Method and device for aesthetic evaluation of wedding dress images based on transfer learning, storage medium

Abstract

The application provides a wedding dress image aesthetic evaluation method and device based on transfer learning, and a storage medium. The method comprises the following steps: dividing wedding dress images meeting preset conditions into several categories, formulating a scoring standard for the wedding dress images of each category, and scoring the wedding dress images according to the scoring standard, wherein the scored wedding dress images and corresponding score values constitute a wedding dress image aesthetic data set; obtaining user click data of the wedding dress images in a platform, and constructing a content quality data set based on the user click data; constructing a content quality evaluation model according to the content quality data set; performing transfer learning based on the wedding dress image aesthetic data set and the content quality evaluation model to obtain a wedding dress image aesthetic evaluation model; and performing aesthetic evaluation on the wedding dress images by using the wedding dress image aesthetic evaluation model. The application performs fine and professional aesthetic scoring on wedding photography photos, and helps users to objectively judge or further correct the wedding dress images.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, in particular to a wedding image aesthetic evaluation method and device based on transfer learning and a storage medium. BACKGROUND

[0002] Nowadays, it is also an indispensable part for the newlyweds to take a set of wedding photos in the wedding preparation process. The bride and groom hope to be recorded in the best state at the important moment to share and keep. How to help people make a relatively objective judgment on the aesthetic of wedding photography images through technical means and help photographers get more beautiful photos during photography or in the post-processing stage is an important problem worth exploring. Therefore, the development of image aesthetic evaluation technology creates the possibility to solve this problem. This technology evaluates the image quality by simulating the perception ability of human beings through computers.

[0003] Image aesthetics evaluation is an important field in computer vision, which extracts image aesthetics features that conform to human perception from images by computer. With the rapid development of machine learning and deep learning technology, many convolutional neural networks are used or created to perform image aesthetics evaluation tasks. For example, RAPID: Rating Pictorial Aesthetics using Deep Learning (ACM Multimedia 2014) designs a network structure that considers both global and local information by extracting features in two branches and then merging them. Composition-preserving deep photo aesthetics assessment (CVPR 2016) proposes to add adaptive spatial pooling operations in the network to eliminate the influence of input picture size adjustment, and extract multi-scale features through multiple parallel adaptive pooling modules. Photo Aesthetics Ranking Network with Attributes and Content Adaptation (ECCV 2016) and Deep Aesthetic Quality Assessment with Semantic Information (TIP 2017) find that incorporating some additional image auxiliary information can also be helpful. The method proposed in NIMA: Neural Image Assessment (TIP 2018) is relatively direct and simple. It does not make too many modifications to the network, nor does it use too much auxiliary information. Instead, it changes the regression method used in the loss function to a classification method. The network outputs a probability distribution from 1 to 10 points, and introduces the Earth Move Distance (EMD) bulldozer distance to calculate the loss. In terms of data sets, there are AVA data set, AADB data set, PCCD data set, etc. The most commonly used is the AVA data set, which contains about 250,000 pictures, and the true value label of each picture is the average score of multiple raters.

[0004] It's worth noting that previous industry work didn't limit the scope of the photos evaluated; photos could include portraits, landscapes, animals, etc. Our focus, however, is specifically on wedding portraits. We sampled 100 portrait photos from a designated platform and tested them with models like NIMA. The results showed no differentiation; most photos scored within a very small range. This might be because previous datasets were global, including portraits, landscapes, animals, etc., and the portraits included artistic shots, film stills, and everyday photos. Different types of portraits objectively have aesthetic differences, and the photos themselves differ significantly from wedding photos. In short, the aesthetic evaluation models built in previous work were crude and intuitive. What's needed now is a refined, professional portrait photography aesthetic evaluation model to accurately assess wedding images. Summary of the Invention

[0005] To address the aforementioned issues, embodiments of this application provide a method, device, and storage medium for evaluating the aesthetics of wedding photos based on transfer learning. This method is capable of scoring the aesthetics of wedding photographs.

[0006] Therefore, one aspect of this application provides a method for evaluating the aesthetics of wedding dress images based on transfer learning, the method comprising the following steps:

[0007] Steps for constructing a wedding dress image aesthetics dataset: Divide wedding dress images that meet preset conditions into several categories, formulate scoring criteria for wedding dress images in each category, and score the wedding dress images according to the scoring criteria. The scored wedding dress images and their corresponding score values ​​constitute a wedding dress image aesthetics dataset.

[0008] Steps for constructing a content quality assessment model: Obtain user click data of the wedding dress images on the platform, construct a content quality dataset based on the user click data; construct a content quality assessment model based on the content quantity dataset;

[0009] Steps for constructing a wedding photography image aesthetics evaluation model: Based on the wedding image aesthetics dataset and the content quality evaluation model, transfer learning is performed to obtain the wedding image aesthetics evaluation model;

[0010] Wedding dress image aesthetic evaluation steps: The wedding dress image is aesthetically evaluated using the aforementioned wedding dress aesthetic evaluation model.

[0011] Optionally, in combination with any of the above aspects, in another implementation of this aspect, a content quality assessment model is constructed based on the content quantity dataset, specifically as follows:

[0012] The BERT model of the transformers library is used as a text encoding model, the input of the text encoding model is the title of the wedding dress image, and the output is a text feature vector; the ResNet model of the torchvision library is used as an image encoding model, the input of the image encoding model is the wedding dress image after compression processing, and the output is an image feature vector;

[0013] The text feature vector and the image feature vector are subjected to a linear layer to obtain two dimension reduction vectors, the two dimension reduction vectors are spliced to obtain a fusion feature vector, and the fusion feature vector is subjected to a linear layer and a softmax layer to obtain a content quality distribution;

[0014] The content quality distribution and the content quality data set are iteratively optimized using a cross-entropy loss function to obtain a content quality evaluation model.

[0015] Optionally, in combination with any of the above aspects, in another implementation manner of the present aspect, based on the wedding dress image aesthetic data set and the content quality evaluation model, transfer learning is performed to obtain a wedding dress image aesthetic evaluation model, specifically,

[0016] The image encoding model is used as a backbone network of the wedding dress aesthetic evaluation model, the input of the backbone network is a target image compressed to a first preset size, and the output is an image aesthetic feature vector; the image aesthetic feature vector is subjected to a linear layer and a softmax layer to obtain an image aesthetic distribution;

[0017] The image aesthetic distribution and the score value of the wedding dress image aesthetic data set are iteratively optimized using a loss function to obtain a wedding dress image aesthetic evaluation model.

[0018] Optionally, in combination with any of the above aspects, in another implementation manner of the present aspect, the loss function of the wedding dress image aesthetic evaluation model is

[0019]

[0020] Wherein N represents the batch size, q ij represents the jth element of the image aesthetic distribution, j = 1, 2, …, 10; y i is the score value in the wedding dress aesthetic data set, the training process is set to epoch = 5, the optimizer uses the Adam optimizer, and the learning rate lr = 1e-4.

[0021] Optionally, in combination with any of the above aspects, in another implementation manner of the present aspect, the loss function of the content quality evaluation model is

[0022]

[0023] where N denotes the batch size, p i denotes the true value of the content quality dataset, p' i The prediction result, the training process sets epoch=30, and the optimizer uses the Adam optimizer with a learning rate of lr=1e-4.

[0024] Optionally, in combination with any of the above aspects, in another implementation form of the aspect, the content quality dataset is constructed based on the user click data, specifically,

[0025] The wedding dress images in which the user click data is in the first preset range and the second preset range in each category are obtained, and part of the samples are selected from the wedding dress images for marking; wherein the samples in the first preset range are marked as high-quality works, and the samples in the second preset range are marked as low-quality works, and the high-quality works and the low-quality works constitute the content quality dataset.

[0026] Optionally, in combination with any of the above aspects, in another implementation form of the aspect, the first preset range is that the user click data is in the top 10%, the second preset range is that the user click data of the wedding dress images is in the tail 10%, and the content quality dataset is D ctr = {(w i ,t i ,p i )|w i ∈W M ,i=1,2,…M}, wherein w i denotes the cover image of the i-th work, t i denotes the title of the i-th work, p i ∈(0,1) denotes the quality of the i-th work, 0 denotes a low-quality work, 1 denotes a high-quality work, and M denotes the number of the dataset.

[0027] Optionally, in combination with any of the above aspects, in another implementation form of the aspect, the step of constructing the wedding dress image aesthetic dataset is specifically,

[0028] The wedding dress images in each category are scored according to a scoring table for a preset number of times, and each scoring is performed by a designated person; each wedding dress image can obtain a score value:

[0029]

[0030] wherein score j denotes the score of the j-th annotator, and K is the number of annotators;

[0031] The wedding photography image aesthetic dataset D={(x i ,y i ) is obtained |x i∈ X N , i = 1, 2, … N}, wherein X N represents a wedding photography image set, and N represents the number of photos in the wedding photography image set.

[0032] In another aspect of the present application, an electronic device is provided, which includes a processor, a memory, and a computer program stored on the memory and executable on the processor, and the processor implements the method for evaluating the aesthetics of wedding images based on transfer learning according to any one of the above aspects when executing the computer program.

[0033] In another aspect of the present application, a storage medium is provided, which stores a computer program, and the computer program implements the method for evaluating the aesthetics of wedding images based on transfer learning according to any one of the above aspects when executed.

[0034] As described above, the present application provides a method for evaluating the aesthetics of wedding images based on transfer learning, an electronic device, and a storage medium, constructs an aesthetic evaluation dataset of wedding photography images, constructs a multi-modal wedding photography content quality evaluation model according to user click data, performs transfer learning based on the wedding image aesthetic dataset and the content quality evaluation model, constructs a wedding image aesthetic evaluation model, and obtains an aesthetic score of a wedding image. The present application can perform fine and professional aesthetic scoring on wedding photography photos, and help users to objectively judge or further correct wedding images.

[0035] The above summary is provided to introduce some concepts in a simplified form, which will be described in further detail in the specific embodiments below. The above summary is neither intended nor should it be construed to identify key or essential features of the claimed subject matter, nor is it intended to be used in determining the scope of the claimed subject matter. The claimed subject matter is not limited to implementations that solve any or all of the disadvantages noted in the background. BRIEF DESCRIPTION OF DRAWINGS

[0036] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and serve to explain the principles of the present application. In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed to be used in the embodiment description will be briefly introduced as follows. Obviously, those skilled in the art can obtain other drawings according to these drawings without any creative effort. These drawings and the written description are not intended to limit the scope of the concept of the present application in any way, but to illustrate the concept of the present application for those skilled in the art by referring to specific embodiments.

[0037] Figure 1 A flowchart of a method for evaluating the aesthetics of wedding images based on transfer learning provided by the present application;

[0038] Figure 2 A mixed image score diagram of a wedding dress image aesthetic evaluation method based on transfer learning provided in the present application. DETAILED DESCRIPTION

[0039] The exemplary embodiments will be described in detail herein with reference to the accompanying drawings. In the following description, unless otherwise indicated, like numbers in the different drawings represent similar or analogous elements. The following exemplary embodiments described in the detailed description section are not meant to be an all-inclusive description of all aspects of the application. Rather, they are merely examples that can illustrate some aspects of the application as detailed in the appended claims.

[0040] It should be noted that, in this document, the terms "comprise", "comprising", or any other variant thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element preceded by "comprises a" does not, without more constraints, foreclose the existence of additional identical elements in the process, method, article, or apparatus that includes the element. Also, components having the same name in different embodiments of the application can have the same meaning or different meanings, which should be determined in light of their explanation in the specific embodiment and further in conjunction with the context in the specific embodiment.

[0041] It should be understood that although the terms first, second, third, etc. can be used herein to describe various information, these terms are not intended to denote a particular order or hierarchy among the information. These terms are merely used to distinguish one category of information from another. For example, a first information can be termed a second information, and similarly, a second information can be termed a first information without departing from the scope of this document. As used herein, the word "if" can be interpreted to mean "when" or "upon" or "in response to determining" depending on the context. Also, as used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context indicates otherwise. It will be further understood that the terms "comprises", "comprising", "includes" and / or "including", as used herein, specify the presence of stated features, steps, operations, elements, components, items, and / or groups thereof, but do not preclude the presence or addition of one or more other features, steps, operations, elements, components, items, and / or groups thereof. As used herein, the term "or", "and / or", "at least one of", and the like as used herein, can be interpreted to be inclusive, i.e., the selection can be additionally based on one or more of the other alternatives. Only under circumstances when explicitly indicated as such, e.g., within specific examples, will the definition of "or" be exclusive.

[0042] It should be understood that, although the steps in the flowcharts in the embodiments of the present application are shown in sequence according to the arrows, the steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, the execution of the steps is not strictly limited in sequence, and they can be executed in other orders. Moreover, at least part of the steps in the figure can include multiple sub-steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution order is not necessarily sequential, but can be alternately executed with at least part of other steps or sub-steps or stages of other steps.

[0043] Depending on the context, the word "if" as used herein can be interpreted as meaning "when" or "while" or "in response to a determination" or "in response to a detection". Similarly, depending on the context, the phrase "if it is determined" or "if (a stated condition or event) is detected" can be interpreted as meaning "when it is determined" or "in response to a determination" or "when (a stated condition or event) is detected" or "in response to a detection (of a stated condition or event)".

[0044] It should be understood that the specific embodiments described herein are merely intended to explain the present application, and are not intended to limit the present application.

[0045] Referring to Figure 1 The present application provides a wedding image aesthetic evaluation method based on transfer learning, which can perform aesthetic scoring on wedding photography photos.

[0046] Specifically, the method includes the following steps:

[0047] Step S1, constructing a wedding image aesthetic dataset step: dividing wedding images meeting predetermined conditions into several categories, formulating scoring standards for wedding images of each category, and scoring the wedding images according to the scoring standards. The scored wedding images and corresponding scores constitute a wedding image aesthetic dataset.

[0048] In order to ensure the objectivity and rationality of the wedding image scoring data, a scoring team composed of 10 ordinary people, 5 wedding photography post-processing retouchers, and 5 wedding photography photographers is first established, and the wedding photography photos are divided into 17 categories, including pure color interior, costume, green plants, sea view, indoor life, city street shooting, film feeling, European architecture, etc. This is because the photos in different categories are difficult to compare. For example, there is a big difference in post-processing between city street shooting and green plant photos, and the evaluation standards are naturally different.

[0049] A small sample of each category is taken by the scoring team, and a set of relatively unified and quantified wedding photography photo 10-point scoring standards is defined based on their professional level and personal aesthetics. Specifically, image color coordination (2 points): refers to the contrast, saturation, and other color aspects of the image; light and shadow (2 points): refers to the exposure level of the image, the use of light by the photographer, etc.; composition (2 points): refers to the composition of the characters and the background when shooting; character modeling and makeup (2 points): refers to character modeling, character value, and makeup; character performance (2 points): refers to character expression, body movement, and character atmosphere. The reason for disassembling the scoring standard is to avoid the bias of different people when scoring due to their different professions. For example, photographers may pay more attention to the composition and performance of the photo, while retouchers may pay more attention to the color and makeup of the photo. This improves the rationality and fairness of the scoring.

[0050] 20000 wedding photos were evenly extracted under 17 categories. The scoring team scored each photo according to the scoring standard, one category at a time. This avoids the problem of introducing noise when the scorer mixes different theme photos together. Finally, a relatively rigorous, objective, and professional wedding photography image aesthetics dataset is obtained, in which the true value of each photo is the average of the annotation results of 20 scorers.

[0051] The score of each wedding photo is:

[0052]

[0053] Where score j represents the score of the jth annotator, and K is the number of annotators.

[0054] The wedding photography image aesthetics dataset D = {(x i ,y i )|x i ∈X N ,i = 1, 2, … N} is obtained, where X N represents the set of wedding photography images, and N represents the number of photos in the set of wedding photography images.

[0055] Step S2, constructing a content quality evaluation model: obtaining user click data of the wedding images in the platform, constructing a content quality dataset based on the user click data; constructing a content quality evaluation model according to the content quantity dataset.

[0056] Click-through rate (CTR) prediction is a method of estimating whether a user will click on a particular product based on user attributes, user behavior records, and product attributes. In the wedding photography industry, "products" refer to the wedding photography portfolios published by businesses, typically displayed to users through titles and cover images. Determining whether a user will click on a particular photo is complex; it's not solely based on the quality of the work. It may also be related to user preferences, location, and the business's brand influence. However, it can be generally assumed that photos with higher CTRs are of higher quality than those with lower CTRs.

[0057] Based on the above ideas, the click-through rate prediction problem was simplified and transformed into a wedding photography content quality assessment problem. Under 17 wedding photography photo categories, wedding images within a first and second preset range of user click data were obtained from each category. A subset of these wedding images was then selected and labeled. Samples within the first preset range were labeled as high-quality works, and samples within the second preset range were labeled as low-quality works. These high-quality and low-quality works constituted the content quality dataset.

[0058] The first preset range is the top 10% of user click data, and the second preset range is the bottom 10% of user click data for wedding dress images. The content quality dataset is D. ctr ={(w i ,t i ,p i )|w i ∈W M ,i=1,2,…M}, where w i Let t represent the cover image of the i-th work. i p represents the title of the i-th work. i ∈(0,1) represents the quality of the i-th work, where 0 represents a low-quality work and 1 represents a high-quality work, and M represents the number of data sets.

[0059] A content quality assessment model is constructed based on the aforementioned content quantity dataset. The content quality assessment model adopts a dual-tower structure. On the text side, a pre-trained BERT model is used, taking the work title as input and outputting a 768-dimensional text encoding vector. On the image side, a pre-trained ResNet101 model is used, taking a work cover image compressed to 512*512 size as input and outputting a 2048-dimensional image encoding vector. The text encoding and image encoding are then subjected to linear layer dimensionality reduction, concatenation and fusion, and linear layer classification operations to output the content quality distribution.

[0060] Specifically, the BERT model of the transformers library is used as the text encoding model, the weight uses the pre-trained bert-base-chinese, the input of the text encoding model is a vector with a size of [N, 64] after the title of the wedding image is encoded by the tokenizer of BERT, and the output is a text feature vector T with a size of [N, 768] BERT The ResNet101 model of the torchvision library is used as the image encoding model, the weight uses the ImageNet pre-training weight built in torchvision, the input of the image encoding model is a wedding image small image processed to [N, 3, 512, 512], and the output is an image feature vector T with a size of [N, 2048] ResNet Where N is the batch size, which is 64 here.

[0061] The text feature vector T BERT is concatenated with the image feature vector T ResNet After the linear layers of [768, 512] and [2048, 512], two [N, 512] dimension reduction vectors are obtained, and the two dimension reduction vectors are concatenated to obtain a [N, 1024] fusion feature vector T Fusion The fusion feature vector T Fusion is input into the [1024, 2] linear layer and the softmax layer to obtain a content quality distribution with a size of [N, 2].

[0062] The content quality distribution and the true value of the content quality dataset are iteratively optimized using the cross-entropy loss function to obtain a content quality evaluation model.

[0063] The loss function of the content quality evaluation model is

[0064]

[0065] Where N represents the batch size, p i represents the true value of the content quality dataset, and p′ i is the prediction result, the training process is set to epoch = 30, the optimizer uses the Adam optimizer, and the learning rate lr = 1e-4.

[0066] Step S3, constructing a wedding photography image aesthetic evaluation model step: based on the wedding image aesthetic dataset and the content quality evaluation model, transfer learning is performed to obtain a wedding image aesthetic evaluation model.

[0067] Based on the image encoding model obtained in step S2 as the backbone network of the wedding dress aesthetic evaluation model, the backbone network input is the target image compressed to the first preset size, that is, [N, 3, 512, 512] size, and the output is an image aesthetic feature vector of [N, 2048]; the image aesthetic feature vector is subjected to a [2048, 10] linear layer and a softmax layer to obtain an image aesthetic distribution of [N, 10] dimensions.

[0068] The image aesthetic distribution and the score value of the wedding dress image aesthetic data set are iteratively optimized using the loss function as shown below to obtain a wedding dress image aesthetic evaluation model.

[0069]

[0070] Where N represents the batch size, q ij The jth element of the image aesthetic distribution, j = 1, 2, …, 10; y i The score value in the wedding dress aesthetic data set, the training process is set as epoch = 5, the optimizer uses the Adam optimizer, and the learning rate lr = 1e-4.

[0071] Step S4, wedding dress image aesthetic evaluation step: performing aesthetic evaluation on the wedding dress image through the wedding dress aesthetic evaluation model.

[0072] To verify the effectiveness of the method, 100 photos taken in a green plant scene were selected to use the image aesthetic evaluation model obtained by the method to predict, Figure 2 Some examples and corresponding model scores are shown.

[0073] More, on the other hand, to verify the effectiveness of the method, the following three image aesthetic evaluation methods are used, specifically 1) directly using the ResNet101 network for training; 2) using the ImageNet pre-trained ResNet101 for transfer learning training, and 3) the method of the present application. The titles of the 100 photos are hidden, and the pictures are packaged in the form of a questionnaire and sent to 50 randomly selected evaluators, and each object is required to select 20 high-quality and 20 low-quality photos from the 100 photos. Then the matching degree between the artificial score and the algorithm score can be calculated by the following formula:

[0074]

[0075] Where c iThe matching number of the 20 high-quality photos selected by the i-th evaluator and the Top-20 algorithm score, where C=20 represents that 20 high-quality photos are selected therefrom, and S=50 represents 50 evaluators. The test results of the three methods are shown in Table 1, and the investigation experiment results verify the effectiveness of the method.

[0076] Table 1 Comparison experiment results of different methods

[0077] High quality match rate Low quality match rate ResNet101 69.63％ 72.77％ ImageNet pre-trained ResNet101 76.92％ 78.59％ Patent method pre-trained ResNet101 81.54％ 83.13％

[0078] The application provides a wedding image aesthetic evaluation method and device based on transfer learning, constructs a wedding photography image aesthetic evaluation dataset, constructs a wedding photography content quality evaluation model based on multi-modal according to user click data, performs transfer learning based on the wedding image aesthetic dataset and the content quality evaluation model, constructs a wedding image aesthetic evaluation model, and obtains an aesthetic score of a wedding image. The application can perform fine and professional aesthetic scoring on wedding photography photos, and help users objectively judge or further correct wedding images.

[0079] The above application embodiment serial numbers are only for description, and do not represent the advantages and disadvantages of the embodiments.

[0080] In the present application, for the same or similar term concept, technical solution and / or application scene description, generally only the first time is described in detail, and for the sake of brevity, the repeated description is not repeated, and for the understanding of the technical solutions of the present application, the same or similar term concept, technical solution and / or application scene description which is not described in detail can be referred to the related description before.

[0081] In the present application, the description of each embodiment has its own emphasis, and the part not described or recorded in a certain embodiment can be referred to the related description of other embodiments.

[0082] The technical features of the technical solutions of the present application can be combined arbitrarily, in order to make the description simple, the possible combinations of each technical feature in the above embodiments are not described, however, as long as the combination of these technical features does not exist contradictory, it should be considered as the range recorded in the present application.

[0083] Through the above description of the embodiments, those skilled in the art can clearly understand that the above-mentioned example method can be realized by means of software and a necessary general hardware platform, and of course, it can also be realized by hardware, but in many cases, the former is a better embodiment. Based on such understanding, the technical solutions of the present application can be embodied in the form of a software product, which is stored in a storage medium (such as a ROM / RAM, a magnetic disk, or an optical disc) and includes a plurality of instructions for causing a terminal device (which can be an electrical device or a network device) to execute the method of each embodiment of the present application.

[0084] The above is only a preferred embodiment of the present application, and does not limit the patent scope of the present application, and any equivalent structure or equivalent process transformation using the contents of the specification and drawings of the present application, or direct or indirect application in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A method for evaluating the aesthetics of wedding dress images based on transfer learning, characterized in that, This method includes the following steps: Steps for constructing a wedding dress image aesthetics dataset: Divide wedding dress images that meet preset conditions into several categories, formulate scoring criteria for wedding dress images in each category, and score the wedding dress images according to the scoring criteria. The scored wedding dress images and their corresponding score values ​​constitute a wedding dress image aesthetics dataset. Steps for constructing a content quality assessment model: Obtain user click data of the wedding dress images on the platform, construct a content quality dataset based on the user click data; construct a content quality assessment model based on the content quantity dataset; Steps for constructing a wedding photography image aesthetics evaluation model: Based on the wedding image aesthetics dataset and the content quality evaluation model, transfer learning is performed to obtain the wedding image aesthetics evaluation model; The steps for evaluating the aesthetics of wedding dress images are as follows: The wedding dress images are evaluated aesthetically using the aforementioned wedding dress aesthetic evaluation model. Specifically, a content quality assessment model is constructed based on the aforementioned content quantity dataset. The BERT model from the transformers library is used as the text encoding model. The input of the text encoding model is the title of the wedding dress image, and the output is a text feature vector. The ResNet model from the torchvision library is used as the image encoding model. The input of the image encoding model is the compressed wedding dress image, and the output is an image feature vector. The text feature vector and the image feature vector are passed through a linear layer to obtain two dimensionality-reduced vectors. The two dimensionality-reduced vectors are concatenated to obtain a fused feature vector. The fused feature vector is passed through a linear layer and a softmax layer to obtain the content quality distribution. The content quality distribution and the content quality dataset are iteratively optimized using the cross-entropy loss function to obtain the content quality assessment model. Based on the aforementioned wedding dress image aesthetics dataset and the aforementioned content quality assessment model, transfer learning is performed to obtain a wedding dress image aesthetics evaluation model, specifically as follows: Based on the image coding model as the backbone network of the wedding dress aesthetic evaluation model, the input of the backbone network is a target image compressed to a first preset size, and the output is an image aesthetic feature vector; the image aesthetic feature vector is passed through a linear layer and a softmax layer to obtain the image aesthetic distribution. The image aesthetics distribution and the rating values ​​of the wedding dress image aesthetics dataset are iteratively optimized using a loss function to obtain the wedding dress image aesthetics evaluation model.

2. The method for evaluating the aesthetics of wedding dress images based on transfer learning as described in claim 1, characterized in that: The loss function of the wedding dress image aesthetic evaluation model is: in Indicates batch size. The first character representing the distribution of image aesthetics One element, ; For the rating values ​​in the wedding dress aesthetics dataset, the training process is set with epoch=5, the optimizer is Adam, and the learning rate is... lr=1e-4 .

3. The method for evaluating the aesthetics of wedding dress images based on transfer learning as described in claim 2, characterized in that: The loss function of the content quality assessment model is in Indicates batch size. This represents the ground truth value of the content quality dataset. The prediction results show that the training process was set with epoch=30, the Adam optimizer was used, and the learning rate lr=1e-4.

4. The method for evaluating the aesthetics of wedding dress images based on transfer learning as described in claim 1, characterized in that: The construction of the content quality dataset based on the user click data specifically involves, Get wedding dress images in each category whose user click data falls within a first preset range and a second preset range, and select a portion of the wedding dress images for labeling; Samples within a first preset range are labeled as high-quality works, and samples within a second preset range are labeled as low-quality works. The high-quality works and low-quality works constitute a content quality dataset.

5. The method for evaluating the aesthetics of wedding dress images based on transfer learning as described in claim 4, characterized in that: The first preset range is the top 10% of user click data, and the second preset range is the bottom 10% of user click data for wedding dress images. The content quality dataset is... ,in Indicates the first The cover image of each work. Indicates the first The title of the work, Indicates the first The quality of a work is represented by 0, indicating a low-quality work, and 1, indicating a high-quality work. Indicates the number of datasets.

6. The method for evaluating the aesthetics of wedding dress images based on transfer learning as described in claim 1, characterized in that: The steps for constructing the wedding dress image aesthetics dataset are as follows: Each category of wedding dress images is scored a predetermined number of times according to a scoring sheet, with each scoring session conducted by designated personnel; each wedding dress image receives a score of: in, Indicates the first The ratings from the labelers, To indicate the number of people; Obtain wedding photography image aesthetics dataset ,in This represents a collection of wedding photography images. This indicates the number of photos in the wedding photography image collection.

7. An electronic device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the method for evaluating the aesthetics of wedding dress images based on transfer learning as described in any one of claims 1 to 6.

8. A storage medium, characterized in that, It stores a computer program, which, when executed, implements a method for evaluating the aesthetics of wedding dress images based on transfer learning as described in any one of claims 1 to 6.

Images