Method and device for constructing data set of systematized color SVG (Scalable Vector Graphics) icons

Abstract

The invention discloses a systematized color SVG icon data set construction method and device, and the method comprises the following steps: collecting an SVG file, and carrying out the data cleaning of the collected SVG file, and obtaining an initial file; wherein the SVG file comprises a plurality of SVG icons of different types, and the initial file is a color SVG file; standardizing the size of a bounding box of the initial file to obtain a sample file; and classifying the sample files, and dividing a training set, a verification set and a test set according to the classified sample files to obtain an SVG data set. According to the method, the construction quality of the color SVG data set can be improved.

Description

Technical Field

[0001] The present invention relates to the field of computer graphics, and in particular to a method and device for constructing a data set of a systematic color SVG icon. Background Art

[0002] There are currently some SVG-related datasets, such as the SVG-Icons8 dataset, the FIGR-8-SVG dataset, and the Icon645 dataset. The following is a detailed description of these datasets and their construction process: (1) SVG-Icons8 dataset: During the compilation process, the authors ensured the consistency and diversity of SVG by ensuring that SVG graphics have similar scales, colors, and styles while capturing different real-world graphics. In the end, the dataset consists of 100,000 SVGs from 56 different categories.

[0003] (2) FIGR-8-SVG dataset: The SVG in this dataset is a black and white representation of an object, concept, pattern, or design, created by designers and artists and compiled into a dataset. There are 1,548,944 SVGs in the dataset, which are divided into 18,409 conceptually different categories, each of which contains at least 8 SVGs and up to several thousand.

[0004] (3) Icon645 dataset: This dataset was retrieved and collected from the Flaticon website to construct an icon dataset containing 377 categories and 645,687 color icons. During the dataset construction process, the author cropped the white space of each icon graphic in the dataset to make it more compact, filtered out black and white icons, and removed redundant instances based on graphic similarity.

[0005] Although the above datasets have made contributions in their respective fields, there are still some problems and shortcomings: (1) Limitations of black-and-white SVG: The SVG-Icons8 and FIGR-8-SVG datasets are mainly composed of black-and-white SVG and lack color SVG, which is insufficient when processing application scenarios that require rich color information.

[0006] (2) Format issues of raster icons: Although the Icon645 dataset contains color icons, these icons are stored in raster image format rather than SVG format. This limits the practical application value of the dataset in applications that require vector graphics.

[0007] (3) Insufficient data cleaning and standardization: Although some datasets have undergone preliminary data cleaning and organization, there are still deficiencies. For example, the sizes of icons in the Icon645 dataset are not uniform, ranging from the smallest 64×64 to the largest 256×256, which leads to problems with the accuracy and consistency of the dataset. At the same time, the collected data may be biased, that is, the category corresponding to the icon is not its true category. Summary of the invention

[0008] In order to overcome the above-mentioned defects of the prior art, the present invention provides a method and device for constructing a data set of a systematic color SVG icon, which can improve the construction quality of the color SVG data set.

[0009] An embodiment of the present invention provides a method for constructing a systematic color SVG icon dataset, comprising the following steps: Collecting SVG files, and performing data cleaning on the collected SVG files to obtain an initial file; wherein the SVG file includes a plurality of SVG icons of different categories, and the initial file is a color SVG file; Standardizing the size of the bounding box of the initial file to obtain a sample file; The sample files are classified, and a training set, a validation set and a test set are divided according to the classified sample files to obtain an SVG data set.

[0010] Furthermore, the collecting of SVG files specifically includes: The icon categories in the Icon645 dataset were filtered and expanded to obtain several collection categories; SVG icons of different collection categories are collected from the open source platform respectively, and corresponding category labels are added to all the SVG icons to obtain SVG files.

[0011] Furthermore, the data cleaning of the collected SVG file to obtain the initial file specifically includes: Confirming the attributes of the SVG files respectively, removing the SVG files that lack the preset fill attribute, and rasterizing the remaining SVG files respectively to obtain corresponding bitmap images; Determine the pixel value of each pixel in the bitmap image respectively, and calculate the proportion of pixels with a pixel value of 0 in all pixels in the bitmap image; The bitmap images whose corresponding proportion is 0 and whose corresponding proportion exceeds a preset ratio threshold are removed, and the SVG files corresponding to the remaining bitmap images are determined as the initial files.

[0012] Furthermore, the step of normalizing the size of the bounding box of the initial file to obtain a sample file specifically includes: The bounding box sizes of all the initial files are uniformly adjusted to a preset border threshold to obtain the sample file.

[0013] Preferably, before the initial file is standardized, the method further comprises: In the initial file <g>The attributes of the tag are transferred to the <g>After the child node of the tag, remove the <g>Label.

[0014] Furthermore, the classifying of the sample files specifically includes: Rasterizing the sample file to obtain a raster image; Inputting all the raster images into an image encoder of a preset CLIP model to obtain image features, and inputting all the category labels into a text encoder of a preset CLIP model to obtain text features; For each of the image features, respectively calculating the similarity between the image feature and each text feature, and determining the category label corresponding to the text feature with the highest similarity as the correct category of the sample file corresponding to the image feature; Counting the number of raster images in each correct category, and taking the first a correct categories with the highest number of corresponding raster images as the data set category; wherein a is a preset ranking threshold; According to the data set category, all the raster images are reclassified through a preset CLIP model, and the confidence between the reclassified raster images and the corresponding categories is calculated, and the raster images with confidence lower than a preset confidence threshold are removed to complete the classification.

[0015] Furthermore, the method of dividing the training set, the validation set and the test set according to the classified sample files to obtain the SVG data set specifically includes: According to the preset number of training sets, number of validation sets and number of test sets, the classified sample files are divided by stratified sampling to obtain the SVG data set.

[0016] Another embodiment of the present invention provides a data set construction device for a systematic color SVG icon, including: a collection module, a standardization module, and a classification module; The acquisition module is used to acquire SVG files and perform data cleaning on the acquired SVG files to obtain an initial file; wherein the SVG file includes a plurality of SVG icons of different categories, and the initial file is a color SVG file; The standardization module is used to standardize the size of the bounding box of the initial file to obtain a sample file; The classification module is used to classify the sample files, and divide the sample files into a training set, a validation set and a test set according to the classified sample files to obtain an SVG data set.

[0017] Furthermore, the acquisition module is used to acquire SVG files, specifically including: The icon categories in the Icon645 dataset were filtered and expanded to obtain several collection categories; SVG icons of different collection categories are collected from the open source platform respectively, and corresponding category labels are added to all the SVG icons to obtain SVG files.

[0018] Furthermore, the standardization module is used to standardize the size of the bounding box of the initial file to obtain a sample file, which specifically includes: The bounding box sizes of all the initial files are uniformly adjusted to a preset border threshold to obtain the sample file.

[0019] Compared with the prior art, the beneficial effects of the present invention are: Through rigorous data cleaning and standardization steps, the quality and consistency of SVG files are improved, making the dataset more reliable in practical applications. At the same time, the data cleaning step also ensures that the final dataset contains rich color SVG graphics, supports the development and training of complex algorithms, enhances the capabilities of graphic design software, and improves data visualization technology.

[0020] In summary, through the technical solution provided by this paper, a high-quality, rich and diverse color SVG dataset can be provided to researchers and developers, promoting technological progress and innovation in related fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 A schematic flow chart of a method for constructing a systematic color SVG icon dataset according to an embodiment of the present invention.

[0022] Figure 2 A statistical diagram of the number of samples corresponding to each category in a training set of a ColorSVG-100K dataset provided by an embodiment of the present invention.

[0023] Figure 3 A statistical diagram of the average number of paths corresponding to each category in a training set of a ColorSVG-100K dataset provided by an embodiment of the present invention.

[0024] Figure 4 A schematic diagram of some samples of a ColorSVG-100K dataset provided by an embodiment of the present invention.

[0025] Figure 5 A schematic diagram of the structure of a device for constructing a systematic color SVG icon data set provided by another embodiment of the present invention. DETAILED DESCRIPTION

[0026] The drawings are for illustrative purposes only and should not be construed as limiting the present patent; It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0027] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0028] Reference Figure 1 , is a flow chart of a method for constructing a systematic color SVG icon dataset according to an embodiment of the present invention, comprising the following steps: S1: collecting SVG files, and performing data cleaning on the collected SVG files to obtain an initial file; wherein the SVG file includes a plurality of SVG icons of different categories, and the initial file is a color SVG file; S2: Standardizing the size of the bounding box of the initial file to obtain a sample file; S3: Classify the sample files, and divide the sample files into a training set, a validation set, and a test set according to the classified sample files to obtain an SVG data set.

[0029] For step S1, specifically, collecting the SVG file includes: The icon categories in the Icon645 dataset were filtered and expanded to obtain several collection categories; SVG icons of different collection categories are collected from the open source platform respectively, and corresponding category labels are added to all the SVG icons to obtain SVG files.

[0030] In a preferred embodiment, the first step in building a color SVG dataset is data collection. This process involves obtaining a large number of SVG files from various online repositories and open source platforms, with the goal of covering graphics of different styles, themes, and complexity. Ensuring diversity is critical to creating a comprehensive dataset that can support a wide range of research and application needs.

[0031] On this basis, the preferred embodiment further screens and expands the categories in the Icon645 dataset to identify common categories as the collection categories. Subsequently, SVG files are collected and downloaded from the Internet according to the collection categories. At this stage, the collected SVG files are relatively extensive, relatively primitive, and of varying quality.

[0032] For step S1, further, the data cleaning of the collected SVG file to obtain the initial file specifically includes: Confirming the attributes of the SVG files respectively, removing the SVG files that lack the preset fill attribute, and rasterizing the remaining SVG files respectively to obtain corresponding bitmap images; Determine the pixel value of each pixel in the bitmap image respectively, and calculate the proportion of pixels with a pixel value of 0 in all pixels in the bitmap image; The bitmap images whose corresponding proportion is 0 and whose corresponding proportion exceeds a preset ratio threshold are removed, and the SVG files corresponding to the remaining bitmap images are determined as the initial files.

[0033] In a preferred embodiment, after data collection is completed, the dataset undergoes a rigorous cleaning process to ensure its quality and consistency. This step is critical because it removes any damaged, incomplete, or low-quality SVG files that may negatively affect algorithm performance. The initially collected data often contains duplicates and colorless SVGs, which is contrary to the core goal of the final required SVG dataset. Therefore, these duplicates and colorless SVGs, as well as any damaged or incomplete files, need to be removed first to ensure the reliability of the dataset in subsequent processing and analysis. The specific process of data cleaning is as follows: First, check whether each SVG file lacks the "fill" attribute (i.e., the preset fill attribute). If the SVG file does not contain the "fill" attribute, it can be preliminarily determined to be an SVG containing only black. Such files are usually rendered in the default black color, so they will be removed. In this process, this preferred embodiment temporarily ignores other possible filling methods.

[0034] Next, the SVG file is rasterized, converted into a bitmap image, and checked to see if the image's colors are only black (value 0) and white (value 255). If this condition is met, the SVG is deleted.

[0035] In addition, to improve the quality of the dataset, it is also necessary to remove SVG files where black is dominant. To do this, the ratio of black pixels to other colored pixels (excluding white background) needs to be evaluated. If the proportion of black pixels exceeds the proportion of other colored pixels (excluding white background), the corresponding SVG will also be removed.

[0036] Thus, the data cleaning step is completed and the initial file is obtained. Such a cleaning process is crucial to improving the overall quality and consistency of the data set.

[0037] For step S2, specifically, the step of normalizing the bounding box size of the initial file to obtain a sample file specifically includes: The bounding box sizes of all the initial files are uniformly adjusted to a preset border threshold to obtain the sample file.

[0038] In a preferred embodiment, data normalization is an important step to ensure the consistency of the dataset. The bounding box (Bbox) is used to define the spatial extent of graphics in SVG, providing key information for positioning and scaling in various applications. However, the collected SVG bounding boxes are inconsistent in size and position, which may increase the complexity of subsequent processing and analysis.

[0039] To solve this problem, the preferred embodiment standardizes all bounding boxes to a size of 100×100 units. This process involves adjusting the size and position of the SVG to fit the standardized bounding box. Through this standardization operation, it is possible to ensure that all SVGs follow a unified spatial framework, making them more reliable and consistent in different research and application contexts. This not only simplifies the processing and manipulation of graphics, but also facilitates more accurate algorithm training.

[0040] For step S2, preferably, before the initial file is standardized, the process further includes: In the initial file <g>The attributes of the tag are transferred to the <g>After the child node of the tag, remove the <g>Label.

[0041] In a preferred embodiment, <g>Tags are used in SVG to group multiple elements together, but this increases the complexity of handling and manipulating individual graphic components. <g>Tags usually carry the influence of <path>Element attributes, so you can't simply delete these tags directly, but you need to <g>The attributes in the tag are transferred to the child nodes and then removed.

[0042] In addition, the preferred embodiment also uses the svglib library provided by DeepSVG and further optimizes it to remove the original SVG <g>The label part finally generates a flat path combination. <g>Tags can simplify the structure of SVG, making it easier to access and process in various computing tasks while preserving the integrity of the graphics.

[0043] For step S3, specifically, the classifying of the sample files specifically includes: Rasterizing the sample file to obtain a raster image; Inputting all the raster images into an image encoder of a preset CLIP model to obtain image features, and inputting all the category labels into a text encoder of a preset CLIP model to obtain text features; For each of the image features, respectively calculating the similarity between the image feature and each text feature, and determining the category label corresponding to the text feature with the highest similarity as the correct category of the sample file corresponding to the image feature; Counting the number of raster images in each correct category, and taking the first a correct categories with the highest number of corresponding raster images as the data set category; wherein a is a preset ranking threshold; According to the data set category, all the raster images are reclassified through a preset CLIP model, and the confidence between the reclassified raster images and the corresponding categories is calculated, and the raster images with confidence lower than a preset confidence threshold are removed to complete the classification.

[0044] In a preferred embodiment, the initially collected SVG files may be misclassified, i.e., the graphics do not match the categories assigned to them. Therefore, in order to ensure accurate classification of the SVG, reclassification is necessary.

[0045] First, we rasterize the SVG to obtain the raster images and use the CLIP model to assist in reclassification. These raster images are input to the CLIP image encoder, while the category labels are input to the CLIP text encoder. The correct category of each SVG is determined by calculating the similarity between the generated features and taking the highest scoring match as the output label.

[0046] Then, the number of SVGs in each category is counted, and these categories are arranged in descending order, and finally the top 500 categories with the largest number are selected as the classification standard of the data set, that is, the data set category.

[0047] Finally, based on the 500 dataset categories, all the raster images are classified, and those SVGs with low confidence scores in the model classification and inconsistent with their original categories are removed from the dataset. At the same time, the same method is used to correct SVGs with high confidence scores but incorrect classification. Finally, in this preferred embodiment, the total number of SVGs obtained after classification is 100,000.

[0048] For step S3, further, dividing the training set, the validation set and the test set according to the classified sample files to obtain the SVG data set specifically includes: According to the preset number of training sets, number of validation sets and number of test sets, the classified sample files are divided by stratified sampling to obtain the SVG data set.

[0049] In a preferred embodiment, the last step of constructing the dataset is to divide it into a training set, a validation set, and a test set. Considering that the SVG or image generation model may take a long time in the generation process, 8,000 samples are finally allocated to the validation set and 2,000 samples are allocated to the test set.

[0050] In order to ensure a balanced distribution of categories and visual features, the preferred embodiment uses a stratified sampling method to divide the original 100,000 SVG data sets. This method ensures that each subset can accurately represent the diversity of the overall data set.

[0051] Therefore, the final training set contains 90,000 SVGs, the validation set contains 8,000 SVGs, and the test set contains 2,000 SVGs.

[0052] Through the above-mentioned method for constructing a dataset of systematic color SVG icons, the embodiment of the present invention constructs the ColorSVG-100K dataset. The following is a statistical analysis of the ColorSVG-100K dataset: First, in the training set, the SVG samples were classified according to their respective categories, and the number of samples in each category was counted. The results are arranged in descending order, and the intermediate results are omitted for simplicity, such as Figure 2 As shown. Figure 2 It can be seen that the category with the largest number of samples contains up to 475 instances, while the category with the smallest number of samples has about 40 instances. This imbalance in the dataset stems from the uneven distribution of different SVG categories in online resources, with more common categories and fewer rare categories.

[0053] In addition, the average number of paths per category in the training set was analyzed to assess the complexity of different categories. The results of this analysis are arranged in descending order, and the intermediate results are omitted for brevity, e.g. Figure 3 As shown. Figure 3 As you can see, the category with the highest average number of paths is "Basket", followed by "Lion", indicating that these categories have more complex designs and contain more lines. Therefore, the complexity is higher. In contrast, the categories with the lowest average number of paths are "Arrow" and "Bookmark", indicating that these SVGs have lower complexity.

[0054] The numerical statistics of different dimensions (categories) in each subset of the SVG dataset are shown in the following table:

[0055] Finally, refer to Figure 4 , is a schematic diagram of some samples of a ColorSVG-100K dataset provided by an embodiment of the present invention. Figure 4 20 categories in the dataset are randomly selected, and 3 SVG samples are randomly selected from each category. By observing these samples, it can be seen that the method for constructing a dataset of systematic color SVG icons provided by the present invention effectively ensures the diversity of the dataset, and the quality of SVG is high, which fully demonstrates the superior effect of the method.

[0056] In summary, the embodiment of the present invention significantly improves the quality and practicality of the SVG dataset through a systematic construction method. The specific functions and effects include: (1) Improvement of dataset quality: Through rigorous data cleaning and standardization steps, the quality and consistency of SVG files are improved, making the dataset more reliable in practical applications.

[0057] (2) Improved data processing efficiency: The simplified SVG structure and unified bounding box standardization make data processing and computational model training more efficient.

[0058] (3) Enhanced category accuracy: The reclassification method of the CLIP model ensures accurate classification of SVG and reduces category errors.

[0059] (4) Improved practicality of the dataset: It covers rich color SVG graphics, supports the development and training of complex algorithms, enhances the capabilities of graphic design software, and improves data visualization technology.

[0060] Reference Figure 5 , is a schematic diagram of the structure of a systematic color SVG icon data set construction device provided by another embodiment of the present invention, comprising: a collection module 101, a standardization module 102 and a classification module 103; The acquisition module 101 is used to acquire SVG files and perform data cleaning on the acquired SVG files to obtain an initial file; wherein the SVG file includes a plurality of SVG icons of different categories, and the initial file is a color SVG file; The standardization module 102 is used to standardize the size of the bounding box of the initial file to obtain a sample file; The classification module 103 is used to classify the sample files, and divide the sample files into a training set, a validation set and a test set according to the classified sample files to obtain an SVG data set.

[0061] Furthermore, the acquisition module 101 is used to acquire SVG files, specifically including: The icon categories in the Icon645 dataset were filtered and expanded to obtain several collection categories; SVG icons of different collection categories are collected from the open source platform respectively, and corresponding category labels are added to all the SVG icons to obtain SVG files.

[0062] Furthermore, the standardization module 102 is used to standardize the size of the bounding box of the initial file to obtain a sample file, which specifically includes: The bounding box sizes of all the initial files are uniformly adjusted to a preset border threshold to obtain the sample file.

[0063] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.< / g> < / g> < / g> < / path> < / g> < / g> < / g> < / g> < / g> < / g> < / g> < / g>

Claims

1. A method for constructing a systematic color SVG icon dataset, characterized in that: The steps include: Collecting SVG files, and performing data cleaning on the collected SVG files to obtain an initial file; wherein the SVG file includes a plurality of SVG icons of different categories, and the initial file is a color SVG file; Standardizing the size of the bounding box of the initial file to obtain a sample file; The sample files are classified, and a training set, a validation set and a test set are divided according to the classified sample files to obtain an SVG data set.

2. The method for constructing a systematic color SVG icon dataset as claimed in claim 1, characterized in that: The collecting of SVG files specifically includes: The icon categories in the Icon645 dataset were filtered and expanded to obtain several collection categories; SVG icons of different collection categories are collected from the open source platform respectively, and corresponding category labels are added to all the SVG icons to obtain SVG files.

3. The method for constructing a systematic color SVG icon dataset as claimed in claim 1, characterized in that: The data cleaning of the collected SVG file to obtain the initial file specifically includes: Confirming the attributes of the SVG files respectively, removing the SVG files that lack the preset fill attribute, and rasterizing the remaining SVG files respectively to obtain corresponding bitmap images; Determine the pixel value of each pixel in the bitmap image respectively, and calculate the proportion of pixels with a pixel value of 0 in all pixels in the bitmap image; The bitmap images whose corresponding proportion is 0 and whose corresponding proportion exceeds a preset ratio threshold are removed, and the SVG files corresponding to the remaining bitmap images are determined as the initial files.

4. The method for constructing a systematic color SVG icon dataset as claimed in claim 1, characterized in that: The standardizing of the bounding box size of the initial file to obtain a sample file specifically includes: The bounding box sizes of all the initial files are uniformly adjusted to a preset border threshold to obtain the sample file.

5. The method for constructing a systematic color SVG icon dataset as claimed in claim 4, characterized in that: Before the initial file is standardized, the following steps are also included: In the initial file <g>The attributes of the tag are transferred to the <g>After the child node of the tag, remove the <g> Label.< / g> < / g> < / g> 6. The method for constructing a systematic color SVG icon dataset as claimed in claim 2, characterized in that: The classifying the sample files specifically includes: Rasterizing the sample file to obtain a raster image; Inputting all the raster images into an image encoder of a preset CLIP model to obtain image features, and inputting all the category labels into a text encoder of a preset CLIP model to obtain text features; For each of the image features, respectively calculating the similarity between the image feature and each text feature, and determining the category label corresponding to the text feature with the highest similarity as the correct category of the sample file corresponding to the image feature; Counting the number of raster images in each correct category, and taking the first a correct categories with the highest number of corresponding raster images as the data set category; wherein a is a preset ranking threshold; According to the data set category, all the raster images are reclassified through a preset CLIP model, and the confidence between the reclassified raster images and the corresponding categories is calculated, and the raster images with confidence lower than a preset confidence threshold are removed to complete the classification.

7. The method for constructing a systematic color SVG icon dataset as claimed in claim 1, characterized in that: The method divides the training set, the validation set and the test set according to the classified sample files to obtain the SVG data set, specifically including: According to the preset number of training sets, number of validation sets and number of test sets, the classified sample files are divided by stratified sampling to obtain the SVG data set.

8. A data set construction device for systematic color SVG icons, characterized in that: include: Acquisition module, standardization module and classification module; The acquisition module is used to acquire SVG files and perform data cleaning on the acquired SVG files to obtain an initial file; wherein the SVG file includes a plurality of SVG icons of different categories, and the initial file is a color SVG file; The standardization module is used to standardize the size of the bounding box of the initial file to obtain a sample file; The classification module is used to classify the sample files, and divide the sample files into a training set, a validation set and a test set according to the classified sample files to obtain an SVG data set.

9. The apparatus for constructing a data set of a systematic color SVG icon according to claim 8, characterized in that: The acquisition module is used to acquire SVG files, specifically including: The icon categories in the Icon645 dataset were filtered and expanded to obtain several collection categories; SVG icons of different collection categories are collected from the open source platform respectively, and corresponding category labels are added to all the SVG icons to obtain SVG files.

10. The apparatus for constructing a data set of a systematic color SVG icon according to claim 8, characterized in that: The standardization module is used to standardize the size of the bounding box of the initial file to obtain a sample file, which specifically includes: The bounding box sizes of all the initial files are uniformly adjusted to a preset border threshold to obtain the sample file.

Images