Rapid product picture loading method and system based on data mining algorithm

Abstract

The invention relates to the technical field of product picture rapid loading, and particularly discloses a product picture rapid loading method and system based on a data mining algorithm, and the method comprises the steps: obtaining product picture demand information inputted by a user; according to the product picture demand information, multi-modal feature extraction is carried out on original picture data in a product picture database to obtain comprehensive feature data, and the comprehensive feature data comprises visual features, text tags and user behavior features; and obtaining a candidate picture set according to the comprehensive feature data. By acquiring the historical preference data of the user and combining the product browsing frequency, the duration, the market popularity and the popularity trend, personalized sorting is performed on the pictures, the real-time picture loading list is generated, and the pictures which the user may be interested in are preferentially displayed, so that the loaded content better conforms to the real-time interest of the user, the situation that the user does not pay attention to is avoided, and the user experience is improved. And the resource utilization efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of product image fast loading, and in particular to a product image fast loading method and system based on a data mining algorithm. Background Art

[0002] In today's digital age, product images have become a crucial element in information dissemination and business promotion. Whether displaying products on e-commerce platforms, promoting products for businesses, or sharing content on social media platforms, fast and accurate loading of product images is crucial.

[0003] Existing image loading systems lack effective utilization of user behavior data and multimodal image data. On the one hand, user behavior data during image browsing, such as browsing time, click preferences, and search history, contains rich information about user interests, but traditional systems fail to fully tap into this data, resulting in a lack of personalized image loading. On the other hand, images themselves contain multimodal data such as visual features (such as color, texture, shape), text tags, and semantic information. Existing technologies mostly utilize only a single or a few types of this data for image screening and loading, failing to fully understand the image content, which affects the accuracy and efficiency of loading. Summary of the Invention

[0004] The purpose of the present invention is to provide a method and system for quickly loading product images based on a data mining algorithm to solve the technical problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions: The product image fast loading method based on data mining algorithm includes: Get product image requirement information input by the user; Performing multimodal feature extraction on original image data in a product image database according to the product image requirement information to obtain comprehensive feature data, wherein the comprehensive feature data includes visual features, text labels, and user behavior features; Acquire a candidate picture set according to the comprehensive feature data; Performing hierarchical screening on the candidate image sets to obtain a high-similarity image set; Acquire user historical preference data, and sort the high-similarity picture set based on the user historical preference data to generate a real-time picture loading list; The product images are loaded according to the real-time image loading list, and interaction information between the user and the loaded images is obtained. A subsequent image loading list is generated according to the interaction information and the real-time image loading list, and the remaining images are loaded according to the subsequent image loading list.

[0006] Preferably, the step of performing multimodal feature extraction on the original image data in the product image database according to the product image requirement information to obtain comprehensive feature data includes: Acquire text data, image data, and voice data according to the product image requirement information; Acquire text keywords according to the text data and voice data, and acquire first positioning-related pictures from a product picture database based on the text keywords; Acquire a second positioning-related picture according to the picture data; Performing multiple visual feature extractions on the first positioning-related image and the second positioning-related image using a CNN model to obtain first multimodal visual features and second multimodal visual features, wherein the first multimodal visual features and the second multimodal visual features both include color histogram features, texture directional gradient features, shape descriptor features, and scale feature transformation features; The first multimodal visual feature and the second multimodal visual feature are subjected to feature fusion to obtain comprehensive feature data.

[0007] Preferably, the step of obtaining a candidate picture set based on the comprehensive feature data includes: Perform principal component analysis on the comprehensive feature data to extract the main image feature data; Constructing a feature index based on the main image feature data; Acquire multiple index images according to the feature index; Classifying the plurality of index images to obtain a preliminary candidate set; A candidate picture set is obtained according to the preliminary candidate set.

[0008] Preferably, the step of performing hierarchical screening on the candidate image sets to obtain a high-similarity image set includes: Classify the candidate image sets according to the preset product categories to obtain multiple product classification image sets; Perform a preliminary screening of each product classification picture set according to product subcategories to obtain a screened picture set; Obtaining visual-text feature similarities of a plurality of filtered images based on the filtered image set; The filtered images are sorted according to the visual-text feature similarity, and the filtered images with the closest visual-text feature similarity are selected to generate a high-similarity image set.

[0009] Preferably, the step of obtaining user historical preference data and sorting the high-similarity picture set based on the user historical preference data to generate a real-time picture loading list includes: Obtain product browsing frequency and browsing duration in the user's historical preference data according to the user's historical preference data; Obtaining preference levels for products corresponding to the high-similarity image set based on the product browsing frequency and browsing duration; Get the current popularity and trend of products in high-similarity image sets; Generate a product popularity evaluation value based on the current popularity and popularity trend; Generate an importance value of the product corresponding to the high-similarity image set according to the preference degree and the product popularity evaluation value; The high-similarity picture sets are sorted according to the importance values to generate a real-time picture loading list.

[0010] Preferably, the step of generating a subsequent image loading list according to the interaction information and the real-time image loading list, and loading the remaining images according to the subsequent image loading list includes: Obtain display loading progress data of each loaded image in the real-time image loading list; Acquiring operation frequency data according to the displayed loading progress data; Acquire real-time interaction data between the user and the loaded image according to the displayed loading progress data, wherein the interaction information includes interaction type, interaction time, and interaction duration; Obtaining the user's attention to the loaded image based on the real-time interaction data and the operation frequency data; Determine whether the attention level is lower than a preset threshold; If it is lower than, it is determined that the user's attention to the high-similarity picture set is reduced, and the high-similarity picture sets with reduced attention are generated into a first loading list; If it is higher, it is determined that the user's attention to the high-similarity picture set has increased, and the high-similarity picture sets with increased attention are generated into a second loading list; Combine the first loading list and the second loading list to generate a subsequent image loading list; Load the remaining images according to the subsequent image loading list.

[0011] The present invention also provides a product image fast loading system based on a data mining algorithm, which is characterized by comprising: The first acquisition module is used to obtain product image requirement information input by the user; an extraction module, configured to perform multimodal feature extraction on the original image data in the product image database according to the product image requirement information to obtain comprehensive feature data, wherein the comprehensive feature data includes visual features, text labels, and user behavior features; A second acquisition module, configured to acquire a candidate picture set based on the comprehensive feature data; A screening module, configured to perform hierarchical screening on the candidate image sets to obtain a high-similarity image set; A third acquisition module is used to obtain user historical preference data, and sort the high-similarity picture set based on the user historical preference data to generate a real-time picture loading list; The loading module is used to load product images according to the real-time image loading list, obtain interaction information between the user and the loaded image, generate a subsequent image loading list according to the interaction information and the real-time image loading list, and load the remaining images according to the subsequent image loading list.

[0012] Preferably, the extraction module includes: A first acquisition unit, configured to acquire text data, image data, and voice data according to the product image requirement information; a second acquiring unit, configured to acquire text keywords according to the text data and voice data, and acquire a first positioning-related image from a product image database based on the text keywords; a third acquiring unit, configured to acquire a second positioning-related image according to the image data; a first extraction unit, configured to perform multiple visual feature extractions on the first positioning-related image and the second positioning-related image using a CNN model to obtain first multimodal visual features and second multimodal visual features, wherein the first multimodal visual features and the second multimodal visual features both include color histogram features, texture directional gradient features, shape descriptor features, and scale feature transformation features; The fusion unit is used to fuse the first multimodal visual feature and the second multimodal visual feature to obtain comprehensive feature data.

[0013] Preferably, the second acquisition module includes: The second extraction unit is used to perform principal component analysis on the comprehensive feature data to extract main image feature data; A building unit, configured to build a feature index based on the main picture feature data; a fourth acquiring unit, configured to acquire a plurality of index images according to the feature index; A first classification unit is used to classify the plurality of index images to obtain a preliminary candidate set; The fourth classification unit is configured to obtain a candidate picture set based on the preliminary candidate set.

[0014] Preferably, the screening module includes: The second classification unit is used to classify the candidate image sets according to preset product categories to obtain multiple product classification image sets; A screening unit is used to preliminarily screen each product classification picture set according to product subcategories to obtain a screened picture set; a fifth classification unit, configured to obtain visual-text feature similarities of a plurality of filtered images based on the filtered image set; The sorting unit is used to sort the screened pictures according to the visual-text feature similarity, and select the screened pictures with the closest visual-text feature similarity to generate a high-similarity picture set.

[0015] The beneficial effects of the present application are as follows: the present invention uses multimodal feature extraction to comprehensively characterize images, integrating visual, textual and user behavioral features, thereby improving screening accuracy. The hierarchical screening mechanism first performs preliminary screening by product category, and then fine-screens based on the similarity of visual-textual features. It can quickly locate high-matching images from a large number of candidate images, thereby improving screening efficiency. In terms of meeting user personalized needs, it obtains user historical preference data, combines product browsing frequency, duration, market popularity, and popular trends, and performs personalized sorting of images to generate a real-time image loading list, giving priority to displaying images that the user may be interested in, thereby enhancing user experience. In addition, during the image loading process, based on the interaction information between the user and the loaded image, such as operation frequency, interaction type, etc., the subsequent image loading list is dynamically adjusted to make the loaded content more in line with the user's real-time interests, avoid loading images that the user does not pay attention to, and improve resource utilization efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a schematic diagram of a method flow chart according to an embodiment of the present application.

[0017] Figure 2 This is a schematic diagram of the system structure of an embodiment of the present application.

[0018] The realization of the objectives, functional features and advantages of this application will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0019] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0020] like Figure 1 As shown, this application provides a method for quickly loading product images based on a data mining algorithm, including: S1. Obtain product image requirement information input by the user; S2. Performing multimodal feature extraction on original image data in a product image database according to the product image requirement information to obtain comprehensive feature data, wherein the comprehensive feature data includes visual features, text labels, and user behavior features; S3. Acquire a candidate picture set based on the comprehensive feature data; S4, performing hierarchical screening on the candidate image set to obtain a high-similarity image set; S5. Obtain user historical preference data, and sort the high-similarity picture set based on the user historical preference data to generate a real-time picture loading list; S6. Load product images according to the real-time image loading list, obtain interaction information between the user and the loaded images, generate a subsequent image loading list according to the interaction information and the real-time image loading list, and load the remaining images according to the subsequent image loading list.

[0021] As described in the above steps S1-S6, the present invention obtains product picture demand information input by the user. In order to solve the problem that the user demand acquisition method in the prior art is single and it is difficult to accurately express complex demands, a multimodal input interface is used to obtain support for text, picture, and voice input. For text input, a format check will be performed to remove redundant symbols, and then the keywords will be identified through part-of-speech analysis; when uploading pictures, the resolution will be detected and features will be pre-extracted; and the voice input will be converted into text for semantic understanding. Finally, the information obtained from these different input methods is integrated to fully and accurately grasp the user needs and provide an accurate basis for subsequent screening of pictures. Then, according to the product picture demand information, multimodal feature extraction is performed on the original picture data in the product picture database to obtain comprehensive feature data. The traditional method is relatively single in feature extraction, focusing only on vision or text, ignoring user behavior characteristics, resulting in an inability to fully describe the picture, affecting the screening accuracy.

[0022] This method first obtains text, pictures, and voice data from the demand information, and uses text keywords and picture data to locate relevant pictures. Then, the CNN model is used to extract visual features such as color histogram features, texture directional gradient features, shape descriptor features, and scale feature transformation from the located pictures multiple times. OCR technology is used to identify text labels, and user historical behavior is analyzed to obtain behavioral features. Finally, the visual features, text labels, and user behavior features are integrated. This results in a more comprehensive and rich description of picture features, greatly improving the accuracy of matching pictures with demand, reducing screening errors, and improving screening efficiency. Afterwards, a candidate picture set is obtained based on the comprehensive feature data. Existing screening methods are inefficient when faced with a large number of pictures. Blind matching not only consumes a lot of time and resources, but also easily misses potential matching pictures. This method first performs principal component analysis on the comprehensive feature data to extract the main picture feature data, and then builds an index based on these key features. Through feature indexing, multiple index pictures can be quickly found, and these index pictures are classified to obtain a preliminary candidate set, and then the preliminary candidate set is optimized to obtain a candidate picture set. This process greatly reduces the scope of picture screening, significantly improves screening efficiency, quickly locates pictures that may meet the needs, saves resources and user time, Then, the candidate image sets are hierarchically screened to obtain high-similarity image sets. In the past, screening lacked hierarchy, making it difficult to distinguish image similarities, resulting in uneven quality of screening results. This method first classifies the candidate image sets according to the preset product categories, and then preliminarily screens each product classification image set according to the product subcategories. Afterwards, the visual-text feature similarity of the screened images is calculated, the screened images are sorted according to the similarity, and the screened images with the closest visual-text feature similarity are selected to generate a high-similarity image set. This hierarchical screening method improves the quality and accuracy of the screened images, so that the screened images are highly matched with user needs, effectively reducing the user's secondary screening costs. Subsequently, the user's historical preference data is obtained, and based on this, the high-similarity image sets are sorted to generate a real-time image loading list. The existing technology lacks personalized sorting, and all users see the same order of images, which cannot meet personalized needs and affects the user experience.

[0023] This method obtains the user's historical browsing frequency and duration to calculate the degree of preference for the products corresponding to the high-similarity picture set. At the same time, it obtains the current popularity and popularity trend of the products corresponding to the high-similarity picture set, generates a product popularity evaluation value, and combines the preference degree and product popularity evaluation value to obtain the importance value of the products corresponding to the high-similarity picture set. Finally, the high-similarity picture set is sorted according to the importance value, and a real-time picture loading list is generated. This process realizes personalized sorting of picture loading, giving priority to displaying pictures that the user may be interested in, increasing the probability of users finding their favorite pictures, and enhancing the user experience. Finally, the product pictures are loaded according to the real-time picture loading list, and the interaction information between the user and the loaded pictures is obtained. Based on this information, a subsequent picture loading list is generated to load the remaining pictures. Traditional loading methods cannot adjust the loading strategy according to real-time user feedback, resulting in an inconsistent user experience.

[0024] In the process of loading images according to the real-time image loading list, this method obtains the display loading progress data of each loaded image, and then obtains the user operation frequency data and the real-time interaction data between the user and the loaded image, including the interaction type, interaction time, and interaction duration. Based on this real-time interaction data, the user's attention to the loaded image is calculated, and the high-similarity image set is screened based on the attention. If the attention is lower than the preset threshold, it is judged that the user's attention to the high-similarity image set has decreased, and this part of the images is generated into a first loading list; if the attention is higher than the preset threshold, it is judged that the user's attention to the high-similarity image set has increased, and the relevant images are generated into a second loading list. Finally, the first loading list and the second loading list are combined to generate a subsequent image loading list, and the remaining images are loaded according to this list. This enables the loading list to be dynamically adjusted according to real-time user feedback, improving the fit between image loading and user needs.

[0025] In one embodiment, the step of performing multimodal feature extraction on the original image data in the product image database according to the product image requirement information to obtain comprehensive feature data includes: S201, acquiring text data, image data, and voice data according to the product image requirement information; S202: Acquire text keywords according to the text data and voice data, and acquire a first positioning-related image from a product image database based on the text keywords; S203, obtaining a second positioning-related image according to the image data; S204, performing multiple visual feature extractions on the first positioning-related image and the second positioning-related image using a CNN model to obtain first multimodal visual features and second multimodal visual features, wherein the first multimodal visual features and the second multimodal visual features both include color histogram features, texture directional gradient features, shape descriptor features, and scale feature transformation features; S205 : Fusing the first multimodal visual feature and the second multimodal visual feature to obtain comprehensive feature data.

[0026] As described in steps S201-S205 above, the present invention obtains text data, image data, and voice data based on product image requirement information. To address the issue of limited user requirement data acquisition in existing technologies, users can enter text through a text input box for real-time recording; upload images, which undergo preliminary format and size checks; and use voice input, which is converted to text using voice recognition technology. This data is then categorized and organized, extracting text data (key descriptive information entered by the user for preliminary image location), image data (user-uploaded images containing intuitive product visual features to aid in accurate screening), and voice data (converted text from voice input to enrich the expression of the requirement). Next, text keywords are extracted from the text and voice data, and based on these text keywords, the first relevant image for location is retrieved from a product image database. Existing text data processing methods struggle to efficiently extract key information, leading to inaccurate search scopes. Therefore, the text and voice data are preprocessed to remove stop words and special symbols. Natural language processing techniques are then used to extract keywords that provide a high-level summary of the user's requirements. Finally, using keywords as search conditions, we searched the product image database through the database indexing mechanism to obtain first-position related images (preliminarily screening out images related to text requirements), narrowing the subsequent screening scope, improving screening efficiency, and obtaining second-position related images based on the image data.

[0027] Since existing technologies cannot fully utilize the image data uploaded by users for accurate screening, and text screening alone cannot capture subtle visual differences in images, basic visual features such as color, shape, and texture are extracted from the image data uploaded by users in this step. Then, using the cosine similarity algorithm, etc., the extracted features are matched with the images in the database to calculate similarity, and the second-positioning related images similar to the uploaded images are screened out, which supplements the shortcomings of text keyword screening alone and improves the screening accuracy. Then, the CNN model is used to perform multiple visual feature extractions on the first-positioning related images and the second-positioning related images (before feature extraction, the first-positioning related images and the second-positioning related images need to be preprocessed. This includes adjusting the image size to meet the CNN model input requirements, such as uniformly scaling images of different sizes to a specific size; the images will also be normalized. The pixel values are mapped to a specific interval (such as [0,1] or [-1,1]) to speed up model training and improve stability. For example, for some product images with large size differences, they are scaled to 224×224 pixels and the pre-processed images are input into the convolution layer of the CNN model. The convolution layer consists of multiple convolution kernels, each of which can be regarded as a feature detector. In the first convolution operation, the convolution kernel slides on the image and performs a convolution operation on the local area, that is, the pixel values covered by the convolution kernel are weighted and summed to generate new feature values, which constitute the preliminary feature map. For example, a 3 ×3 convolution kernel can detect simple features such as edges and textures in the image. As the convolution layer deepens, the subsequent convolution layer will perform convolution again based on the feature map of the previous layer to extract more complex and abstract features. For example, in the subsequent convolution layer, edge features may be combined to form more advanced shape features. Multiple convolution operations can extract features from images at different scales and angles. After each convolution operation, an activation function (such as ReLU function) will be introduced. By using the activation function after multiple convolution layers, the model can learn various nonlinear visual feature relationships in the image. After multiple convolution and activation function processing, Use the pooling layer to reduce the dimensionality of the feature map. Common pooling methods include maximum pooling and average pooling. Maximum pooling selects the maximum value in each pooling window as the output, while average pooling calculates the average value within the window as the output. Multiple visual feature extractions will repeat the convolution, activation, and pooling process. Each repetition allows the model to learn more advanced and abstract visual features. As the network depth increases, the simple edge and texture features extracted by the early convolutional layers are gradually combined and abstracted into more complex shape and structural features. For example, from initially recognizing lines in an image, to subsequently recognizing the outline of an object, and finally recognizing the complete object.Through multiple iterations, the CNN model can extract multi-level and multi-dimensional visual features from the image, providing rich feature representations for subsequent image classification, retrieval and other tasks), obtain the first multimodal visual feature and the second multimodal visual feature, and finally, fuse the first multimodal visual feature and the second multimodal visual feature to obtain comprehensive feature data. Because the existing technology processes different types of image features independently and cannot give full play to their advantages, feature splicing or weighted fusion is adopted. Feature splicing is to connect the two feature vectors in sequence, and weighted fusion is to assign weights according to the importance of the features and then sum them up. Through fusion, the features obtained by locating the image based on text keywords and locating the image based on image similarity matching are integrated to form comprehensive feature data, which fully utilizes the advantages of multimodal data and improves the matching degree between the image and user needs.

[0028] In one embodiment, the step of obtaining a candidate picture set based on the comprehensive feature data includes: S301, perform principal component analysis on the comprehensive feature data to extract main image feature data; S302, constructing a feature index based on the main image feature data; S303, obtaining multiple index images according to the feature index; S304, classifying the plurality of index images to obtain a preliminary candidate set; S305: Obtain a candidate picture set according to the preliminary candidate set.

[0029] As described in steps S301-S305 above, the present invention extracts primary image feature data by performing principal component analysis on the comprehensive feature data. The comprehensive feature data encompasses multimodal information such as visual, textual, and user behavior information from the image. Next, the eigenvalues and eigenvectors of the covariance matrix are determined through eigenvalue decomposition or singular value decomposition. The top principal components are selected based on their eigenvalues, and the original comprehensive feature data is projected onto these principal components to obtain the primary image feature data. This step effectively reduces data dimensionality and removes redundant information, providing a more streamlined and efficient data foundation for subsequent screening. The comprehensive feature data is the original multimodal data set that comprehensively describes the image; the primary image feature data is the key feature data extracted through principal component analysis, preserving the key information of the original data. A feature index is constructed based on the primary image feature data. In large-scale image databases, image retrieval can be extremely time-consuming without an effective indexing mechanism. The primary image feature data is quantized and encoded, converting it into a data structure suitable for index storage, such as a hash value or vector index. The encoded feature data is then associated with the corresponding image identifier and stored in an index table or index database. This allows for quick retrieval of image identifiers using the feature index when searching for specific features, without having to search the entire database. The primary image feature data serves as the basis for indexing, while the feature index is a data structure that stores the correspondence between image feature codes and image identifiers, enabling fast searches and retrieving multiple indexed images based on the feature index.

[0030] Traditional image search methods are inefficient and difficult to quickly find images relevant to user needs from massive amounts of images. Upon receiving a user's filtering request, the request features are converted into a format that matches the index. Matching image identifiers are searched in the index, and then, based on these identifiers, multiple corresponding index images are retrieved from the image database. This process, leveraging feature indexing, enables rapid location of relevant images from a massive image pool, significantly improving search efficiency. Feature indexing facilitates rapid locating. Index images are a collection of images relevant to the user's requested features retrieved from the database through index search. This facilitates subsequent filtering, and multiple index images are classified to form a preliminary candidate set. Index images retrieved from the database often contain redundant and irrelevant content, and the lack of further filtering and classification mechanisms increases the user's screening burden. Classification rules are developed based on attributes and features such as image color, shape, style, and product category. Each index image is matched to the rules to determine its category, eliminating images that do not meet the requirements and grouping images that meet specific criteria into a single category to form a preliminary candidate set. Index images are the objects of classification, while the preliminary candidate set is a collection of images that, after classification and screening, better meet the user's needs, narrowing the scope of the search. The candidate set is then obtained based on the preliminary candidate set. The preliminary candidate set may still contain images that do not meet user requirements or have varying quality, which can affect the user experience. Based on the user's specific needs, we further refine the screening criteria, such as imposing stricter requirements on image color, resolution, and product details. We then examine each image in the preliminary candidate set, eliminating those that do not meet the criteria. We then perform a quality assessment and sorting of the remaining images to finalize the candidate set. The preliminary candidate set is the basis for obtaining the candidate set. The candidate set is a collection of images that, after further screening and optimization, more accurately meets user needs and is used for subsequent image loading and display.

[0031] In one embodiment, the step of performing hierarchical screening on the candidate image sets to obtain a high-similarity image set includes: S401, classifying the candidate image sets according to preset product categories to obtain multiple product classification image sets; S402, preliminarily screening each product classification picture set according to product subcategories to obtain a screened picture set; S403, obtaining visual-text feature similarities of a plurality of filtered images according to the filtered image set; S404 , sorting the filtered images according to the visual-text feature similarity, and selecting the filtered images with the closest visual-text feature similarity to generate a high-similarity image set.

[0032] As described in steps S401-S404 above, the present invention classifies candidate image sets according to preset product categories to obtain multiple product classification image sets. In the prior art, the lack of preliminary classification when processing a large number of candidate images leads to an inefficient and chaotic screening process. The candidate image sets are classified according to preset product classification criteria, such as common categories such as electronics, clothing, and household items. The image's text tags, visual features, and metadata are analyzed to determine its category. For example, if an image's text tag contains the keyword "computer" or visually exhibits computer features, it is classified as an electronics product. In this way, the mixed candidate image set is divided into multiple product classification image sets. The candidate image set is a preliminary screening of images that may meet user needs; the preset product categories are pre-set classification criteria; and the product classification image set is a collection of images belonging to the same category, resulting from classification. This narrows the screening scope and improves screening efficiency. Each product classification image set is then preliminarily screened according to product subcategories to obtain a filtered image set. Classification by broad categories alone cannot meet the needs of accurate screening. Products in the same broad category vary greatly, and not subdividing them will interfere with the screening results. For each product classification image set, filter again based on the product subcategory standards. For example, taking the electronic product classification image set as an example, its subcategories can be divided into mobile phones, computers, cameras, etc. Filtering is performed by analyzing detailed information in the image, such as product details and text descriptions. For example, when filtering images in the mobile phone subcategory, pay attention to whether the product shape resembles a mobile phone and whether the text description contains words such as "mobile phone model". Images that meet the requirements are filtered out to form a filtered image set for the mobile phone subcategory.

[0033] The product classification image set is the result of the previous classification step and is also the target of this screening. Product subcategories are subcategories based on the main categories. The filtered image set is a collection of images with higher relevance after filtering by subcategories, further narrowing the screening scope. The visual-text feature similarity of multiple filtered images is obtained based on the filtered image set. Existing screening methods often fail to comprehensively utilize the visual and textual features of images, resulting in the filtered images not being well matched with user needs. For each image in the filtered image set, visual features and textual features are extracted separately. Convolutional neural networks and other technologies are used to obtain visual features such as color, texture, and shape of the images. Natural language processing is used to extract textual features such as keywords and semantics from the image descriptions and labels. Then, the cosine similarity algorithm is used to calculate the similarity between the visual feature vector and the textual feature vector of each image, and the similarity is recorded in association with the image to form a dataset containing the images and their corresponding similarities. The filtered image set is the subject of this feature extraction and similarity calculation. Visual-text feature similarity measures the degree of similarity between images in both the visual and textual dimensions, providing a basis for subsequent screening. This step comprehensively considers the multi-dimensional information of the image, improving the accuracy of the screening. Finally, the filtered images are sorted according to their visual-text feature similarity, and the filtered images with the closest visual-text feature similarity are selected to generate a high-similarity image set. Existing screening methods lack an effective sorting and precise selection mechanism for image similarity, resulting in low-quality screening results. Based on the visual-text feature similarity obtained in the previous step, the images in the filtered image set are sorted in descending or ascending order (usually descending to select images with high similarity).

[0034] After sorting, the images with the closest similarity are selected according to predefined rules. This rule can include selecting the top N images or images with a similarity greater than a certain threshold. These selected images are then combined to form a high-similarity image set. Visual-text feature similarity serves as the basis for sorting and filtering; the filtered images are the objects of operation. High-similarity image sets are high-quality screening results, where the images closely match user requirements. This ensures the high quality and usability of the screening results.

[0035] In one embodiment, the step of obtaining user historical preference data and sorting the high-similarity picture set based on the user historical preference data to generate a real-time picture loading list includes: S501, obtaining product browsing frequency and browsing duration in the user's historical preference data according to the user's historical preference data; S502: Obtaining preference levels for products corresponding to the high-similarity image set based on the product browsing frequency and browsing duration; S503: Obtain the current popularity and trend of the product corresponding to the high-similarity image set; S504: Generate a product popularity evaluation value based on the current popularity and popularity trend; S505: Generate an importance value for the product corresponding to the high-similarity image set based on the preference degree and the product popularity evaluation value; S506: Sort the high-similarity picture sets according to the importance values to generate a real-time picture loading list.

[0036] As described in steps S501-S506 above, the present invention obtains product browsing frequency and browsing duration based on user historical preference data. Existing techniques often ignore user historical preferences when sorting images. This method analyzes and compiles user historical preference data stored in a database, counting the number of times users browsed different products to obtain browsing frequency, and summing the duration of each browse to obtain browsing duration. User historical preference data is a collection of information about a user's past product browsing history; product browsing frequency reflects the frequency of a user's attention to a specific product; and browsing duration reflects the depth of a user's interest in a specific product. This step mines historical user behavior data, laying the foundation for subsequent evaluation of user preferences. Next, based on product browsing frequency and browsing duration, the preference level for products in a set of highly similar images is determined. Browsing frequency and browsing duration alone cannot intuitively reflect a user's overall preference for a product, and existing techniques lack a mechanism for comprehensively considering these two metrics. A preset algorithm assigns weights to browsing frequency and browsing duration, multiplying them by their respective weights and adding them together to determine a product preference score. For example, we assign a weight of 0.6 to browsing frequency and a weight of 0.4 to browsing duration to calculate the preference level for products in a high-similarity image set. Here, product browsing frequency and browsing duration are important indicators for calculating preference level, which is a numerical value derived from combining the two to measure the user's preference for the product. This step more accurately assesses user preferences for products and determines the current popularity and trends of products in a high-similarity image set. Existing image sorting methods often ignore product market popularity and trends, causing users to miss out on popular product images.

[0037] By collecting data such as product discussion popularity and search volume on social media platforms and referencing industry reports and sales data, we analyze changes in product market attention and thus determine its popularity trends. Highly similar image sets correspond to products, i.e., the products to which the highly similar images correspond. Current popularity reflects the current level of interest in the product, while trends reflect the product's market development. To make image sorting more comprehensive, a product popularity assessment is generated based on current popularity and trends. Current popularity and trends are distinct metrics, making them difficult to directly use for image sorting. Existing technologies lack methods to integrate them. Evaluation rules are developed to quantify and comprehensively calculate current popularity and trends, assigning scores to different levels of popularity and trend status, and then performing a weighted summation. For example, current popularity can be categorized as high, medium, and low, with scores of 3, 2, and 1, respectively; and trends can be categorized as rising, stable, and declining, with scores of 3, 2, and 1, respectively, to derive a product popularity assessment. Current popularity and popularity trends are key factors in calculating a product's popularity rating. The popularity rating is a quantitative value derived from combining these two factors, facilitating subsequent unified consideration. Based on preference and popularity, a product importance rating is generated for each highly similar image set. Existing image ranking methods fail to integrate user preferences and market factors, resulting in overly biased rankings. Weights are assigned to preference and popularity, multiplied by their respective weights, and then added together to derive the product importance rating. Preference reflects user preference for a product, while popularity reflects its market appeal. The importance rating combines these two factors to more comprehensively reflect a product's appeal to users, enabling more rational image ranking. Finally, the highly similar image sets are sorted according to importance ratings to generate a real-time image loading list. Improper sorting of the importance ratings can negatively impact the user's viewing experience. The highly similar image sets are sorted in descending order based on importance ratings, placing images corresponding to products with higher importance ratings at the top of the list to generate a real-time image loading list. The importance value is used as the sorting basis, and the high-similarity image set is the sorting object. The real-time image loading list arranges images according to the comprehensive importance of the product to the user, making it convenient for users to quickly find product images that are of interest and in line with market trends, thereby improving the user experience.

[0038] In one embodiment, the step of generating a subsequent image loading list according to the interaction information and the real-time image loading list, and loading the remaining images according to the subsequent image loading list includes: S601, obtaining display loading progress data of each loaded image in the real-time image loading list; S602, acquiring operation frequency data according to the displayed loading progress data; S603: Acquire real-time interaction data between the user and the loaded image based on the displayed loading progress data, wherein the interaction information includes interaction type, interaction time, and interaction duration; S604: Obtain the user's attention to the loaded image based on the real-time interaction data and the operation frequency data; S605: Determine whether the attention level is lower than a preset threshold; If it is lower than, it is determined that the user's attention to the high-similarity picture set is reduced, and the high-similarity picture sets with reduced attention are generated into a first loading list; If it is higher, it is determined that the user's attention to the high-similarity picture set has increased, and the high-similarity picture sets with increased attention are generated into a second loading list; S606: Generate a subsequent image loading list by combining the first loading list and the second loading list; S607: Load the remaining pictures according to the subsequent picture loading list.

[0039] As described in steps S601-S607 above, the present invention obtains display loading progress data for each loaded image in the real-time image loading list. It is difficult for the existing technology to accurately grasp the image loading progress, resulting in the inability to flexibly adjust the loading strategy. When the image is loading, by monitoring the loading process, the number of bytes loaded for each loaded image, the loading percentage and other information are recorded to form display loading progress data. The real-time image loading list contains information about the images to be loaded and is the data source; the loaded images are the specific images being loaded in the list; the display loading progress data is used to reflect the image loading progress and provide a basis for subsequent adjustment strategies. This step realizes real-time tracking of the image loading progress. Then, the operation frequency data is obtained based on the display loading progress data. The existing technology does not fully utilize the loading progress data to analyze the user operation frequency, which makes the loading strategy mismatched with the user's operating habits. Combined with the display loading progress data, the number of user operations such as sliding and clicking is counted within the specific time window of image loading to obtain the operation frequency data. Display loading progress data provides a time reference, while operation frequency data reflects the frequency of user operations, providing a reference for optimizing loading strategies. Real-time interaction data between users and loading images, including interaction type, time, and duration, is obtained based on display loading progress data. Existing technologies fail to collect sufficient information about user-image interactions, making it difficult to accurately identify user interests. By combining display loading progress data with real-time interaction data, the time, specific actions, and duration of user interactions with the loading image are recorded, generating real-time interaction data. Display loading progress data assists in determining the image loading status at the time of interaction. The interaction type, time, and duration included in real-time interaction data provide important insights for analyzing user attention. This allows for comprehensive collection of user-image interactions, and the degree of user attention towards the loading image can be determined based on real-time interaction data and operation frequency data. Existing technologies lack effective methods for assessing user attention, resulting in a mismatch between image loading and user interest. Evaluation rules can be developed based on factors such as interaction type, time, and duration in real-time interaction data to calculate user attention. For example, different interaction types can be weighted differently, with longer interaction durations receiving higher scores.

[0040] The attention metric quantifies the user's level of interest in an image, providing a clear indicator for subsequent image screening and sorting. Highly similar image sets are screened based on attention, and the level of interest is determined to be below a preset threshold. Existing technologies do not filter images based on user attention, potentially loading images that the user is not interested in, wasting resources. Each image's attention metric is compared with a preset threshold. Highly similar image sets containing images below the threshold are generated into a first load list, while those above the threshold are generated into a second load list. Attention metric serves as the screening criterion, while high-similarity image sets serve as the screening targets. The preset threshold is used to distinguish between image sets of interest and those of indifference. The first load list contains image sets with decreasing attention, while the second load list contains image sets with increasing attention. The first and second load lists are combined to generate a subsequent image load list. Existing technologies lack a mechanism for integrating images after classification, resulting in an irrational loading order. The first and second load lists are integrated, typically prioritizing the loading of images in the second load list, while images in the first load list are prioritized or loaded in smaller numbers, to generate a subsequent image load list. The first loading list and the second loading list provide picture set information, and the subsequent picture loading list determines the picture loading order, which improves the rationality of loading. Finally, the remaining pictures are loaded according to the subsequent picture loading list. The existing technology may not be able to load pictures accurately in the order of the list, resulting in loading confusion. Strictly follow the order of the pictures in the subsequent picture loading list and load the remaining pictures in sequence to ensure the correct loading order. The subsequent picture loading list specifies the loading order, and the remaining pictures are the objects to be loaded. This step ensures that the pictures are loaded in a reasonable order.

[0041] like Figure 2 As shown, the present invention also provides a product image fast loading system based on data mining algorithm, including: The first acquisition module 1 is used to obtain product image requirement information input by the user; Extraction module 2, configured to perform multimodal feature extraction on the original image data in the product image database according to the product image requirement information to obtain comprehensive feature data, wherein the comprehensive feature data includes visual features, text labels, and user behavior features; A second acquisition module 3, configured to acquire a candidate picture set based on the comprehensive feature data; Screening module 4, used for performing hierarchical screening on the candidate image sets to obtain high-similarity image sets; The third acquisition module 5 is used to obtain user historical preference data, and sort the high-similarity picture set based on the user historical preference data to generate a real-time picture loading list; The loading module 6 is used to load product images according to the real-time image loading list, obtain interaction information between the user and the loaded images, generate a subsequent image loading list according to the interaction information and the real-time image loading list, and load the remaining images according to the subsequent image loading list.

[0042] In one embodiment, the extraction module 2 includes: A first acquisition unit, configured to acquire text data, image data, and voice data according to the product image requirement information; a second acquiring unit, configured to acquire text keywords according to the text data and voice data, and acquire a first positioning-related image from a product image database based on the text keywords; a third acquiring unit, configured to acquire a second positioning-related image according to the image data; a first extraction unit, configured to perform multiple visual feature extractions on the first positioning-related image and the second positioning-related image using a CNN model to obtain first multimodal visual features and second multimodal visual features, wherein the first multimodal visual features and the second multimodal visual features both include color histogram features, texture directional gradient features, shape descriptor features, and scale feature transformation features; The fusion unit is used to fuse the first multimodal visual feature and the second multimodal visual feature to obtain comprehensive feature data.

[0043] In one embodiment, the second acquisition module 3 includes: The second extraction unit is used to perform principal component analysis on the comprehensive feature data to extract main image feature data; A building unit, configured to build a feature index based on the main picture feature data; a fourth acquiring unit, configured to acquire a plurality of index images according to the feature index; A first classification unit is used to classify the plurality of index images to obtain a preliminary candidate set; The fourth classification unit is configured to obtain a candidate picture set based on the preliminary candidate set.

[0044] In one embodiment, the screening module 4 includes: The second classification unit is used to classify the candidate image sets according to preset product categories to obtain multiple product classification image sets; A screening unit is used to preliminarily screen each product classification picture set according to product subcategories to obtain a screened picture set; a fifth classification unit, configured to obtain visual-text feature similarities of a plurality of filtered images based on the filtered image set; The sorting unit is used to sort the screened pictures according to the visual-text feature similarity, and select the screened pictures with the closest visual-text feature similarity to generate a high-similarity picture set.

[0045] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, value library or other media provided in this application and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct RAM bus dynamic RAM (DRDRAM), and RAM bus dynamic RAM (RDRAM).

[0046] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, apparatus, article, or method comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, apparatus, article, or method. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, apparatus, article, or method comprising the element.

[0047] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent results or equivalent process transformations made using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for quickly loading product images based on a data mining algorithm, characterized in that: include: Get product image requirement information input by the user; Performing multimodal feature extraction on original image data in a product image database according to the product image requirement information to obtain comprehensive feature data, wherein the comprehensive feature data includes visual features, text labels, and user behavior features; Acquire a candidate picture set according to the comprehensive feature data; Performing hierarchical screening on the candidate image sets to obtain a high-similarity image set; Acquire user historical preference data, and sort the high-similarity picture set based on the user historical preference data to generate a real-time picture loading list; The product images are loaded according to the real-time image loading list, and interaction information between the user and the loaded images is obtained. A subsequent image loading list is generated according to the interaction information and the real-time image loading list, and the remaining images are loaded according to the subsequent image loading list.

2. The method for quickly loading product images based on data mining algorithm according to claim 1, characterized in that: The step of performing multimodal feature extraction on the original image data in the product image database according to the product image requirement information to obtain comprehensive feature data includes: Acquire text data, image data, and voice data according to the product image requirement information; Acquire text keywords according to the text data and voice data, and acquire first positioning-related pictures from a product picture database based on the text keywords; Acquire a second positioning-related picture according to the picture data; Performing multiple visual feature extractions on the first positioning-related image and the second positioning-related image using a CNN model to obtain first multimodal visual features and second multimodal visual features, wherein the first multimodal visual features and the second multimodal visual features both include color histogram features, texture directional gradient features, shape descriptor features, and scale feature transformation features; The first multimodal visual feature and the second multimodal visual feature are subjected to feature fusion to obtain comprehensive feature data.

3. The method for quickly loading product images based on data mining algorithm according to claim 1, characterized in that: The step of obtaining a candidate picture set according to the comprehensive feature data includes: Perform principal component analysis on the comprehensive feature data to extract the main image feature data; Constructing a feature index based on the main image feature data; Acquire multiple index images according to the feature index; Classifying the plurality of index images to obtain a preliminary candidate set; A candidate picture set is obtained according to the preliminary candidate set.

4. The method for quickly loading product images based on data mining algorithm according to claim 1, characterized in that: The step of performing hierarchical screening on the candidate picture sets to obtain a high-similarity picture set comprises: Classify the candidate image sets according to the preset product categories to obtain multiple product classification image sets; Perform a preliminary screening of each product classification picture set according to product subcategories to obtain a screened picture set; Obtaining visual-text feature similarities of a plurality of filtered images based on the filtered image set; The filtered images are sorted according to the visual-text feature similarity, and the filtered images with the closest visual-text feature similarity are selected to generate a high-similarity image set.

5. The method for quickly loading product images based on data mining algorithm according to claim 1, characterized in that: The step of obtaining user historical preference data and sorting the high-similarity picture set based on the user historical preference data to generate a real-time picture loading list includes: Obtain product browsing frequency and browsing duration in the user's historical preference data according to the user's historical preference data; Obtaining preference levels for products corresponding to the high-similarity image set based on the product browsing frequency and browsing duration; Get the current popularity and trend of products in high-similarity image sets; Generate a product popularity evaluation value based on the current popularity and popularity trend; Generate an importance value of the product corresponding to the high-similarity image set according to the preference degree and the product popularity evaluation value; The high-similarity picture sets are sorted according to the importance values to generate a real-time picture loading list.

6. The method for quickly loading product images based on data mining algorithm according to claim 1, characterized in that: The step of generating a subsequent image loading list according to the interaction information and the real-time image loading list, and loading the remaining images according to the subsequent image loading list includes: Obtain display loading progress data of each loaded image in the real-time image loading list; Acquiring operation frequency data according to the displayed loading progress data; Acquire real-time interaction data between the user and the loaded image according to the displayed loading progress data, wherein the interaction information includes interaction type, interaction time, and interaction duration; Obtaining the user's attention to the loaded image based on the real-time interaction data and the operation frequency data; Determine whether the attention level is lower than a preset threshold; If it is lower than, it is determined that the user's attention to the high-similarity picture set is reduced, and the high-similarity picture sets with reduced attention are generated into a first loading list; If it is higher, it is determined that the user's attention to the high-similarity picture set has increased, and the high-similarity picture sets with increased attention are generated into a second loading list; Combine the first loading list and the second loading list to generate a subsequent image loading list; Load the remaining images according to the subsequent image loading list.

7. Product image fast loading system based on data mining algorithm, characterized by: include: The first acquisition module is used to obtain product image requirement information input by the user; an extraction module, configured to perform multimodal feature extraction on the original image data in the product image database according to the product image requirement information to obtain comprehensive feature data, wherein the comprehensive feature data includes visual features, text labels, and user behavior features; A second acquisition module, configured to acquire a candidate picture set based on the comprehensive feature data; A screening module, configured to perform hierarchical screening on the candidate image sets to obtain a high-similarity image set; A third acquisition module is used to obtain user historical preference data, and sort the high-similarity picture set based on the user historical preference data to generate a real-time picture loading list; The loading module is used to load product images according to the real-time image loading list, obtain interaction information between the user and the loaded image, generate a subsequent image loading list according to the interaction information and the real-time image loading list, and load the remaining images according to the subsequent image loading list.

8. The product image fast loading system based on data mining algorithm according to claim 7 is characterized in that: The extraction module comprises: A first acquisition unit, configured to acquire text data, image data, and voice data according to the product image requirement information; a second acquiring unit, configured to acquire text keywords according to the text data and voice data, and acquire a first positioning-related image from a product image database based on the text keywords; a third acquiring unit, configured to acquire a second positioning-related image according to the image data; a first extraction unit, configured to perform multiple visual feature extractions on the first positioning-related image and the second positioning-related image using a CNN model to obtain first multimodal visual features and second multimodal visual features, wherein the first multimodal visual features and the second multimodal visual features both include color histogram features, texture directional gradient features, shape descriptor features, and scale feature transformation features; The fusion unit is used to fuse the first multimodal visual feature and the second multimodal visual feature to obtain comprehensive feature data.

9. The product image fast loading system based on data mining algorithm according to claim 7 is characterized in that: The second acquisition module includes: The second extraction unit is used to perform principal component analysis on the comprehensive feature data to extract main image feature data; A building unit, configured to build a feature index based on the main picture feature data; a fourth acquiring unit, configured to acquire a plurality of index images according to the feature index; A first classification unit is used to classify the plurality of index images to obtain a preliminary candidate set; The fourth classification unit is configured to obtain a candidate picture set based on the preliminary candidate set.

10. The product image fast loading system based on data mining algorithm according to claim 7, characterized in that: The screening module comprises: The second classification unit is used to classify the candidate image sets according to preset product categories to obtain multiple product classification image sets; A screening unit is used to preliminarily screen each product classification picture set according to product subcategories to obtain a screened picture set; a fifth classification unit, configured to obtain visual-text feature similarities of a plurality of filtered images based on the filtered image set; The sorting unit is used to sort the screened pictures according to the visual-text feature similarity, and select the screened pictures with the closest visual-text feature similarity to generate a high-similarity picture set.

Images