced, but also a clear and standardized semantic basis is provided for subsequent graphic processing and application. The beneficial effect of this process is that, through a systematic approach, the recognition accuracy and semantic parsing capability of the target graphic are significantly improved, providing more reliable data support for subsequent applications.

[0061] Based on any of the above embodiments, in this method, the metaphorical semantics and trigger words corresponding to any of the target graphics are determined based on the following method: Based on the visual semantics corresponding to the target graphic, a target text matching the visual semantics is retrieved from a preset text information set; Counting the frequency of each word in the target text, and taking the word with the highest frequency as the metaphorical semantics corresponding to the target graphic; Based on preset word extraction rules, the visual semantics corresponding to the target graphic and the metaphorical semantics corresponding to the target graphic are extracted to determine the trigger word corresponding to the target graphic.

[0062] It should be noted that the preset text information set refers to a collection of various text information that is related to a specific topic, emotion or semantics. The preset text information set provides rich background data for the extraction of metaphorical semantics, enabling more accurate identification of metaphorical expressions related to the target graphic.

[0063] In addition, the preset word extraction rules are a series of standardized methods for identifying and extracting key words from visual semantics and metaphorical semantics. Through these rules, text data can be systematically processed to screen out the most representative and effective words, ensuring the accuracy and relevance of the extraction results.

[0064] In one embodiment, based on the visual semantics corresponding to the target graphic, a target text matching the target graphic is retrieved from a preset text information set. This process can obtain text information related to the target graphic, thereby laying a foundation for subsequent metaphorical semantic extraction. Further, after obtaining the target text, the frequency of each word in the text is counted. This statistical analysis can help identify which words are most closely related to the visual semantics of the target graphic. The words with the highest frequency of occurrence are determined as the metaphorical semantics of the target graphic, which can reflect the deep semantic meaning and association of the target graphic. Further, based on the preset word extraction rules, the visual semantics and metaphorical semantics of the target graphic are extracted to determine the trigger word of the target graphic. The trigger word in this process mainly plays the role of quickly stimulating or guiding the user to associate with specific concepts. The words that best represent the graphic are extracted from the visual semantics and metaphorical semantics through the preset word extraction rules as trigger words.

[0065] For example, for the entire training set, as the number of target graphics in the training set increases, the number of high-frequency words extracted also increases, ensuring that each graphic has a unique and representative metaphorical meaning label. The labeling strategy based on vocabulary frequency can ensure that each target graphic in the training set has corresponding metaphorical semantics and trigger words, forming effective annotation data, thereby improving the model's understanding and recognition ability of graphics.

[0066] Exemplarily, the visual semantics are extracted from the target graphic, which is usually achieved by analyzing the features of the graphic, such as shape, color, constituent elements, etc.; these visual semantics are used to retrieve the target text matching it in a preset text information set.

[0067] The method for determining a graphic training set provided by an embodiment of the present invention effectively combines the graphic with relevant language information by extracting the visual semantics of the target graphic and retrieving the matching target text in a preset text information set; by counting the frequency of words in the target text, the most representative metaphorical semantics can be identified. This process not only enhances the semantic depth of the target graphic, but also provides richer context; in addition, the trigger words extracted using the preset word extraction rules provide important guidance for the user's interaction with the graphic.

[0068] Based on any of the above embodiments, in the method, the graphic training set includes the target graphic set and the graphic label set. After determining the graphic training set, the method further includes: The graphic training set is iteratively updated at least once until a tuning training set that meets preset requirements is obtained; wherein any iterative update in the at least one iterative update includes: Determine the associated elements corresponding to each target graphic in the target graphic set, and superimpose the associated elements corresponding to each target graphic into the corresponding target graphic to obtain an extended target graphic set; Determine the expanded target graphics set and the target graphics set as the iteratively updated tuning target graphics set; Determine the trigger words, visual semantics and metaphorical semantics corresponding to each tuning target graphic in the tuning target graphic set as graphic labels of each tuning target graphic to obtain a tuning graphic label set; A tuning training set is determined based on the tuning target graph set and the tuning graph label set.

[0069] Here, preset requirements refer to specific target standards for the quality of the training set or model performance. These standards are usually pre-set during model training and application to measure whether the training effect meets the usage requirements.

[0070] Here, the extended target graph set refers to a new set of graphs generated by superimposing associated elements on the original target graph. The tuned training set is a training dataset that has been iteratively optimized and expanded, and contains more diverse graphs and their label information.

[0071] In one embodiment, the associated elements corresponding to each target graphic in the target graphic set are determined, and the associated elements are superimposed on the original target graphic to form an extended target graphic set, so that a large number of new graphic combinations can be generated, thereby increasing the diversity and quantity of training data; further, the trigger words, visual semantics and metaphorical semantics are re-labeled for the extended graphics, thereby ensuring that the generated new graphics also have clear label information; the original target graphic set and the extended target graphic set are combined to generate a tuned target graphic set, and the tuned target graphic set and the tuned graphic label set are combined to determine the tuned training set; so as to obtain more training data for the model to use.

[0072] Exemplarily, the 20 collected initial graphics are used as the first training set, which is represented by the number "20*1"; next, the training set is expanded based on the "unit increment" method: the second training set consists of 20 initial graphics and 20 extended graphics after superimposing the associated elements once, totaling 40 graphics, numbered "20*2"; the third training set contains 20 initial graphics and 40 extended graphics after superimposing the associated elements twice, totaling 60 graphics, numbered "20*3"; and so on, the nth training set contains 20 initial graphics and extended graphics obtained after superimposing the associated elements n times, numbered "20*n". The training set is expanded to "20*20", that is, it contains 400 graphics, including 20 initial graphics and extended graphics after superimposing the associated elements 20 times. These training sets are used for training, and each trained model is qualitatively evaluated. The evaluation criteria include goodness of fit, graphic quality and generalization ability. The training set with the highest comprehensive score is the optimal training set.

[0073] The method for determining a graphics training set provided in an embodiment of the present invention can continuously optimize the training data by gradually iterating and updating the graphics training set, so that the training effect of the model meets the expected requirements. First, for the target graphics set, the corresponding associated elements are superimposed to generate an extended target graphics set, which increases the richness and diversity of the training data. Then, these extended graphics and the original target graphics are used together as the tuning target graphics set, and more accurate graphic labels are generated by matching trigger words, visual semantics, and metaphorical semantics for each graphic, thereby further improving the quality of the training data. Finally, through repeated iterations, it is ensured that the training set meets the predetermined standards after multiple rounds of updates, so that the model performs well in multiple dimensions such as fit, graphics quality, and generalization ability, meeting application requirements.

[0074] The graphic training set determination device provided by the present invention is described below. The graphic training set determination device described below and the graphic training set determination method described above can refer to each other.

[0075] FIG5 is a schematic diagram of the structure of a device for determining a graphic training set provided by the present invention. As shown in FIG5 , the device for determining a graphic training set includes: The associated element superposition module 510 is used to superimpose the associated elements corresponding to each initial graphic in the initial graphic set onto the corresponding initial graphic to determine the extended graphic set; A target graphics set determining module 520, configured to determine a target graphics set based on the initial graphics set and the extended graphics set; A graphic label determination module 530, which is used to determine the trigger words, visual semantics and metaphorical semantics corresponding to each target graphic in the target graphic set as the graphic label of each target graphic to obtain a graphic label set; The training set determination module 540 is used to determine a graphic training set based on the target graphic set and the graphic label set.

[0076] Figure 6 An example of a physical structure diagram of an electronic device is shown in FIG. Figure 6 As shown, the electronic device may include: a processor 610, a communication interface 620, a memory 630 and a communication bus 640, wherein the processor 610, the communication interface 620 and the memory 630 communicate with each other through the communication bus 640. The processor 610 may call the logic instructions in the memory 630 to execute the method for determining a graphic training set, which includes: superimposing the associated elements corresponding to each initial graphic in the initial graphic set onto the corresponding initial graphic to determine an extended graphic set; determining a target graphic set based on the initial graphic set and the extended graphic set; determining the trigger words, visual semantics and metaphorical semantics corresponding to each target graphic in the target graphic set as graphic labels of each target graphic to obtain a graphic label set; determining a graphic training set based on the target graphic set and the graphic label set.

[0077] In addition, the logic instructions in the above-mentioned memory 630 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0078] On the other hand, the present invention also provides a computer program product, which includes a computer program, and the computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the graphic training set determination method provided by the above-mentioned methods, and the method includes: superimposing the associated elements corresponding to each initial graphic in the initial graphic set onto the corresponding initial graphic to determine an extended graphic set; determining a target graphic set based on the initial graphic set and the extended graphic set; determining the trigger words, visual semantics and metaphorical semantics corresponding to each target graphic in the target graphic set as graphic labels of each target graphic to obtain a graphic label set; determining a graphic training set based on the target graphic set and the graphic label set.

[0079] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the graphic training set determination method provided by the above-mentioned methods, the method comprising: superimposing the associated elements corresponding to each initial graphic in the initial graphic set onto the corresponding initial graphic to determine an extended graphic set; determining a target graphic set based on the initial graphic set and the extended graphic set; determining the trigger words, visual semantics and metaphorical semantics corresponding to each target graphic in the target graphic set as graphic labels of each target graphic to obtain a graphic label set; determining a graphic training set based on the target graphic set and the graphic label set.

[0080] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0081] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0082] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for determining a graphic training set, characterized in that: include: Overlaying the associated elements corresponding to each initial graphic in the initial graphic set onto the corresponding initial graphic to determine an extended graphic set; Determine a target graphics set based on the initial graphics set and the extended graphics set; Determine the trigger words, visual semantics and metaphorical semantics corresponding to each target graphic in the target graphic set as graphic labels of each target graphic to obtain a graphic label set; A graphic training set is determined based on the target graphic set and the graphic label set.

2. The method for determining a graphic training set according to claim 1, characterized in that: The associated element corresponding to any of the initial graphics is determined based on the following method: Extracting features of the initial graphic to obtain a first visual feature vector; Determine first similarity results between the first visual feature vector and feature vectors of each scene element in a preset scene element data set; The scene element corresponding to the first similarity result having the greatest similarity among the first similarity results is determined as the associated element corresponding to the initial graphic.

3. The method for determining a graphic training set according to claim 1, characterized in that: The step of superimposing the associated elements corresponding to each initial graphic in the initial graphic set onto the corresponding initial graphic to determine the extended graphic set includes: Based on a preset layout rule, determining position information and proportion information of each of the associated elements on the corresponding initial graphic; Based on the position information and proportion information of each of the associated elements on the corresponding initial graphic, each of the associated elements is superimposed on the corresponding initial graphic to determine an extended graphic set.

4. The method for determining a graphic training set according to claim 1, characterized in that: The visual semantics corresponding to any of the target graphics is determined based on the following method: Extracting the features of the target graphic to obtain a second visual feature vector; Determine a second similarity result between the second visual feature vector and a feature vector of each preset graphic in a preset graphic information set; the preset graphic information set includes a plurality of preset graphics and label information corresponding to each preset graphic; Determine a preset graphic corresponding to a second similarity result having the greatest similarity among the second similarity results as a similar graphic; The label information corresponding to the similar graphics is determined as the visual semantics of the target graphics.

5. The method for determining a graphic training set according to claim 4, characterized in that: The metaphorical semantics and trigger words corresponding to any of the target graphics are determined based on the following method: Based on the visual semantics corresponding to the target graphic, a target text matching the visual semantics is retrieved from a preset text information set; Counting the frequency of each word in the target text, and taking the word with the highest frequency as the metaphorical semantics corresponding to the target graphic; Based on preset word extraction rules, the visual semantics corresponding to the target graphic and the metaphorical semantics corresponding to the target graphic are extracted to determine the trigger word corresponding to the target graphic.

6. The method for determining a graphic training set according to claim 1, characterized in that: The graphic training set includes the target graphic set and the graphic label set. After determining the graphic training set, the method further includes: The graphic training set is iteratively updated at least once until a tuning training set that meets preset requirements is obtained; wherein any iterative update in the at least one iterative update includes: Determine the associated elements corresponding to each target graphic in the target graphic set, and superimpose the associated elements corresponding to each target graphic into the corresponding target graphic to obtain an extended target graphic set; Determine the expanded target graphics set and the target graphics set as the iteratively updated tuning target graphics set; Determine the trigger words, visual semantics and metaphorical semantics corresponding to each tuning target graphic in the tuning target graphic set as graphic labels of each tuning target graphic to obtain a tuning graphic label set; A tuning training set is determined based on the tuning target graph set and the tuning graph label set.

7. A device for determining a graphic training set, characterized in that: include: An associated element superposition module, which is used to superimpose the associated elements corresponding to each initial graphic in the initial graphic set onto the corresponding initial graphic to determine an extended graphic set; A target graphics set determination module, configured to determine a target graphics set based on the initial graphics set and the extended graphics set; A graphic label determination module, which is used to determine the trigger words, visual semantics and metaphorical semantics corresponding to each target graphic in the target graphic set as the graphic label of each target graphic to obtain a graphic label set; The training set determination module is used to determine a graphic training set based on the target graphic set and the graphic label set.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the method for determining a graphic training set according to any one of claims 1 to 6 is implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for determining a graphic training set according to any one of claims 1 to 6 is implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the method for determining a graphic training set according to any one of claims 1 to 6 is implemented.

Images