Limited sample product crowdsourcing design method based on perceptual engineering and GAN

Abstract

The invention relates to the technical field of product design, in particular to a finite sample product crowdsourcing design method based on perceptual engineering and GAN. Comprising the following steps: S1, collecting product images, and constructing a sample data set of the product images; s2, performing perceptual image vocabulary collection on the product, performing KMO principal component factor analysis on scores in the established perceptual word space, extracting an average score of each perceptual image word to obtain a plurality of groups of perceptual image vocabularies, and scoring the sample data set through the plurality of groups of perceptual image vocabularies; s3, constructing an image generation model based on the perceptual image vocabularies and the GAN adversarial network, inputting an average score and a product image of each perceptual image vocabulary, and performing iterative learning for multiple times to obtain a mapping relationship between perceptual scores and element lines; and S4, acquiring a user emotion demand, outputting a product feature element image corresponding to the user emotion demand through the trained model, and outputting an image design product according to the feature element image.

Description

Technical Field

[0001] The present invention relates to the field of product design technology, and in particular to a crowdsourcing design method for limited sample products based on Kansei engineering and GAN. Background Art

[0002] Product design is crucial to a company's success, playing a vital role in meeting user needs, enhancing the user experience, shaping product image, strengthening market competitiveness, and boosting sales. However, in today's society, the general consumer product landscape presents a diverse landscape. With such a rich array of products, consumers and designers often encounter a mismatch between their emotional and physical needs, resulting in a limited number of products. While the market is flooded with a wide variety of products and merchandise, few truly resonate with consumers and satisfy their unique emotional needs. This phenomenon leaves consumers feeling bewildered and lost when choosing, and also makes them more likely to become loyal to a single product or brand over the long term.

[0003] Compared to existing designs, "function" is the primary design consideration, with no breakthroughs in styling or emotional expression. Creating products that truly resonate with consumers and satisfy their deepest emotional needs is key for companies to stand out in the fiercely competitive market, and it's a challenge designers are constantly striving to address. Consequently, product form design approaches centered around user emotional needs are attracting significant attention.

[0004] In the current product design field, crowdsourcing design, as an emerging model, aims to improve innovation efficiency and reduce costs by integrating a wide range of designer resources. However, the existing crowdsourcing design model has significant limitations. First, due to the huge differences in participants' abilities, the quality of design solutions fluctuates greatly, making it difficult to ensure the consistency of the output. Second, crowdsourcing design relies heavily on the designer's subjective experience and lacks a systematic method to translate user emotional needs into product design requirements, resulting in the final solution being out of touch with the user's actual needs. In addition, traditional crowdsourcing design requires relying on massive user feedback data to optimize design parameters. However, in actual applications, especially in the context of niche products or start-ups, sample data is often limited, which seriously restricts the efficiency of design iteration. Summary of the Invention

[0005] The technical problem solved by the present invention is to provide a crowdsourcing design method for limited-sample products based on Kansei engineering and GAN, which has excellent technical performance under the conditions that the original product design is not novel and the product data set has a small sample, and provides an effective solution for product innovation design under limited sample conditions.

[0006] The basic solution provided by the present invention is a crowdsourcing design method for limited sample products based on Kansei Engineering and GAN, which includes the following steps:

[0007] S1. Collect product images of existing products and build a sample dataset of product images;

[0008] S2. Collect perceptual imagery words for the product and perform KMO principal component factor analysis on the scores in the established perceptual word space. Extract the average score of each perceptual imagery word and perform data analysis and score calculation to screen perceptual imagery words. Obtain several groups of perceptual imagery words and score the sample data set using these groups of perceptual imagery words. The KMO principal component factor analysis is calculated using the following formula:

[0009]

[0010] Where r represents the simple correlation coefficient and a represents the partial correlation coefficient;

[0011] S3. Build an image generation model based on perceptual image vocabulary and GAN adversarial network. Input the average score of each perceptual image vocabulary and product image, perform multiple iterative learning, and obtain the mapping relationship between perception score and element line.

[0012] S4. Obtain user emotional needs, output product feature element images corresponding to user emotional needs through the trained model, and design products based on the feature element image output

[0013] S5. Distribute questionnaires to multiple evaluators and obtain the questionnaire results for the output image design product, and verify it using the results of the perceptual image vocabulary weight score.

[0014] Furthermore, the step S3 includes the following steps:

[0015] S31. Obtain the average scores of six different groups of perceptual image vocabulary and input them into the DCGAN generator. At the same time, input the corresponding product images into the generator and discriminator.

[0016] S32. In each training cycle, first train the generator to improve the discriminator's score for the generated images through backpropagation, and then train the discriminator to improve the discriminator's score for the real images and reduce the score for the generated images;

[0017] S33. Through the above steps, the generator and discriminator are continuously updated. When the training is finally completed, the generative adversarial network reaches the Nash equilibrium:

[0018]

[0019] Represents the goal of optimizing the discriminator D to maximize its output ability to distinguish between real samples and generated samples;

[0020] Represents the goal of optimizing the generator G so that the samples it generates can deceive the discriminator as much as possible;

[0021] L D Represents the objective function (loss function) of the discriminator D;

[0022] L G Represents the objective function (loss function) of the generator G;

[0023] represents p from the noise distribution z Calculate the expected value when sampling z, where z is usually the noise vector input to the generator to generate new data samples;

[0024] Indicates that p is obtained from the real data distribution x Calculation of the expected value of x when sampling data;

[0025] D (x) Represents the discriminant output of the discriminator D for the real sample x;

[0026] D(G(z)) represents the discriminant output of the discriminator D on the sample G(z) generated by the generator G;

[0027] f D (D(x)) represents the discriminator loss function for the real sample x;

[0028] f G (D(G(z))) represents the discriminator loss function used to generate sample G(z);

[0029] g G (D(G(z))) represents the loss function of the generator G, which attempts to deceive the discriminator D by generating samples G(z).

[0030] Furthermore, in the process of updating the discriminator in S32, the following steps are also included:

[0031] By incorporating the LC regularization parameter into the discriminator update process:

[0032]

[0033] R LC Represents the LeCam divergence term, which is used to limit the update amplitude of the discriminator to prevent the discriminator from overfitting to limited sample data;

[0034] represents p from the noise distribution z Calculate the expected value when sampling z, where z is usually the noise vector input to the generator to generate new data samples;

[0035] Indicates that p is obtained from the real data distribution x Calculation of the expected value of x when sampling data;

[0036] D (x) Represents the discriminant output of the discriminator D for the real sample x;

[0037] D(G(z)) represents the discriminant output of the discriminator D on the sample G(z) generated by the generator G;

[0038] a F It represents the target value of the real sample discriminant output, which is generally set as a constant to regulate the degree of matching between the discriminator output and the real sample;

[0039] a R Represents the target value of the generated sample discriminant output, which is also a constant used to regulate the degree of match between the discriminator output and the generated sample;

[0040] L D represents the loss function of the discriminator D, including the adversarial loss and LeCam divergence terms;

[0041] V D represents the standard adversarial loss of the discriminator D;

[0042] λ represents the weight coefficient of the LeCam divergence term, which is used to balance the standard adversarial loss V D and LeCam divergence R LC The impact between

[0043] By introducing a regularization parameter, the KL divergence between the generated image and the real image is converted into a Lecam divergence distribution.

[0044] Furthermore, the step S5 includes the following steps:

[0045] S51. Normalize the index, convert the absolute value of the index into a relative value, and convert the positive index using the following formula:

[0046]

[0047] Convert negative indicators using the following formula:

[0048]

[0049] S52. Calculate the index value weight of i under index j:

[0050]

[0051] S53. Calculate the information entropy value of index j:

[0052]

[0053] in

[0054] S54. Calculate the information entropy redundancy of index j:

[0055] g j =1-e j

[0056] S55. Calculate the weight of index j:

[0057]

[0058] S56. Calculate the comprehensive score of evaluator i:

[0059]

[0060] The principles and advantages of the present invention are:

[0061] By integrating perceptual work with generative adversarial networks (GANs), we propose a product innovation design method for limited sample conditions. The core principle is to transform user emotional needs into a quantifiable perceptual word space through Kansei engineering, which is then used to drive an improved GAN model to generate product design elements that align with emotional characteristics. First, existing product image designs are collected. This is done by using automated tools such as selnium and beautifulSuop to construct a sample dataset from e-commerce websites. KMO principal component factor analysis and SPSS software are used to obtain perceptual imagery vocabulary scores. Six core perceptual imagery vocabulary groups are then selected using semantic differential analysis and TOPSIS. Using a DCGAN as the baseline model, perceptual word scores and product images are input. Through an iterative learning training strategy and the introduction of an LC regularization parameter to optimize the discriminator update process, the model stably learns the mapping between perceptual scores and design elements under limited sample conditions.

[0062] Compared with the prior art, the present invention has the following advantages:

[0063] 1. Break through the traditional GAN's dependence on massive data, achieve high-quality generation under limited sample conditions, and avoid mode collapse and training instability.

[0064] 2. Through quantitative modeling of the finite element perceptual word space, user emotional needs are converted into objective design parameters, solving the pain points of traditional crowdsourcing design, such as reliance on experience and large quality fluctuations. This provides designers with precise creative prompts and reduces communication and trial-and-error costs.

[0065] 3. Compatible with multiple scenarios, it is not only suitable for small-sample enterprises, but can also empower the crowdsourcing design ecosystem, accelerate multi-role collaboration by generating prototypes, and improve user satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 This is a flow chart of an embodiment of a crowdsourcing design method for a limited sample product based on Kansei engineering and GAN in the present invention;

[0067] Figure 2 A schematic diagram of a sample data set of an embodiment of a crowdsourcing design method for limited-sample products based on Kansei engineering and GAN in the present invention;

[0068] Figure 3 A schematic diagram of a thermometer product according to an embodiment of a limited sample product crowdsourcing design method based on Kansei engineering and GAN in the present invention;

[0069] Figure 4 A schematic diagram of a thermometer product converted into a contour line according to an embodiment of a limited sample product crowdsourcing design method based on Kansei Engineering and GAN of the present invention;

[0070] Figure 5 A schematic diagram of the styling feature elements of various product design categories for a thermometer product in an embodiment of a finite sample product crowdsourcing design method based on Kansei engineering and GAN of the present invention;

[0071] Figure 6 A schematic diagram of the styling feature elements of various product design categories for a thermometer product in an embodiment of a finite sample product crowdsourcing design method based on Kansei engineering and GAN of the present invention;

[0072] Figure 7 This is a schematic diagram of the final output of a thermometer product design according to an embodiment of the limited sample product crowdsourcing design method based on Kansei Engineering and GAN in the present invention. DETAILED DESCRIPTION

[0073] The following is further described in detail through specific implementation methods:

[0074] The embodiment is basically as shown in the attached Figure 1 As shown:

[0075] A limited sample product crowdsourcing design method based on Kansei Engineering and GAN includes the following steps:

[0076] S1. Collect product images of existing products and build a sample data set of product images. Specifically, use Selenium to automate the browser, open the JD.com page through Selenium, and automatically scroll the page until all the required product information or images are loaded. Selenium will perform these operations and obtain the HTML content of the page. Use BeautifulSoup to parse HTML: After obtaining the HTML of the page, pass it to BeautifulSoup for parsing. Use BeautifulSoup to find the Tags, extract the URL of the product image (usually the link in the src attribute). Download images: Download images one by one from the extracted image URL list. Use Python's requests or urllib library to save the images to the local disk. Finally, collect the following Figure 2 The sample dataset is shown.

[0077] S2. Collect perceptual image vocabulary for the product, score the perceptual image vocabulary, and perform KMO principal component factor analysis on the scores in the established perceptual word space. Extract the average score of each perceptual image word, perform data analysis and score calculation on it, and screen the perceptual image vocabulary to obtain several groups of perceptual image vocabulary. Score the sample data set using these groups of perceptual image vocabulary. The KMO principal component factor analysis is calculated using the following formula:

[0078]

[0079] Where r represents the simple correlation coefficient and represents the partial correlation coefficient;

[0080] In this embodiment, Figure 3 Taking the thermometer product shown as an example, the product design categories of the thermometer product include body shape, temperature port shape, case shape, sound device shape, display screen shape, and temperature column shape. By distributing questionnaires, we collected sensory image vocabulary and ratings about the thermometer product.

[0081] Table 1 Questionnaire survey on small vertical body temperature detectors

[0082]

[0083] The component matrix after KMO rotation is shown in Table 2:

[0084] Table 2 Component matrix after KMO rotation

[0085]

[0086] Finally, we selected the perceptual image words shown in Table 3:

[0087] Table 3 Selected perceptual image vocabulary

[0088]

[0089]

[0090] Finally, the sensory image vocabulary scores of each element of the product are obtained as shown in Table 4.

[0091] Table 4 Perceptual image vocabulary scores for each part of product elements

[0092]

[0093] S3. Build an image generation model based on perceptual image vocabulary and GAN adversarial network, input the average score of each perceptual image vocabulary and product image, perform multiple iterative learning, and obtain the mapping relationship between perception score and element line. Figure 4 As shown, the product image is first converted into contour lines, and then the Figure 5 、 Figure 6 The styling feature elements of each product design category of the thermometer products shown.

[0094] S4. Obtain user emotional needs, and output product feature element images corresponding to user emotional needs through the trained model. Figure 7 Images shown are designed products.

[0095] S5. Distribute questionnaires to multiple evaluators and obtain the questionnaire results for the output image design product, and verify it using the results of the perceptual image vocabulary weight score.

[0096] The S3 includes the following steps:

[0097] S31. Obtain the average scores of six different groups of perceptual image vocabulary and input them into the DCGAN generator. At the same time, input the corresponding product images into the generator and discriminator.

[0098] S32. In each training cycle, first train the generator to improve the discriminator's score for the generated images through backpropagation, and then train the discriminator to improve the discriminator's score for the real images and reduce the score for the generated images;

[0099] S33. Through the above steps, the generator and discriminator are continuously updated. When the training is finally completed, the generative adversarial network reaches the Nash equilibrium:

[0100]

[0101] Represents the goal of optimizing the discriminator D to maximize its output ability to distinguish between real samples and generated samples;

[0102] Represents the goal of optimizing the generator G so that the samples it generates can deceive the discriminator as much as possible;

[0103] L D Represents the objective function (loss function) of the discriminator D;

[0104] L G Represents the objective function (loss function) of the generator G;

[0105] represents p from the noise distribution z Calculate the expected value when sampling z, where z is usually the noise vector input to the generator to generate new data samples;

[0106] Indicates that p is obtained from the real data distribution x Calculation of the expected value of x when sampling data;

[0107] D (x) Represents the discriminant output of the discriminator D for the real sample x;

[0108] D(G(z)) represents the discriminant output of the discriminator D on the sample G(z) generated by the generator G;

[0109] f D (D(x)) represents the discriminator loss function for the real sample x;

[0110] f G (D(G(z))) represents the discriminator loss function used to generate sample G(z);

[0111] g G (D(G(z))) represents the loss function of the generator G, which attempts to deceive the discriminator D by generating samples G(z).

[0112] The process of updating the discriminator in S32 further includes the following steps:

[0113] By incorporating the LC regularization parameter into the discriminator update process:

[0114]

[0115] R LC Represents the LeCam divergence term, which is used to limit the update amplitude of the discriminator to prevent the discriminator from overfitting to limited sample data;

[0116] represents p from the noise distribution zCalculate the expected value when sampling z, where z is usually the noise vector input to the generator to generate new data samples;

[0117] Indicates that p is obtained from the real data distribution x Calculation of the expected value of x when sampling data;

[0118] D (x) Represents the discriminant output of the discriminator D for the real sample x;

[0119] D(G(z)) represents the discriminant output of the discriminator D on the sample G(z) generated by the generator G;

[0120] a F It represents the target value of the real sample discriminant output, which is generally set as a constant to regulate the degree of matching between the discriminator output and the real sample;

[0121] a R Represents the target value of the generated sample discriminant output, which is also a constant used to regulate the degree of match between the discriminator output and the generated sample;

[0122] L D represents the loss function of the discriminator D, including the adversarial loss and LeCam divergence terms;

[0123] V D represents the standard adversarial loss of the discriminator D;

[0124] λ represents the weight coefficient of the LeCam divergence term, which is used to balance the standard adversarial loss V D and LeCam divergence R LC The impact between

[0125] By introducing a regularization parameter, the KL divergence between the generated image and the real image is converted into a Lecam divergence distribution.

[0126] The S5 comprises the following steps:

[0127] S51. Normalize the index, convert the absolute value of the index into a relative value, and convert the positive index using the following formula:

[0128]

[0129] Convert negative indicators using the following formula:

[0130]

[0131] S52. Calculate the index value weight of i under index j:

[0132]

[0133] S53. Calculate the information entropy value of index j:

[0134]

[0135] in

[0136] S54. Calculate the information entropy redundancy of index j:

[0137] g j =1-e j

[0138] S55. Calculate the weight of index j:

[0139]

[0140] S56. Calculate the comprehensive score of evaluator i:

[0141]

[0142] Comparative experiments show that existing GAN methods typically require large training datasets to generate high-quality and stable results. When sample data is limited, traditional GAN architectures often encounter the following problems: 1. Mode collapse due to insufficient training samples; 2. Instability in the training process; 3. Poor generation quality; 4. Difficulty capturing complex product features.

[0143] As shown in Table 1, traditional GAN variants such as LSGAN (FID: 17.07), WGAN (FID: 16.26), and RSGAN (FID: 16.09) perform poorly when sample data is limited. In contrast, our proposed method (FID: 12.72) overcomes these limitations by introducing specific mechanisms, achieving stable training and generating higher quality results even in the case of limited datasets.

[0144] Table 1 Comparative experiments

[0145]

[0146] To ensure model training stability under limited sample conditions and prevent overfitting, we implemented several key strategies: 1. Progressive Growing, where the model starts training with low-resolution images and gradually increases the resolution. 2. Adaptive Learning Rate Scheduling:

[0147]

[0148] Our experimental results verify the effectiveness of these stability measures:

[0149]

[0150] These results show that the generation quality continues to improve during training while maintaining stable training dynamics. The gradual reduction of the gradient norm and the stable G / D loss ratio indicate that the method successfully prevents mode collapse and training instability.

[0151] To provide a clearer illustration of how design parameters are handled during model training and validation, we provide a detailed case study to generate a temperature detector design with the “diversity” style:

[0152] As shown in Table 2, taking the "single-diversity" parameter as an example, we track its optimization during training:

[0153] Table 2 Training parameters

[0154]

[0155] This paper proposes an innovative product form design method based on Kansei Engineering (KE) and Generative Adversarial Network (GAN).

[0156] By integrating perceptual work with generative adversarial networks (GANs), we propose a product innovation design method for limited sample conditions. The core principle is to transform user emotional needs into a quantifiable perceptual word space through Kansei engineering, which is then used to drive an improved GAN model to generate product design elements that align with emotional characteristics. First, existing product image designs are collected. This is done by using automated tools such as selnium and beautifulSuop to construct a sample dataset from e-commerce websites. KMO principal component factor analysis and SPSS software are used to obtain perceptual imagery vocabulary scores. Six core perceptual imagery vocabulary groups are then selected using semantic differential analysis and TOPSIS. Using a DCGAN as the baseline model, perceptual word scores and product images are input. Through an iterative learning training strategy and the introduction of an LC regularization parameter to optimize the discriminator update process, the model stably learns the mapping between perceptual scores and design elements under limited sample conditions.

[0157] Using KE technology, we study the relationship between product features and user emotions. We then train an image generation network based on the sentiment scores of the product features' perceptual imagery. By inputting emotional value requirements, we generate corresponding product images with emotional characteristics. Designers use these generated product feature images, which contain creative cues, to design product features that meet user emotional needs and generate product design concept sketches.

[0158] The above are only embodiments of the present invention. Common knowledge such as the known specific structures and characteristics in the scheme are not described in detail here. Ordinary technicians in the field are aware of all common technical knowledge in the technical field of the invention before the application date or priority date, can obtain all existing technologies in the field, and have the ability to apply conventional experimental means before that date. Ordinary technicians in the field can improve and implement this scheme in combination with their own abilities under the inspiration given by this application. Some typical known structures or known methods should not become obstacles for ordinary technicians in the field to implement this application. It should be pointed out that for those skilled in the art, without departing from the structure of the present invention, several variations and improvements can be made, which should also be regarded as the scope of protection of the present invention. These will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. A crowdsourcing design method for limited-sample products based on Kansei Engineering and GAN, characterized by: The following steps are involved: S1. Collect product images of existing products and build a sample dataset of product images; S2. Collect perceptual imagery words for the product, perform KMO principal component factor analysis on the scores in the established perceptual word space, extract the average score of each perceptual imagery word, perform data analysis and score calculation to screen perceptual imagery words, obtain several groups of perceptual imagery words, and score the sample data set using these groups of perceptual imagery words. The KMO principal component factor analysis is calculated using the following formula: Where r represents the simple correlation coefficient and a represents the partial correlation coefficient; S3. Build an image generation model based on perceptual image vocabulary and GAN adversarial network. Input the average score of each perceptual image vocabulary and product image, perform multiple iterative learning, and obtain the mapping relationship between perceptual score and element line. S4. Obtain user emotional needs, output product feature element images corresponding to user emotional needs through the trained model, and design products based on the feature element image output; S5. Distribute questionnaires to multiple evaluators for the output image design product and obtain the questionnaire results. Use the perceptual image vocabulary weight score to verify the results.

2. The crowdsourcing design method for limited-sample products based on Kansei Engineering and GAN according to claim 1, characterized in that: The S3 includes the following steps: S31. Obtain the average scores of six different groups of perceptual image vocabulary and input them into the DCGAN generator. At the same time, input the corresponding product images into the generator and discriminator. S32. In each training cycle, first train the generator to improve the discriminator's score for the generated images through backpropagation, and then train the discriminator to improve the discriminator's score for the real images and reduce the score for the generated images; S33. Through the above steps, the generator and discriminator are continuously updated. When the training is finally completed, the generative adversarial network reaches the Nash equilibrium: Represents the goal of optimizing the discriminator D to maximize its output ability to distinguish between real samples and generated samples; Represents the goal of optimizing the generator G so that the samples it generates can deceive the discriminator as much as possible; L D Represents the objective function (loss function) of the discriminator D; L G Represents the objective function (loss function) of the generator G; represents p from the noise distribution z Calculate the expected value when sampling z, where z is usually the noise vector input to the generator to generate new data samples; Indicates that p is obtained from the real data distribution x Calculation of the expected value of x when sampling data; D (x) Represents the discriminant output of the discriminator D for the real sample x; D(G(z)) represents the discriminant output of the discriminator D on the sample G(z) generated by the generator G; f D (D(x)) represents the discriminator loss function for the real sample x; f G (D(G(z))) represents the discriminator loss function used to generate sample G(z); g G (D(G(z))) represents the loss function of the generator G, which attempts to deceive the discriminator D by generating samples G(z).

3. The crowdsourcing design method for limited sample products based on Kansei Engineering and GAN according to claim 2, characterized in that: The process of updating the discriminator in S32 further includes the following steps: By incorporating the LC regularization parameter into the discriminator update process: R LC Represents the LeCam divergence term, which is used to limit the update amplitude of the discriminator to prevent the discriminator from overfitting to limited sample data; represents p from the noise distribution z Calculate the expected value when sampling z, where z is usually the noise vector input to the generator to generate new data samples; Indicates that p is obtained from the real data distribution x Calculation of the expected value of x when sampling data; D (x) Represents the discriminant output of the discriminator D for the real sample x; D(G(z)) represents the discriminant output of the discriminator D on the sample G(z) generated by the generator G; a F It represents the target value of the real sample discriminant output, which is generally set as a constant to regulate the degree of matching between the discriminator output and the real sample; a R Represents the target value of the generated sample discriminant output, which is also a constant used to regulate the degree of match between the discriminator output and the generated sample; L D represents the loss function of the discriminator D, including the adversarial loss and LeCam divergence terms; V D represents the standard adversarial loss of the discriminator D; λ represents the weight coefficient of the LeCam divergence term, which is used to balance the standard adversarial loss V D and LeCam divergence R LC The impact between By introducing a regularization parameter, the KL divergence between the generated image and the real image is converted into a Lecam divergence distribution.

4. The limited sample product crowdsourcing design method based on Kansei Engineering and GAN according to claim 3, characterized in that: The S5 comprises the following steps: S51. Normalize the index, convert the absolute value of the index into a relative value, and convert the positive index using the following formula: Convert negative indicators using the following formula: S52. Calculate the index value weight of i under index j: S53. Calculate the information entropy value of index j: in e>0; S54. Calculate the information entropy redundancy of index j: g j =1-e j S55. Calculate the weight of index j: S56. Calculate the comprehensive score of evaluator i:

Images