A method, apparatus, equipment, and medium for filtering product scene images.

Abstract

This invention provides a method, apparatus, device, and medium for filtering product scene images. The method includes: acquiring a set of images to be processed; performing edge pixel sampling on each image in the set of images to be processed; calculating the proportion of white pixels in the edge region of the image; if the proportion of white pixels is greater than or equal to a preset first threshold, then marking the image as a white background image and removing it; otherwise, marking the image as a non-white background candidate image; inputting the obtained non-white background candidate images into a multimodal large model; performing semantic classification on the images using preset hierarchical prompts; the hierarchical prompts at least instruct the multimodal large model to determine whether the image belongs to the category of puzzle, infographic, size image, packaging, texture, lifestyle scene image, or creative studio image according to priority order, and outputting formatted data containing the classification results; based on the classification results, extracting images classified as lifestyle scene images or creative studio images that do not contain superimposed text, and outputting them as target product scene images; this method can significantly reduce the cost of calling large models, reduce API call fees, and reduce system response latency.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method, apparatus, device, and medium for screening product scene images. Background Technology

[0002] In the fields of e-commerce and AI-generated content (AIGC), high-quality "product scene images" are a crucial data foundation for training image generation models (such as LoRA and ControlNet) and building e-commerce shopping guide systems. Product scene images refer to pictures showcasing products in real-world environments or creative settings, distinct from traditional white-background product images. Currently, the conventional method for obtaining this type of data is to scrape it from massive amounts of e-commerce product images.

[0003] However, the original e-commerce product image dataset is filled with a lot of noise, and its direct use has the following drawbacks: (i) Low data cleaning efficiency. Traditional methods rely on manual screening, which is too costly and extremely inefficient when dealing with millions of data points, making it difficult to meet the needs of large-scale data production.

[0004] (ii) White background image obfuscation. A large number of product images in e-commerce datasets are pure white background images. These images lack light and shadow and background information. If they are mixed into the training set, it will reduce the generative model's ability to understand the scene and affect the model's performance in practical applications.

[0005] (iii) Non-target image interference. Data sets often contain non-target images such as jigsaw puzzles, images with a lot of marketing text (commonly known as "psoriasis ads"), packaging detail images, and ruler measurement images. Existing simple classification models have difficulty accurately distinguishing between "clean scene images" and "scene images with text interference," especially failing to accurately identify whether the text in the image belongs to "product text" or "marketing overlay text," resulting in a high misclassification rate.

[0006] (iv) Waste of computational resources. If a large model is directly called to filter all images, it will lead to a huge waste of computing power and API call costs, since white background images account for a large proportion of the dataset, which is not economically feasible.

[0007] Therefore, how to efficiently and cost-effectively extract high-quality product scene images from massive e-commerce images while ensuring the accuracy of the filtering is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0008] The technical problem to be solved by the present invention is to provide a method, apparatus, device and medium for filtering product scene images, which can significantly reduce the cost of calling large models, reduce API call fees, and reduce system response latency.

[0009] In a first aspect, the present invention provides a method for filtering product scene images, comprising the following steps: Step S1: Obtain the image set to be processed, perform edge pixel sampling on each image in the image set to be processed, calculate the proportion of white pixels in the edge region of the image, if the proportion of white pixels is greater than or equal to a preset first threshold, then mark the image as a white background image and remove it; otherwise, mark the image as a non-white background candidate image. Step S2: Input the obtained non-white background candidate images into the multimodal large model, and perform semantic classification on the images through preset hierarchical prompt words. The hierarchical prompt words at least instruct the multimodal large model to determine whether the image belongs to the category of puzzle, infographic, size image, packaging, texture, life scene image or creative studio image according to priority order, and output formatted data containing the classification results. Step S3: Based on the classification results, extract the images classified as either "Life Scene Image" or "Creative Studio Image" that do not contain superimposed text, and output them as the target product scene images.

[0010] Secondly, the present invention provides a product scene image filtering device, comprising: The preliminary screening module obtains a set of images to be processed, performs edge pixel sampling on each image in the set, calculates the proportion of white pixels in the edge region of the image, and if the proportion of white pixels is greater than or equal to a preset first threshold, the image is marked as a white background image and removed; otherwise, the image is marked as a non-white background candidate image. The fine screening module inputs the obtained non-white background candidate images into the multimodal large model, performs semantic classification on the images through preset hierarchical prompt words, and the hierarchical prompt words at least instruct the multimodal large model to determine whether the image belongs to the category of puzzle, information image, size image, packaging, texture, life scene image or creative studio image according to priority order, and outputs formatted data containing the classification results. The output module extracts images classified as either "life scene images" or "creative studio images" without overlaying text, based on the classification results, and outputs them as target product scene images.

[0011] Thirdly, the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in the first aspect.

[0012] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the first aspect.

[0013] One or more technical solutions provided by this invention have at least the following technical effects or advantages: 1. This invention uses Python scripts to quickly remove the vast majority of white background images, sending only a small number of complex images to the large model, which greatly reduces API call costs and processing time.

[0014] 2. This invention adopts a hierarchical prompt word logic, which prioritizes excluding puzzle images and text images, effectively solving the pain point of traditional models that have difficulty distinguishing between "scene images with text" and "clean scene images".

[0015] 3. The strict text detection protocol of this invention ensures that the final obtained scene image is free from annoying advertisements and is directly adapted to the training requirements of AIGC models.

[0016] 4. The JSON data output by this invention includes the reasoning process and detailed attributes, which facilitates subsequent data management and strategy adjustment.

[0017] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0018] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0019] Figure 1 This is a flowchart of the method in Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the device in Embodiment 2 of the present invention. Detailed Implementation

[0020] The overall concept of the technical solution in this application is as follows: This application provides a product scene image filtering method based on edge detection and multimodal large model collaboration, aiming to efficiently and accurately clean high-quality product scene images without text and without white backgrounds from massive e-commerce images. The main steps include: Step S1: Rapid Coarse Screening of White Background Image Based on Edge Pixel Sampling To reduce the cost of calling large models later, we first use a rule-based algorithm based on computer vision to remove obvious white background images.

[0021] Edge sampling: The system reads the image to be processed and extracts pixel data from the edges of the image. The edge sampling thickness is set as a percentage of the shortest side of the image (e.g., 5%).

[0022] White pixel determination: Traverse the pixels in the edge region and calculate the RGB value of each pixel. If the R, G, and B channel values ​​of a pixel are all greater than a preset brightness threshold (e.g., 245), then the pixel is determined to be a "white pixel".

[0023] Proportional threshold filtering: Calculate the proportion of "white pixels" in the edge region. If the proportion exceeds the preset white background threshold (e.g., 97%), the image is determined to be a "white background image" and is directly rejected; otherwise, it is marked as a "non-white background candidate image" and proceeds to the next stage.

[0024] Step S2: Semantic Fine Screening Based on Multimodal Large Model The "non-white background candidate images" selected in step S1 are input into a multimodal large language model (MLLM, such as Nova, GPT-4o, etc.) and classified in detail using preset hierarchical prompts.

[0025] Pre-computation Protocols: Instruct the model to prioritize detecting the presence of "overlay text" (such as watermarks and promotional labels) and "real human" (distinguishing between real people and printed faces) in images.

[0026] Hierarchical classification logic: The model strictly classifies images according to priority order, from highest to lowest: Collage: Recognizes and stitches multiple images together in a split screen or grid pattern.

[0027] Infographic / Selling Point Image: Identifies images containing marketing copy, watermarks, or graphic design elements.

[0028] Dimension: Identifies images containing rulers, measuring tapes, or dimension arrows.

[0029] Packaging: Identifies a flat inventory image of the product, its packaging, and accessories.

[0030] Texture / Close-up: Identifies images that only show material details.

[0031] Main Image: A secondary check to confirm whether it is a white background image that was missed.

[0032] Lifestyle: Identifies products that are located in real-life spaces or where there is human interaction (RealScene).

[0033] Creative Studio: Identifies photographs with non-white backgrounds, artistic backgrounds, or solid color backgrounds.

[0034] Target extraction: Only retain images categorized as "Lifestyle" and "Creative Studio" that do not contain overlaid text.

[0035] Step S3: Structured output of classification results and manual verification The model outputs standardized JSON data, including category labels, confidence scores, and attribute analysis (such as whether there are people or text).

[0036] The system automatically distributes images to different categories of folders based on the JSON results.

[0037] Human-in-the-loop: For images with low model confidence or specific categories (such as "life scene images"), a manual sampling process is introduced to quickly verify whether there are any misjudgments and to feed back the correction results to optimize the prompt words.

[0038] This embodiment provides a screening system based on the combination of Python scripts and the Nova large model.

[0039] 1. Initial screening module (removal of images with white background) The system runs Python scripts based on the NumPy and PIL image processing libraries.

[0040] Input: The folder path containing mixed product images.

[0041] Core algorithm logic: ; Where Ⅱ is an indicator function, which is 1 if the condition is met; otherwise it is 0.

[0042] Execution flow: The script recursively traverses the folder, automatically processes images with alpha channels (RGBA) (compositing them with a white background), and calculates the percentage of white at the edges. If Ratio... white If the value is ≥0.97, it is classified as white_bg; otherwise, it is classified as not_white_bg.

[0043] Output: Generates a list of not_white_bg images, which serve as input to the fine screening module.

[0044] 2. Fine Screening Module (Scene Graph Extraction) The system calls the Nova multimodal large model interface and loads the system prompt word defined below: Role Definition: Senior E-commerce Visual Data Analyst.

[0045] Testing tasks: Text detection: Distinguish between "product text" (retained) and "marketing overlay text" (removed). For example, the logo on the bottle is not considered an impurity, but the "Sale" icon in the center of the image is considered an impurity.

[0046] Person detection: Distinguish between "real human limbs" (retained as Lifestyle) and "printed faces / mannequins" (ignore).

[0047] Classification Decision Tree: If multiple image stitching is detected, output [Collage]; If overlaid marketing text is detected, output [Infographic]; If a measuring tool is detected, output [Dimension]; ... (and so on) If the background is realistic and there are character interactions, output [Lifestyle]; If the background is an artistic one with no depth of field, output [Creative Studio].

[0048] Output format: 3. Post-processing and manual verification The system parses the returned JSON and moves images marked as "Lifestyle" and "Creative Studio" with "has_overlay_text" set to false to the final dataset directory. Staff only need to quickly browse and review this directory, eliminating the need to examine each image in the original dataset, thus achieving efficient image cleaning.

[0049] Example 1 like Figure 1 As shown, this embodiment provides a method for filtering product scene images, characterized by the following steps: Step S1: Obtain the image set to be processed, perform edge pixel sampling on each image in the image set to be processed, calculate the proportion of white pixels in the edge region of the image, if the proportion of white pixels is greater than or equal to a preset first threshold, then mark the image as a white background image and remove it; otherwise, mark the image as a non-white background candidate image. Step S2: Input the obtained non-white background candidate images into the multimodal large model, and perform semantic classification on the images through preset hierarchical prompt words. The hierarchical prompt words at least instruct the multimodal large model to determine whether the image belongs to the category of puzzle, infographic, size image, packaging, texture, life scene image or creative studio image according to priority order, and output formatted data containing the classification results. Step S3: Based on the classification results, extract the images classified as either "Life Scene Image" or "Creative Studio Image" that do not contain superimposed text, and output them as the target product scene images.

[0050] In this embodiment, preferably, the edge pixel sampling specifically includes: extracting the edge regions around the image, wherein the sampling thickness of the edge regions is a preset proportion of the shortest side length of the image; in a preferred embodiment, the sampling thickness of the edge regions is set to 5% of the shortest side length of the image. Let the width of the image be W pixels and the height be H pixels, then the shortest side length S = min(W, H), and the sampling thickness T = S × 5%. The edge regions include: Top edge: A rectangular area from (0,0) to (W,T); Bottom edge: A rectangular area from (0,HT) to (W,H); Left edge: A rectangular region from (0,T) to (T,HT); Right edge: A rectangular region from (WT,T) to (W,HT); Those skilled in the art will understand that the preset ratio can be adjusted according to the actual application scenario, for example, set to 3%, 5% or 10%, all of which can achieve the technical effect of the present invention.

[0051] In this embodiment, preferably, the white pixel ratio is calculated as follows: the number of pixels in the edge region that meet the white pixel determination criteria is counted and divided by the total number of pixels in the edge region; the white pixel determination criteria is that the RGB value of the pixel meets the preset brightness threshold requirement; the white pixel determination criteria is that the RGB three channel values ​​of the pixel are all greater than the preset second threshold; or the maximum RGB value of the pixel is greater than the preset second threshold.

[0052] In this embodiment, preferably, the hierarchical prompt words also instruct the multimodal large model to perform the following operations: Detect whether there is superimposed text in the image, and distinguish between the product text and the marketing text superimposed on it; Detect the presence of real people in an image and distinguish between real human limbs and printed faces or mannequins; Images with detected marketing overlay text are categorized as infographics and removed, while images with detected real human bodies are categorized as everyday scene images and retained. The formatted data includes at least a category label, an attribute identifier indicating whether it contains overlay text, and an attribute identifier indicating whether it contains real people.

[0053] In this embodiment, preferably, the priority order from high to low is as follows: puzzle, infographic, size chart, packaging, texture, everyday scene, and creative studio.

[0054] In this embodiment, preferably, step S3 further includes: storing the target product scene images into corresponding folders according to their classification categories, and generating a classification report.

[0055] In this embodiment, preferably, it further includes step S4: obtaining the classification confidence score output by the multimodal large model, manually sampling images with a classification confidence score lower than a preset third threshold, and feeding back the manual correction results to the hierarchical prompt words of the multimodal large model for optimization.

[0056] Based on the same inventive concept, this application also provides an apparatus corresponding to the method in Embodiment 1, as detailed in Embodiment 2.

[0057] Example 2 like Figure 2 As shown, this embodiment provides a product scene image filtering device, including: The preliminary screening module obtains a set of images to be processed, performs edge pixel sampling on each image in the set, calculates the proportion of white pixels in the edge region of the image, and if the proportion of white pixels is greater than or equal to a preset first threshold, the image is marked as a white background image and removed; otherwise, the image is marked as a non-white background candidate image. The fine screening module inputs the obtained non-white background candidate images into the multimodal large model, performs semantic classification on the images through preset hierarchical prompt words, and the hierarchical prompt words at least instruct the multimodal large model to determine whether the image belongs to the category of puzzle, information image, size image, packaging, texture, life scene image or creative studio image according to priority order, and outputs formatted data containing the classification results. The output module extracts images classified as either "life scene images" or "creative studio images" without overlaying text, based on the classification results, and outputs them as target product scene images.

[0058] In this embodiment, preferably, the edge pixel sampling specifically includes: extracting the edge regions around the image, wherein the sampling thickness of the edge regions is a preset ratio of the shortest side length of the image.

[0059] In this embodiment, preferably, the white pixel ratio is calculated as follows: the number of pixels in the edge region that meet the white pixel determination criteria is counted and divided by the total number of pixels in the edge region; the white pixel determination criteria is that the RGB value of the pixel meets the preset brightness threshold requirement; the white pixel determination criteria is that the RGB three channel values ​​of the pixel are all greater than the preset second threshold; or the maximum RGB value of the pixel is greater than the preset second threshold.

[0060] In this embodiment, preferably, the hierarchical prompt words also instruct the multimodal large model to perform the following operations: Detect whether there is superimposed text in the image, and distinguish between the product text and the marketing text superimposed on it; Detect the presence of real people in an image and distinguish between real human limbs and printed faces or mannequins; Images with detected marketing overlay text are categorized as infographics and removed, while images with detected real human bodies are categorized as everyday scene images and retained. The formatted data includes at least a category label, an attribute identifier indicating whether it contains overlay text, and an attribute identifier indicating whether it contains real people.

[0061] In this embodiment, preferably, the priority order from high to low is as follows: puzzle, infographic, size chart, packaging, texture, everyday scene, and creative studio.

[0062] In this embodiment, preferably, the output module further includes: storing the target product scene images into corresponding folders according to the classification categories, and generating a classification report.

[0063] In this embodiment, preferably, it also includes a prompt word adjustment module, which obtains the classification confidence score output by the multimodal large model, performs manual sampling inspection on images with a classification confidence score lower than a preset third threshold, and feeds back the manual correction results to the hierarchical prompt words of the multimodal large model for optimization.

[0064] Since the apparatus described in Embodiment 2 of the present invention is an apparatus used to implement the method of Embodiment 1 of the present invention, those skilled in the art can understand the specific structure and variations of the apparatus based on the method described in Embodiment 1 of the present invention, and therefore will not be described again here. All apparatuses used in the method of Embodiment 1 of the present invention fall within the scope of protection of the present invention.

[0065] Based on the same inventive concept, this application provides an electronic device embodiment corresponding to Embodiment 1, as detailed in Embodiment 3.

[0066] Example 3 This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it can implement any of the implementation methods in Embodiment 1.

[0067] Since the electronic device described in this embodiment is the device used to implement the method in Embodiment 1 of this application, those skilled in the art can understand the specific implementation method and various variations of the electronic device in this embodiment based on the method described in Embodiment 1 of this application. Therefore, how the electronic device implements the method in the embodiment of this application will not be described in detail here. Any device used by those skilled in the art to implement the method in the embodiment of this application falls within the scope of protection of this application.

[0068] Based on the same inventive concept, this application provides a storage medium corresponding to Embodiment 1, as detailed in Embodiment 4.

[0069] Example 4 This embodiment provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it can implement any of the implementation methods in Embodiment 1.

[0070] The technical solutions provided in this application embodiment have at least the following technical effects or advantages: 1. In this embodiment, the rule-based algorithm in step S1 is used for initial screening, pre-removing white-background images, which typically account for 70%-80%, and only inputting non-white-background candidate images into the multimodal large model for fine screening. Compared to the prior art's method of directly calling the large model for all images, this invention can significantly reduce the cost of calling the large model, reduce API call fees, and reduce system response latency.

[0071] 2. The edge detection algorithm in this application is based on traditional image processing technology, which has fast computation speed and low resource consumption. It adopts a two-stage collaborative architecture of "rule-based algorithm initial screening + large model fine screening", which enables the computationally intensive large model to be called only for a small-scale candidate set, thus greatly improving the overall processing throughput.

[0072] 3. In existing technologies, general image classification models struggle to distinguish between "product text" (such as bottle logos, which should be retained) and "marketing overlay text" (such as promotional labels, which should be removed). This application's embodiment, through a pre-calculation protocol in preset hierarchical prompts, explicitly instructs the multimodal large model to distinguish between the two types of text, achieving accurate identification and removal of overlay text and avoiding the accidental deletion of product images containing legitimate text.

[0073] 4. In existing technologies, models are prone to misclassifying printed faces and mannequins as real people, leading to classification errors. This application's embodiments explicitly distinguish between "real human limbs" and "printed faces / mannequins" using prompts, ensuring that images containing only real human interactions are classified as everyday scene images, thus improving classification accuracy.

[0074] 5. The embodiments of this application adopt a hierarchical classification logic from high to low priority (puzzle class → infographic class → size image class → packaging class → texture class → life scene image class → creative studio image class), which solves the classification conflict problem of multi-labeled images and ensures that each image is assigned to a unique and most appropriate category.

[0075] 6. This application's embodiments introduce a manual sampling mechanism to manually review images with low classification confidence levels, and feed the correction results back to the prompt words for optimization. This mechanism enables the system to continuously iterate and improve, gradually increasing classification accuracy and adapting to the ever-changing characteristics of e-commerce images.

[0076] 7. In the embodiments of this application, key parameters such as edge sampling thickness, white pixel threshold, white background ratio threshold, and classification confidence threshold can be flexibly configured according to actual business scenarios, and have good adaptability and scalability.

[0077] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0078] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0079] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0080] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1The steps of the function specified in one or more boxes.

[0081] While specific embodiments of the present invention have been described above, those skilled in the art should understand that the specific embodiments described are merely illustrative and not intended to limit the scope of the present invention. Equivalent modifications and variations made by those skilled in the art in accordance with the spirit of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A merchandise scene graph screening method, characterized by, Includes the following steps: Step S1: Obtain the image set to be processed, perform edge pixel sampling on each image in the image set to be processed, calculate the proportion of white pixels in the edge region of the image, if the proportion of white pixels is greater than or equal to a preset first threshold, then mark the image as a white background image and remove it; otherwise, mark the image as a non-white background candidate image. Step S2: Input the obtained non-white background candidate images into the multimodal large model, and perform semantic classification on the images through preset hierarchical prompt words. The hierarchical prompt words at least instruct the multimodal large model to determine whether the image belongs to the category of puzzle, infographic, size image, packaging, texture, life scene image or creative studio image according to priority order, and output formatted data containing the classification results. Step S3: Based on the classification results, extract the images classified as either "Life Scene Image" or "Creative Studio Image" that do not contain superimposed text, and output them as the target product scene images.

2. The method according to claim 1, characterized in that, The edge pixel sampling specifically includes: extracting the edge regions around the image, wherein the sampling thickness of the edge regions is a preset ratio of the shortest side length of the image.

3. The method according to claim 1, characterized in that, The white pixel ratio is calculated as follows: the number of pixels in the edge region that meet the white pixel determination criteria is counted and divided by the total number of pixels in the edge region; the white pixel determination criteria is that the RGB value of the pixel meets the preset brightness threshold requirement; the white pixel determination criteria is that the RGB three channel values ​​of the pixel are all greater than the preset second threshold; or the maximum RGB value of the pixel is greater than the preset second threshold.

4. The method according to claim 1, characterized in that, The hierarchical prompts also instruct the multimodal large model to perform the following operations: Detect whether there is superimposed text in the image, and distinguish between the product text and the marketing text superimposed on it; Detect whether there are real people in the image and distinguish between real human limbs and printed faces or mannequins; Images with detected marketing overlay text are categorized as infographics and removed, while images with detected real human bodies are categorized as everyday scene images and retained. The formatted data includes at least a category label, an attribute identifier indicating whether it contains overlay text, and an attribute identifier indicating whether it contains real people.

5. The method according to claim 1, characterized in that, The priority order from highest to lowest is as follows: puzzle, infographic, size chart, packaging, texture, everyday scene, and creative studio.

6. The method according to claim 1, characterized in that, Step S3 also includes: storing the target product scene images into corresponding folders according to the category, and generating a category report.

7. The method according to claim 1, characterized in that, It also includes step S4: obtaining the classification confidence score output by the multimodal large model, manually sampling images with a classification confidence score lower than a preset third threshold, and feeding back the manual correction results to the hierarchical prompt words of the multimodal large model for optimization.

8. A product scene image screening device, characterized in that: include: The preliminary screening module obtains a set of images to be processed, performs edge pixel sampling on each image in the set, calculates the proportion of white pixels in the edge region of the image, and if the proportion of white pixels is greater than or equal to a preset first threshold, the image is marked as a white background image and removed; otherwise, the image is marked as a non-white background candidate image. The fine screening module inputs the obtained non-white background candidate images into the multimodal large model, performs semantic classification on the images through preset hierarchical prompt words, and the hierarchical prompt words at least instruct the multimodal large model to determine whether the image belongs to the category of puzzle, information image, size image, packaging, texture, life scene image or creative studio image according to priority order, and outputs formatted data containing the classification results. The output module extracts images classified as either "life scene images" or "creative studio images" without overlaying text, based on the classification results, and outputs them as target product scene images.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 7.

Images