An image processing method and system for an indoor design scene

Abstract

The application relates to the technical field of image processing, in particular to an image processing method and system for an indoor design scene. The method comprises the following steps: in response to a search request of a user, determining a target style label, mapping the target style label into a multi-dimensional aesthetic weight vector, the multi-dimensional aesthetic weight vector at least including weight values corresponding to order sense, precision sense, artistic conception sense and innovation sense dimensions; acquiring a personalized aesthetic preference vector of the user, and fusing the multi-dimensional aesthetic weight vector and the personalized aesthetic preference vector to generate a search query vector; performing matching in a search candidate set based on the search query vector, sorting candidate images according to a matching result, and outputting a search result; collecting feedback behaviors of the user on the search result, updating the personalized aesthetic preference vector according to the feedback behaviors, and storing the updated vector into historical behavior data of the user. The application can improve search efficiency and accuracy, and meet the acquisition of high-quality inspiration in the modern indoor design industry.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to an image processing method and system for interior design scenarios. Background Technology

[0002] Interior design plays a vital role in enhancing the functionality and artistic atmosphere of a space. Throughout the design process, from concept and planning to construction drawings, designers rely heavily on inspirational reference images to aid decision-making. As design styles become increasingly diverse, the industry demands higher efficiency and accuracy in acquiring and matching inspirational images.

[0003] Currently, the main ways to obtain inspiration include the following: 1. Image retrieval based on keyword tags (such as Huaban.com and Pinterest) relies on manual annotation for text matching, making it difficult to deeply analyze the differences in multiple aesthetic dimensions such as order, craftsmanship, artistic conception, and innovation of images; 2. General image search systems (such as Baidu Image Search) are based on common visual features for retrieval, but they do not model interior design scenarios and cannot take into account both project stages and personalized aesthetic needs.

[0004] 3. General recommendation algorithms rely on collaborative filtering mechanisms, lack consideration for the aesthetic dimensions of interior design professionals, and have limited retrieval accuracy.

[0005] Therefore, given the shortcomings of existing technologies, how to design an interior design inspiration image processing method that can improve retrieval efficiency and accuracy and meet the urgent need of the modern interior design industry for high-quality inspiration is a technical problem that urgently needs to be solved. Summary of the Invention

[0006] The purpose of this application is to overcome the above-mentioned technical problems and provide an image processing method and system for interior design scenarios, which can improve retrieval efficiency and accuracy and meet the needs of the modern interior design industry for obtaining high-quality inspiration.

[0007] Firstly, one embodiment of this application discloses an image processing method for interior design scenarios, employing the following scheme: An image processing method for interior design scenarios includes: responding to a user's search request, determining a target style tag, mapping the target style tag to a multi-dimensional aesthetic weight vector, wherein the multi-dimensional aesthetic weight vector includes at least weight values ​​corresponding to the dimensions of order, craftsmanship, artistic conception, and innovation; obtaining the user's personalized aesthetic preference vector, and fusing the multi-dimensional aesthetic weight vector with the personalized aesthetic preference vector to generate a search query vector; matching the search query vector in a search candidate set, and sorting the candidate images according to the matching results, and outputting the search results; collecting the user's feedback behavior on the search results, updating the personalized aesthetic preference vector according to the feedback behavior, and storing the updated vector in the user's historical behavior data.

[0008] By adopting the above technical solutions, this approach addresses the problems of crude retrieval methods, low aesthetic matching, and lack of personalization in the traditional design process when designers obtain inspiration reference images. It maps target style tags into a multi-dimensional aesthetic weight vector containing weights for dimensions such as order, craftsmanship, artistic conception, and innovation, accurately grasping the user's style needs. To address the issue of existing platforms failing to recognize designers' personalized aesthetic preferences, the multi-dimensional aesthetic weight vector is fused with the user's personalized aesthetic preference vector to generate a retrieval query vector, making the retrieval results more aligned with the user's individual aesthetic preferences. Furthermore, to address the low retrieval efficiency of existing platforms, the retrieval results are matched and sorted within the retrieval candidate set based on the retrieval query vector, improving both accuracy and efficiency. Finally, to address the issue of existing platforms failing to use user behavior to inform image library construction and model training, the approach collects and stores user feedback on retrieval results to update personalized aesthetic preference vectors, enabling effective utilization of user behavior data, promoting continuous system optimization, and providing designers with image retrieval services that better meet their needs.

[0009] Optionally, before determining the target style tag in response to the user's search request and mapping the target style tag to a multidimensional aesthetic weight vector, the method further includes: acquiring a candidate interior design image set, and performing multidimensional aesthetic evaluation on each image in the candidate interior design image set using an aesthetic scoring model to filter out images that meet preset aesthetic standards and store them in a vertical inspiration image library to construct the search candidate set; wherein, the acquisition method of the candidate interior design image set includes at least one of batch import, targeted supplementary acquisition, and user upload; the multidimensional aesthetic evaluation includes at least validity detection, multidimensional aesthetic scoring, horizontal comparison screening, and manual screening and review; the horizontal comparison screening is a comparison between multiple images in the same batch in the candidate interior design image set.

[0010] By adopting the above technical solution, a candidate interior design image set is obtained. An aesthetic scoring model is used to perform multi-dimensional aesthetic evaluation on each image in the candidate interior design image set, which can filter out images that meet preset aesthetic standards, thereby improving the quality rate of the images in the database, making the quality rate significantly higher than that of general platforms. The filtered images that meet the preset aesthetic standards are stored in a vertical inspiration image library to construct a retrieval candidate set, forming a targeted vertical domain image library, providing high-quality image resources for subsequent searches. The candidate interior design image set is obtained through various methods such as batch import, targeted supplementary acquisition, and user uploads, which can enrich the image sources and ensure the diversity and sufficiency of the image library. The multi-dimensional aesthetic evaluation includes validity detection, multi-dimensional aesthetic scoring, horizontal comparison screening, and manual screening and review. Through multiple rounds of evaluation, the aesthetic quality of the images in the database can be effectively guaranteed, establishing a quality threshold for the vertical image library. The horizontal comparison screening uses a method of comparing multiple images in the same batch, which can rank the images in the same batch and eliminate the lowest proportion of images, further improving the quality of the images in the database.

[0011] Optionally, the step of performing multi-dimensional aesthetic evaluation on each image in the candidate interior design image set using an aesthetic scoring model includes: performing validity detection on the images in the candidate interior design image set based on a preset sample prototype to obtain a valid first target design image set; extracting image features from the images in the first target design image set and performing multi-dimensional aesthetic scoring on the image features of each image using multiple independent scoring heads to obtain a corresponding multi-dimensional aesthetic scoring vector, wherein the multi-dimensional aesthetic scoring vector includes at least scores for order, craftsmanship, artistic conception, and innovation; performing horizontal comparison screening on the images in the first target design image set and eliminating the lower-ranked image set according to a preset elimination ratio to obtain a second target design image set; performing manual screening and review on the images in the second target design image set, storing the reviewed images in the vertical inspiration image library to construct the retrieval candidate set, and using the review results to update the aesthetic scoring model, wherein the review results include labels of pass, rejection, or doubt for each image.

[0012] By adopting the above technical solutions, validity testing of the candidate interior design image set can filter out non-interior design images and obtain a valid image set; multi-dimensional aesthetic scoring can obtain scores for images in dimensions such as sense of order, craftsmanship, artistic conception, and innovation; horizontal comparison screening can eliminate images with lower rankings and improve the overall image quality; manual screening and review further ensure the aesthetic quality of the images in the database, and the review results are used to update the aesthetic scoring model, which can make the model continuously evolve with the accumulation of labeled data, improve the quality of automatic image scoring, reduce the workload of manual review, and at the same time build a vertical inspiration image library with a significantly higher rate of high-quality images than general platforms as a search candidate set.

[0013] Optionally, the step of matching the retrieval query vector in the retrieval candidate set, sorting the candidate images according to the matching results, and outputting the retrieval results includes: identifying the current project design stage of the user, wherein the project design stage includes at least the concept stage, scheme stage, detailed stage, and construction drawing stage; constraining the scope of the retrieval candidate set according to the project design stage, wherein the images in the retrieval candidate set have a pre-established mapping relationship with the project design stage; calculating the similarity between the retrieval query vector and the candidate images in the constrained retrieval candidate set, sorting the candidate images according to the similarity, and outputting the retrieval results.

[0014] By adopting the above technical solution, the current project design stage of the user can be identified and the scope of the search candidate set can be constrained accordingly, so that the search results can better meet the needs of the user at different design stages. Calculating the similarity between the search query vector and the candidate image in the constrained search candidate set and sorting the results can improve the accuracy and relevance of the search results, providing users with interior design inspiration images that better meet their needs.

[0015] Optionally, the step of performing validity detection on the images in the candidate interior design image set based on a preset sample prototype to obtain a valid first target design image set includes: extracting feature information from each image in the candidate interior design image set to calculate the similarity with the preset sample prototype and obtain a validity score for each image, wherein the feature information includes global features and local semantic features, and the local semantic features include at least spatial layout, furniture components, wall-ceiling-floor relationship, and lighting structure; and based on the comparison result of the validity score and a preset validity score threshold, eliminating invalid images and obtaining a valid image set as the first target design image set.

[0016] By adopting the above technical solution, the validity of the candidate interior design image set can be detected by using a preset sample prototype. The validity score is obtained by extracting image feature information and calculating the similarity with the preset sample prototype. Then, invalid images are removed by combining the preset valid score threshold, thereby obtaining a valid first target design image set. This ensures that the images that subsequently enter the multi-dimensional aesthetic evaluation process are all valid interior design images, thereby improving the quality rate of images entering the database.

[0017] Optionally, before obtaining the user's personalized aesthetic preference vector and fusing the multidimensional aesthetic weight vector with the personalized aesthetic preference vector to generate a retrieval query vector, the method further includes: receiving user input to adjust the weight values ​​of at least one dimension of the multidimensional aesthetic weight vector to update the multidimensional aesthetic weight vector and associate it with the target style tag and save it.

[0018] By adopting the above technical solution, the system receives user input to adjust the weight values ​​of at least one dimension of the multidimensional aesthetic weight vector, allowing users to flexibly change the weights according to their own needs, making the multidimensional aesthetic weight vector more in line with the user's actual needs; the system updates the multidimensional aesthetic weight vector and saves it in association with the target style tag, making it convenient for users to quickly call up the weight configuration that matches their own preferences, thereby improving retrieval efficiency and accuracy.

[0019] Optionally, it also includes: monitoring the supply-demand ratio of images corresponding to different style tags, and triggering a targeted supplementary acquisition task for the style tag when it is detected that the supply of images corresponding to any style tag is insufficient.

[0020] By adopting the above technical solution, the supply and demand of images corresponding to different style tags can be monitored. When insufficient supply is detected, a targeted replenishment task is triggered, so that the image library can be automatically expanded and optimized according to user behavior, avoiding the absence of images of specific styles, ensuring that the image library covers a sufficient number of images of various styles, and forming a positive flywheel that gets better with use.

[0021] Optionally, the feedback behavior includes positive feedback behavior and negative feedback behavior, wherein the positive feedback behavior includes at least favorite, long stay and download, and the negative feedback behavior includes at least quick skip and ignore.

[0022] By adopting the above technical solution, the types of feedback behaviors are clarified, and the personalized aesthetic preference vector can be updated more accurately based on the user's positive feedback (collection, long stay and download) and negative feedback (quick skip and ignore) behaviors. This allows the system to better capture the user's aesthetic preferences and thus provide the user with search results that better meet their personalized needs.

[0023] Secondly, one embodiment of this application discloses an intelligent retrieval and aesthetic evaluation system for inspirational image sets, which adopts the following scheme: A smart image retrieval and aesthetic evaluation system for inspirational images includes: a retrieval response module, used to respond to a user's retrieval request, determine a target style tag, and map the target style tag into a multi-dimensional aesthetic weight vector, wherein the multi-dimensional aesthetic weight vector includes at least weight values ​​corresponding to the dimensions of order, craftsmanship, artistic conception, and innovation; a retrieval query vector generation module, used to obtain the user's personalized aesthetic preference vector, and fuse the multi-dimensional aesthetic weight vector with the personalized aesthetic preference vector to generate a retrieval query vector; a retrieval result output module, used to match the retrieval query vector in a retrieval candidate set, sort the candidate images according to the matching results, and output the retrieval results; and a feedback behavior acquisition module, used to acquire the user's feedback behavior on the retrieval results, update the personalized aesthetic preference vector according to the feedback behavior, and store the updated vector in the user's historical behavior data.

[0024] By adopting the above technical solution, the target style tags corresponding to user search requests can be mapped into a multi-dimensional aesthetic weight vector containing weight values ​​for dimensions such as order, craftsmanship, artistic conception, and innovation. This vector is then fused with the user's personalized aesthetic preference vector to generate a search query vector. The search results are matched and sorted in the search candidate set and output. Furthermore, user feedback behavior can be collected to update and store the personalized aesthetic preference vector, thereby meeting users' personalized search needs, improving the matching degree between search results and user aesthetics, and enhancing the user experience.

[0025] Thirdly, one embodiment of this application discloses an electronic device that adopts the following solution: An electronic device includes 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 computer program, it implements the method described in any one of the above-mentioned methods.

[0026] In summary, this application includes at least one of the following beneficial technical effects: 1. By mapping target style tags to multi-dimensional aesthetic weight vectors and combining them with user-personalized aesthetic preference vectors to generate search query vectors for matching and ranking, this approach can meet users' personalized aesthetic needs, solve the problem of existing platforms being generic and lacking personalization, and improve search accuracy. Furthermore, by receiving user input to adjust the weight values ​​of at least one dimension of the multi-dimensional aesthetic weight vector, updating the multi-dimensional aesthetic weight vector, and saving it in association, this approach allows users to fine-tune search conditions and meet the diverse needs of different users. 2. Identify the current project design stage of the user and constrain the scope of the search candidate set according to the stage. This can take into account the needs of the project stage, solve the problem of the lack of design stage awareness in the existing platform, and improve search efficiency. 3. The images entering the database undergo a multi-dimensional aesthetic evaluation process, including validity testing, multi-dimensional aesthetic scoring, horizontal comparison screening, and manual review. This ensures the aesthetic quality of the images entering the database and solves the problem of existing platforms lacking aesthetic quality screening, evaluation, and elimination mechanisms. As a result, the rate of high-quality images entering the database is significantly higher than that of general platforms. 4. Acquire candidate interior design image sets through batch import, targeted supplementary acquisition, and user uploads. After multi-dimensional aesthetic evaluation, store them in a vertical inspiration image library. This can build a self-growing image database for vertical fields, solving the problem of existing platforms lacking vertical image library accumulation. It can also monitor the supply and demand ratio of images corresponding to different style tags. When insufficient supply is detected, a targeted supplementary acquisition task is triggered to ensure a sufficient supply of images of various styles in the image library, so that the image library can automatically expand and optimize according to user behavior. 5. Collecting user feedback on search results updates the personalized aesthetic preference vector, and the review results of multi-dimensional aesthetic evaluation update the aesthetic scoring model. This enables the aesthetic scoring model to continuously evolve with the accumulation of labeled data and user behavior data, building a self-evolving closed loop and solving the problem of no data flywheel in existing platforms. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating an image processing method for interior design scenarios disclosed in an embodiment of this application; Figure 2 for Figure 1 A schematic diagram of the specific process of step S30 in an image processing method for interior design scenarios disclosed herein; Figure 3-5 This is a schematic diagram of the interface display for the system's automatic evaluation when performing multi-dimensional aesthetic evaluation on an image. Figure 6 This is a schematic diagram of the interface for manual review during multi-dimensional aesthetic evaluation of images. Figure 7-8 This is a schematic diagram of the interface display during the search operation; Figure 9 This is a schematic diagram of the structure of an intelligent retrieval and aesthetic evaluation system for an inspirational image collection disclosed in another embodiment of this application; Figure 10 This is a schematic diagram of the structure of an electronic device disclosed in another embodiment of this application. Detailed Implementation

[0028] The present application will be further described in detail below with reference to the accompanying drawings.

[0029] Embodiments of this application will now be described in more detail with reference to the accompanying drawings. While embodiments of this application are shown in the drawings, it should be understood that this application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to make this application more thorough and complete, and to fully convey the scope of this application to those skilled in the art.

[0030] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a” and “the” as used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0031] It should be understood that although the terms "first," "second," etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0032] The technical solutions of the embodiments of this application are described in detail below with reference to the accompanying drawings.

[0033] [First Embodiment] See Figure 1 The first embodiment of this application discloses an image processing method for interior design scenarios. By mapping target style tags to multidimensional aesthetic weight vectors and then fusing them with personalized aesthetic preference vectors, we can accurately match user needs and improve the relevance and accuracy of search results.

[0034] Specifically, the steps include the following: S10. In response to the user's search request, determine the target style tag and map the target style tag into a multi-dimensional aesthetic weight vector. In response to a user's search request, when determining the target style tag, the user can fill in the style type they want to search for on the display interface. For example, if the user requests to search for interior design images of "Japanese minimalist style," the system will determine "Japanese minimalist" as the target style tag. The target style tag can be various common interior design styles. Users can also directly select one or more of the keywords displayed on the display interface, such as "Nordic, Modern, Japanese, New Chinese, Industrial, Light Luxury, Wabi-sabi, Minimalist," as their style search terms.

[0035] In mapping target style tags to a multi-dimensional aesthetic weight vector, the multi-dimensional aesthetic weight vector includes at least the weight values ​​corresponding to the dimensions of order, craftsmanship, artistic conception, and innovation. Taking "Japanese minimalist style" as an example, the possible mapped multi-dimensional aesthetic weight vector is [order: 0.85, craftsmanship: 0.60, artistic conception: 0.90, innovation: 0.35]. These weight values ​​can be preset according to the characteristics of different styles and can be adjusted according to the actual situation.

[0036] S20. Obtain the user's personalized aesthetic preference vector, and fuse the multi-dimensional aesthetic weight vector with the personalized aesthetic preference vector to generate a retrieval query vector; When obtaining a user's personalized aesthetic preference vector, the system generates it based on the user's historical behavioral data. For example, if a user frequently collects images with a high sense of order and atmosphere, then the weight values ​​of these two dimensions in their personalized aesthetic preference vector will be relatively high. By fusing the multi-dimensional aesthetic weight vector with the personalized aesthetic preference vector, and adjusting the weight values ​​of order and atmosphere in the multi-dimensional aesthetic weight vector based on the personalized aesthetic preference vector, a search query vector is generated. This comprehensively considers stylistic characteristics and user preferences, making the search more accurate.

[0037] It is worth mentioning that the system can also receive user input to adjust the weight value of at least one dimension in the multidimensional aesthetic weight vector. For example, by sliding the slider on the interface, the [Innovation] value can be increased to 0.60 to update the multidimensional aesthetic weight vector and associate it with the currently selected target style tag to save it as "My Japanese Template" for direct retrieval in the next search.

[0038] S30. Match the candidate images in the retrieval candidate set based on the retrieval query vector, sort the candidate images according to the matching results, and output the retrieval results; Specifically, based on the retrieval query vector, multi-dimensional weight values ​​are matched in the retrieval candidate set. The candidate images are then sorted by relevance according to the matching results, with images of high relevance placed first, and the retrieval results are then output.

[0039] See Figure 2 It is worth mentioning that step S30, when executed, includes: S31. Identify the current project design stage the user is in; The project design phase includes at least the concept phase, scheme phase, detailed design phase, and construction drawing phase.

[0040] S32. Constrain the scope of the search candidate set according to the project design phase; Among them, the images in the candidate search set have a pre-established mapping relationship with the project design stage; S33. In the constrained retrieval candidate set, calculate the similarity between the retrieval query vector and the candidate images, sort the candidate images according to the similarity, and output the retrieval results.

[0041] Therefore, based on the above steps S31-S33, the system can filter the candidate set in combination with the current project stage (solution stage), and output the search results after fusing with personal preference vectors.

[0042] S40. Collect user feedback behavior on search results, update personalized aesthetic preference vector based on feedback behavior, and store the updated vector in the user's historical behavior data.

[0043] When collecting user feedback on search results, feedback includes both positive and negative behaviors. Positive feedback includes at least saving, prolonged engagement, and downloading, while negative feedback includes at least quickly skipping and ignoring. The system updates the personalized aesthetic preference vector based on these feedback behaviors and stores the updated vector in the user's historical behavior data. For example, if a user saves an image, the aesthetic dimension weight corresponding to that image will increase accordingly in the personalized aesthetic preference vector; the amount of this increase can be preset and is not limited here.

[0044] Furthermore, prior to step S10, the following is also included: S01. Obtain a candidate interior design image set, and perform multi-dimensional aesthetic evaluation on each image in the candidate interior design image set through an aesthetic scoring model to filter out images that meet the preset aesthetic standards and store them in a vertical inspiration image library to construct a retrieval candidate set. The acquisition of candidate interior design image sets can be achieved through at least one of the following methods: batch import, targeted supplementary acquisition, or user upload. For example, the platform operations team can import a large number of candidate interior design images in batches, or trigger targeted supplementary acquisition tasks when insufficient supply is detected based on the supply-demand ratio of images under different style tags. Users can also upload their own images. The aesthetic scoring model can extract image features based on convolutional neural networks or visual Transformers, followed by multiple fully connected layers as independent scoring heads.

[0045] The multidimensional aesthetic evaluation process includes at least validity testing, multidimensional aesthetic scoring, horizontal comparison screening, and manual review. Specifically: During validity testing, images in the candidate interior design image set are tested based on a preset sample prototype. Specifically, feature information is extracted from each image in the candidate interior design image set. The feature information includes global features and local semantic features. Local semantic features include at least spatial layout, furniture components, wall-ceiling-floor relationship, and lighting structure.

[0046] By calculating the similarity between these feature information and the preset sample prototype, a validity score is obtained for each image. Based on the comparison between the validity score S_valid and the preset validity score threshold T_valid, invalid images with S_valid < T_valid are removed, and the set of valid images is obtained as the first target design image set. For example, some product images or portrait images may not meet the requirements of interior design and will be removed.

[0047] When scoring in a multidimensional aesthetic sense, image features are extracted from the images in the first target design image set, and the image features of each image are scored in a multidimensional aesthetic sense using multiple independent scoring heads to obtain the corresponding multidimensional aesthetic score vector. The multidimensional aesthetic score vector includes at least scores for sense of order, sense of craftsmanship, sense of artistic conception, and sense of innovation.

[0048] During horizontal comparison and screening, images acquired within the same batch in the first target design image set are compared, and images ranked lower are eliminated according to a preset elimination ratio, leaving the second target design image set. Specifically, multiple pairs / lists of comparison units are constructed within the same batch. Pair comparison sorting and aggregation methods (such as the Elo or Bradley-Terry model) can be used to obtain the relative ranking of the batches, and images are eliminated according to the bottom-ranking ratio. For example, horizontal ranking is performed by batch (50 images per batch), and the bottom 20% of images in each batch are automatically eliminated.

[0049] During manual screening and review, the images in the second target design image set are manually screened and reviewed. The images that pass the review are stored in the vertical inspiration image library to build a retrieval candidate set. The review results are used to update the aesthetic scoring model. The review results include the labels marked as passed, rejected, or questionable for each image.

[0050] Furthermore, when conducting multi-dimensional aesthetic evaluations, it also includes: The system monitors the supply-demand ratio of images corresponding to different style tags, and triggers a targeted re-acquisition task for any style tag when a shortage of images is detected. It should be noted that the re-acquisitioned images also undergo multi-dimensional aesthetic evaluation before being added to the database.

[0051] In this way, by monitoring the image supply-demand ratio and triggering targeted re-collection tasks, the image library can be automatically expanded and optimized according to user behavior, forming a positive flywheel that gets better with use. This further ensures that there are enough images in the image library that meet aesthetic standards for users to search, improving search efficiency and accuracy, and meeting users' personalized needs.

[0052] To more clearly explain the image processing method for interior design scenarios disclosed in this application, please refer to the following examples: Scenario 1, see Figure 3-5 and Figure 6 A multi-dimensional aesthetic evaluation was performed on each image in the candidate interior design image set using an aesthetic scoring model: The platform operations team imported 200 candidate interior design images in batches; Step A1: Validity check – Filter 15 non-interior design images (product images / people images), leaving 185 images passed; Step A2: Four-dimensional aesthetic scoring - Extract four-dimensional scores (0-100) for order, craftsmanship, artistic conception, and innovation from 185 images to generate an aesthetic score vector; Step A3: Horizontal comparison of multiple batches - Rank the batches (50 images per batch) horizontally, automatically eliminate the bottom 20% of each batch (37 images in total), leaving 148 images; Step A4: Manual Screening and Review – Operations staff review 148 images (the interface may display a 4D radar chart to assist in the judgment): 118 images passed, 22 rejected, and 8 with doubts. Figure 6 The diagram shown is a schematic of the operation process; Step A5: Training Sample Generation – 118 rejected samples (+22 rejected samples) are stored in the training library of the aesthetic scoring model; Final inclusion in the database: Through a four-layer funnel (validity check, four-dimensional aesthetic scoring, multiple horizontal comparisons, and manual screening and review), the inclusion rate of high-quality products is about 68%, which is much higher than the natural high-quality product rate when there is no screening on the general platform.

[0053] It is worth mentioning that in practical applications, the aesthetic scoring model will continue to evolve with manual screening and review, and the consistency rate with manual scoring can be increased to 81%, while the amount of manual review can be reduced by 40%.

[0054] Scenario 2, see Figure 7-8 In response to the user's search request, the search results are output: The designer is in the initial design phase of a [Japanese minimalist style] bedroom: Step B1: The designer clicks the quick search button [Japanese Minimalist]; Step B2: The system reads the pre-configured multi-dimensional aesthetic weight vector V_weight = [Sense of Order: 0.50, Sense of Craftsmanship: 0.60, Sense of Artistic Conception: 0.70, Sense of Innovation: 0.65], and directly initiates a search; Step B3: The system filters the candidate set based on the current project stage (solution stage), and outputs the search results after fusing them with personal preference vectors; Step B4: The designer increases the [Innovation] value to 0.60 using the slider, and the search results are updated in real time; Step B5: The designer saves the current weight configuration as [My Japanese Template] for easy access next time.

[0055] In practical applications, designers can significantly improve search efficiency (by over 60%) by triggering professional searches with a single click, lowering the barrier to entry for professional searches. Furthermore, under the same search task, the traditional method of "repeated keyword trial and error + manual page-turning and filtering" takes an average of 12.5 minutes, while this system's "quick button + weight fine-tuning" takes an average of 4.8 minutes, resulting in an efficiency improvement of (12.5-4.8) / 12.5 = 61.6%).

[0056] The implementation principle of this embodiment is as follows: by associating target style tags with multi-dimensional aesthetic weight vectors and combining them with users' personalized aesthetic preferences, it can more accurately match user needs and output search results that meet user expectations. Simultaneously, the multi-dimensional aesthetic evaluation mechanism ensures the aesthetic quality of the images in the database, providing a high-quality candidate image set for subsequent searches. By collecting user feedback behavior to update the preference vector, the system can continuously adapt to changes in user behavior, improving the accuracy and efficiency of searches, and solving problems in existing technologies such as limited search accuracy, inability to cater to personalized needs, lack of aesthetic quality filtering, and lack of vertical image library accumulation.

[0057] [Second Embodiment] See Figure 9 The present application discloses a second embodiment of an intelligent retrieval and aesthetic evaluation system for inspirational image sets. The system includes: a retrieval response module 210, a retrieval query vector generation module 220, a retrieval result output module 230, and a feedback behavior acquisition module 240.

[0058] The retrieval response module 210 is used to respond to the user's retrieval request, determine the target style tag, and map the target style tag into a multi-dimensional aesthetic weight vector. The multi-dimensional aesthetic weight vector includes at least the weight values ​​corresponding to the dimensions of order, craftsmanship, artistic conception and innovation. The retrieval query vector generation module 220 is used to obtain the user's personalized aesthetic preference vector and fuse the multidimensional aesthetic weight vector with the personalized aesthetic preference vector to generate a retrieval query vector; The retrieval result output module 230 is used to perform matching in the retrieval candidate set based on the retrieval query vector, sort the candidate images by matching degree according to the matching results, and output the retrieval results; The feedback behavior collection module 240 is used to collect user feedback behavior on the search results, update the personalized aesthetic preference vector based on the feedback behavior, and store the updated vector in the user's historical behavior data.

[0059] It should be noted that the intelligent retrieval and aesthetic evaluation system for inspirational image sets disclosed in this embodiment implements an image processing method for interior design scenarios as described in the above embodiment, and therefore will not be described in detail here. Optionally, the various modules and other operations or functions in this embodiment are respectively for implementing the methods in the aforementioned embodiments.

[0060] [Third Embodiment] See Figure 10In the third embodiment of this application, an electronic device is disclosed, which includes a memory 310 and a processor 320. The memory 310 is used to store computer programs; the processor 320 is used to execute the computer programs to implement the steps of the image processing method for interior design scenarios described in the second embodiment above. For details, please refer to the above description and will not be described in detail here.

[0061] The technical effect of the electronic device provided in this embodiment in practical application is the same as the technical effect of the image processing method for interior design scenarios in the second embodiment.

[0062] [Fourth Embodiment] A computer-readable storage medium is disclosed in the fourth embodiment of this application. The computer-readable storage medium is, for example, a non-volatile memory, such as magnetic media (e.g., hard disks, floppy disks, and magnetic tapes), optical media (e.g., CD-ROMs and DVDs), magneto-optical media (e.g., optical discs), and hardware devices specifically configured to store and execute computer-executable instructions (e.g., read-only memory (ROM), random access memory (RAM), flash memory, etc.). A computer program is stored on the computer-readable storage medium. The computer-readable storage medium can be executed by one or more processors or processing devices to implement an image processing method for an interior design scene as described in the foregoing embodiments.

[0063] Furthermore, it is understood that the foregoing embodiments are merely illustrative examples of the present invention. Provided that the technical features do not conflict, the structure is not contradictory, and the purpose of the invention is not violated, the technical solutions of the various embodiments can be arbitrarily combined and used.

[0064] In the embodiments provided by this invention, it should be understood that the disclosed methods, systems, and devices can be implemented in other ways. For example, the modules included in the systems described above are merely illustrative, and the division of modules is only a logical functional division. In actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0065] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0066] Furthermore, in the various embodiments of the present invention, the functional units / modules can be integrated into one processing unit / module, or each unit / module can exist physically separately, or two or more units / modules can be integrated into one unit / module. The integrated unit / module described above can be implemented in hardware or in the form of hardware plus software functional units / modules.

[0067] The integrated units / modules implemented as software functional units / modules described above can be stored in a computer-readable storage medium. These software functional units, stored in a storage medium, include several instructions to cause one or more processors of a computer measurement device (which may be a personal computer, server, or network measurement device, etc.) to execute some steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0068] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An image processing method for interior design scenarios, characterized in that, include: In response to a user's search request, a target style tag is determined, and the target style tag is mapped to a multi-dimensional aesthetic weight vector. The multi-dimensional aesthetic weight vector includes at least the weight values ​​corresponding to the dimensions of order, craftsmanship, artistic conception, and innovation. Obtain the user's personalized aesthetic preference vector, and fuse the multidimensional aesthetic weight vector with the personalized aesthetic preference vector to generate a retrieval query vector; Based on the retrieval query vector, the search candidate set is matched, and the candidate images are sorted according to the matching results to output the retrieval results; Collect user feedback on the search results, update the personalized aesthetic preference vector based on the feedback, and store the updated vector in the user's historical behavior data.

2. The method according to claim 1, characterized in that, Before determining the target style tag and mapping the target style tag to a multidimensional aesthetic weight vector in response to the user's search request, the method further includes: A candidate interior design image set is obtained, and a multi-dimensional aesthetic evaluation is performed on each image in the candidate interior design image set through an aesthetic scoring model. Images that meet the preset aesthetic standards are selected and stored in a vertical inspiration image library to construct the retrieval candidate set. The acquisition method of the candidate interior design image set includes at least one of batch import, targeted supplementary acquisition, and user upload; The multidimensional aesthetic evaluation process includes at least validity testing, multidimensional aesthetic scoring, horizontal comparison screening, and manual screening and review. The horizontal comparison screening involves comparing multiple images from the same batch within the candidate interior design image set.

3. The method according to claim 2, characterized in that, The process of performing multi-dimensional aesthetic evaluation on each image in the candidate interior design image set using an aesthetic scoring model includes: The validity detection is performed on the images in the candidate interior design image set based on the preset sample prototype to obtain a valid first target design image set. Image features are extracted from the images in the first target design image set, and the image features of each image are scored using multiple independent scoring heads to obtain a corresponding multidimensional aesthetic score vector. The multidimensional aesthetic score vector includes at least scores for order, craftsmanship, artistic conception, and innovation. The images in the first target design image set are subjected to the horizontal comparison screening, and the images ranked lower are eliminated according to the preset elimination ratio, leaving the second target design image set. The manual screening and review is performed on the images in the second target design image set. The images that pass the review are stored in the vertical inspiration image library to construct the retrieval candidate set. The review results are used to update the aesthetic scoring model, wherein the review results include a label of pass, rejection or doubt for each image.

4. The method according to claim 1, characterized in that, The process of matching the candidate images in the retrieval candidate set based on the retrieval query vector, sorting the candidate images according to the matching results, and outputting the retrieval results includes: Identify the current project design stage the user is in, which includes at least the concept stage, scheme stage, detailed stage, and construction drawing stage; The scope of the retrieval candidate set is constrained according to the project design phase, wherein the images in the retrieval candidate set have a pre-established mapping relationship with the project design phase; In the constrained retrieval candidate set, the similarity between the retrieval query vector and the candidate images is calculated, and the candidate images are sorted according to the similarity and the retrieval results are output.

5. The method according to claim 3, characterized in that, The step of performing validity detection on the images in the candidate interior design image set based on a preset sample prototype to obtain a valid first target design image set includes: The feature information of each image in the candidate interior design image set is extracted to calculate the similarity with the preset sample prototype and obtain the validity score of each image. The feature information includes global features and local semantic features. The local semantic features include at least spatial layout, furniture components, wall-ceiling-floor relationship, and lighting structure. Based on the comparison result between the validity score and the preset validity score threshold, invalid images are eliminated, and the set of valid images is obtained as the first target design image set.

6. The method according to claim 1, characterized in that, Before obtaining the user's personalized aesthetic preference vector and fusing the multidimensional aesthetic weight vector with the personalized aesthetic preference vector to generate a retrieval query vector, the method further includes: The system receives user input to adjust the weight values ​​of at least one dimension of the multidimensional aesthetic weight vector, updates the multidimensional aesthetic weight vector, and saves it in association with the target style tag.

7. The method according to claim 3, characterized in that, Also includes: Monitor the supply-demand ratio of images corresponding to different style tags, and trigger a targeted supplementary acquisition task for the style tag when it is detected that the supply of images corresponding to any style tag is insufficient.

8. The method according to claim 1, characterized in that, The feedback behaviors include positive feedback behaviors and negative feedback behaviors. The positive feedback behaviors include at least bookmarking, staying on the page for a long time, and downloading. The negative feedback behaviors include at least skipping quickly and ignoring.

9. A system for intelligent retrieval and aesthetic evaluation of inspirational image collections, characterized in that, include: The retrieval response module is used to respond to the user's retrieval request, determine the target style tag, and map the target style tag into a multi-dimensional aesthetic weight vector. The multi-dimensional aesthetic weight vector includes at least the weight values ​​corresponding to the dimensions of order, craftsmanship, artistic conception and innovation. The retrieval query vector generation module is used to obtain the user's personalized aesthetic preference vector and fuse the multidimensional aesthetic weight vector with the personalized aesthetic preference vector to generate a retrieval query vector. The retrieval result output module is used to perform matching in the retrieval candidate set based on the retrieval query vector, sort the candidate images by matching degree according to the matching results, and output the retrieval results; The feedback behavior collection module is used to collect user feedback behavior on the search results, update the personalized aesthetic preference vector based on the feedback behavior, and store the updated vector in the user's historical behavior data.

10. 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 computer program, it implements the method described in any one of claims 1 to 8.

Images