Research on the Development of Innovative Paths of Public Service Advertisements in the Context of New Media on the Internet

In this paper, visual and attribute feature vectors are generated using image and attribute encoders as inputs, respectively. These feature vectors are combined by two fully connected layers to generate a 128-dimensional feature vector y, which adjusts the layout of the generated network.

Image coding: given an image, it is fed to a pre-trained VGG16 model to extract image features [25]. Using the 14×14×512 output of the last convolutional layer as the image representation, a 512-dimensional vector is formed using spatial global average pooling, which is then fed into three fully connected layers to generate a 128-dimensional image vector.

Attribute coding: In this paper, we consider four design attributes: ad style, appropriate population, subject categorization, and location information of the ad body and the main and secondary banners. In this paper, an integer from 0 to 5 is used to describe the style attribute of PSAs to represent six different PSA styles in the dataset. Meanwhile, an integer from 0 to 3 is used to represent the classification of appropriate population groups, which represent children, teenagers, middle-aged and elderly. For topic categorization, the dataset in this paper contains 10 first-level categories and five second-level categories under each first-level category, using integers from 0 to 9 to represent the first-level categories and integers from 0 to 4 to represent the second-level categories. For location information, this paper uses coordinate representation, and then encodes these four attributes into a hot vector and repeats each vector 10 times to increase its importance, and then inputs it into a fully connected layer to output a 48-dimensional attribute vector. Finally, all outputs are fused through a fully connected layer to form a 32-dimensional attribute vector.

Layout generation networks are based on GANs, which consist of a generator and a discriminator. Generator G learns to generate samples with the same distribution as the training data, while discriminator D learns to determine whether a given sample is real or generated. In the layout generation network of this paper, the generator maps a 128-dimensional potential vector to a layout. Discriminator D outputs a confidence value to indicate whether layout x is real or generated.

GANs typically generate layouts from sampled latent vectors. In contrast, an additional encoder E is added to the model in this paper in order to map the distribution p(z^|x) from the actual layout distribution p(x) to the feature space. Meanwhile, the generator G induces the distribution q(x˜|z) to map the samples from the prior distribution q(z) to the layout space. In this paper, the standard normal distribution is used as q(z). The discriminator D is trained to recognize joint pairs (x,z^=E(x))(x˜=G(z),z) in the layout space (x,x˜). Thus, the adversarial goal of this paper becomes: the generator and the encoder are trained together to deceive the discriminator by generating joint pairs from either the generator (x˜=G(z),z)) or the encoder (x,z^=E(x)) that are not recognizable by the discriminator, which learns to distinguish between the two types of joint pairs. The encoder processes the layout samples through a series of convolutional layers to generate two vectors representing the mean and standard deviation of a Gaussian distribution. The samples are then extracted from the Gaussian distribution as feature vectors.

To generate layouts, this paper inputs 128-dimensional vectors y embedded in the network as conditional information into the layout network and sends them to an encoder, a generator, and a discriminator so that the encoder E can induce a distribution p(z^|x,y) that maps the layout examples x from p(x) to the feature space conditional on y, the generator G can induce a distribution q(x˜|z,y) that maps the examples in q(z) to the layout space conditional on y, and the discriminator D learns to to recognize input joint pairs conditional on y. Connecting y directly to the generator G as an input to G is applied to the encoder E and discriminator D while replicating y along the spatial dimension to form the feature map of 60 × 45 × 128. Then, connect this function map with x or x˜ as input to D and connect it to the fourth layer of E.

Based on Least Squares GAN (LSGAN) [26], in this paper, the loss function of generator G and discriminator D is expressed as: 1LGAND=12(D(x,E(x,y),y)−1)2+12(D(G(z,y),z,y))2 2LGANG=12(D(G(z,y),z,y)−1)2

In order to generate higher quality images, this paper adds reconstruction loss L_rec and KL divergence loss L_KL to the loss function: 3Lrec=‖x−G(E(x,y),y)‖2 4LKL=DKL(p(z∧|x,y)‖q(z))

where D_KL is the KL difference. Meanwhile, in order to produce diversity in layout, this paper adds loss L_variety to the loss function to represent the diversity: 5Lvariety=mink∈1,2,…,K‖x−G(zk,y)‖2

where k is a hyperparameter. Therefore, the loss functions of the generator and encoder become: 6LG=LGANG+Lrec+Lvariety 7LE=Lrec+LKL

To generate the layout, this paper converts the output of the generator to an initial layout (60 × 45 × 3) by removing the filler elements and quantizing each value as 0 or 1. This paper uses a post-optimization step to optimize element boundaries and correct small deviations between elements, and then extracts individual elements from the initial layout by connecting component markers. This paper employs a series of morphological image processing operations to approximate the image boundaries to address their jaggedness. To solve small deviations between certain elements, this paper performs top-bottom, left-right alignment on them. To perform top alignment, firstly, if the top boundary coordinates of the element bounding boxes differ by less than 2 cells, the elements will be aggregated into a group. For the same group of elements, this article adjusts their top bounding boxes to align with the lowest top bounding box, thus creating sufficient spacing between the elements. Similarly, alignment of the bottom, left, or right side is performed in a similar manner.

Based on the input PSA styles, this paper samples 16 sets of information about the relative positions of advertisement themes and advertisement descriptors in the whole image from the dataset. For each set of sampled values, this paper randomly generates 32 layouts. And according to the appropriateness of the aspect ratio of the PSA subject and the aspect ratio of the input image in the generated layouts, if the difference is too large, they are filtered out. Finally, in order to diversify the generated layouts, this paper uses the Maximum Marginal Relevance criterion (MMR) to process the filtered layouts. Specifically, this paper uses the discriminator output (classification probability) as the quality score of the generated layouts, and uses the L2 distance in the feature space to compute the similarity score between layouts. For layouts ι of each other, this paper computes the ranking score ri=Qt−maxe∈LS(ι,e), where Q_t is the quality score of ι and S(ι, e) is the similarity score between ι and e. The layouts with the highest ranking scores are sequentially added to the ranked list L and all layouts are sorted by the balance of quality and similarity. Finally, this paper returns the top three layouts as the generated layouts.

The implementation of the personalized PSA recommendation model requires the collection and processing of user and PSA data, including user behavioral data, interest data, and social relationship data, as well as data on the content, structure, and language style of PSAs. The specific data collection and preprocessing operation flow is shown in Figure 3. 1)

Acquire raw data from target data sources (e.g., databases, websites, etc.), and perform cleaning operations on the raw data in order to ensure the accuracy and completeness of the data.

Figure 3.

Data acquisition and preprocessing process

On the basis of user data analysis, users can be tagged into different interest categories or feature labels. The core step of user label modeling involves feature extraction. Feature extraction is performed by analyzing the user’s historical behavioral data and other relevant information. This study uses a logistic regression algorithm to extract representative and distinguishing features [27]. Namely: 8P(y=1|x)=1/(1+exp(−(wT×x+b)))

Equation (8) expresses the probability that the target variable takes the value of 1. x is the input eigenvector, w and b are the model parameters. w^T denotes the transpose of w and exp denotes the natural exponential function. The core idea of the logistic regression algorithm is to minimize the loss function by combining the input features in a linearly weighted manner and mapping the result to the probability range between 0 and 1 through a specific sigmoid function, where a gradient descent algorithm is used to iteratively update the values of parameters w and b so that they can minimize the loss function.

The implementation of personalized PSA recommendations requires the selection of appropriate similarity calculation methods and recommendation algorithms, so as to calculate similarity between PSAs and users, sort and recommend PSAs. Choosing a suitable algorithm can improve the accuracy and effectiveness of advertisement recommendations. In this study, the cosine similarity-based collaborative filtering recommendation algorithm is selected for implementation. The cosine similarity-based collaborative filtering recommendation algorithm uses the historical behavioral data of users to measure the degree of similarity between users by calculating the cosine similarity between the rating vectors, and then realizes personalized recommendation [28].

First construct a user-item rating matrix: create a rating matrix ratings, where each row represents a user’s rating of an item. This rating matrix can be defined and populated according to the actual situation.

The next step is to find a number of users who are most similar to the target user by calculating the cosine similarity of the rating vectors between users, which is a measure of similarity between vectors, and is calculated as shown below: 9cosθ=A⋅B‖A‖×‖B‖=∑i=1n(Ai×Bi)∑i=1n(Ai)2×∑i=1n(Bi)2

In Eq. (9) A and B denote two n-dimensional vectors, A is [A1,A2,⋯,An], B is [B1,B2,⋯,Bn], and ‖A‖ and ‖B‖ denote the lengths of vector A and vector B, respectively. The process of calculating the cosine similarity of the angle between A and B the cosine similarity of θ involves mapping the vectors to a higher dimensional space and calculating the angle between them, with the result ranging from -1 to 1. When the angle between two vectors is close to 0°, i.e., their directions are almost identical, the cosine similarity is close to 1, indicating that the two vectors are very similar. When the angle is close to 90°, i.e., their directions are almost perpendicular, the cosine similarity is close to 0, indicating that there is no significant similarity between the two vectors. When the angle is close to 180°, i.e. their directions are completely opposite, the cosine similarity is close to -1, indicating that the two vectors are not similar at all.

Once the most similar users are found, the algorithm can use the preferences and behavioral patterns of these users to predict the target user’s preference for items that have not yet been evaluated. Typically, the algorithm recommends the items to the target user based on the average or weighted average ratings of the most similar users for the items that have not yet been evaluated. The formula for calculating the similarity of users is as follows: 10wnn′=rnTrn′rnrn′

In Eq. (10) wnn′ is the similarity between user n and user n′, andr_n represents the rating vector of user n.

The weighted average of similar users is calculated as follows: 11rnm=∑n′∈Unwnn′∑n″∈Unwnn″rn″m

In Eq. (11) r_nm is the rating of user n for item m, i_nm ∈ (0, 1) indicates whether user n has a rating for item m or not, 1 indicates a rating and 0 indicates no rating. U_n is the set of similar users for user n.

A collaborative filtering recommendation algorithm based on cosine similarity can utilize historical data to compute the similarity between users and combine it with user rating data for training and predicting the content of advertisements that users may be interested in. The model training uses a matrix decomposition algorithm, the core of which is to decompose the user-item rating matrix into two low-dimensional matrices, and the implied feature vector obtained through learning represents the relationship between users and items. This implied feature vector can capture the potential association between the user and the item, so that it can accurately score prediction and recommendation, and the matrix decomposition algorithm is widely used in collaborative filtering recommendation to improve the recommendation accuracy and personalization. The matrix decomposition process is shown below: 12R≈P×QT=R^

The P matrix in Eq. (12) is the relationship of N users to K topics, and the Q matrix is the relationship of K topics to M items. How to measure the error between the decomposed matrix and the original rating matrix, this study uses a loss function to solve, which is the minimum of the sum of the losses of all non-”-” items (i.e., unrated items in the original matrix) in Eq. (13): 13minloss=∑ri,j≠−ei,j2

This paper uses the expert scoring method to conduct a questionnaire research for a group of graduate students and above. 60 questionnaires were distributed to the respondents, and 52 valid responses were received. 1)

Evaluation of PSA image quality

Through the research, the importance of PSA image synthesized category diversity, style diversity, contour integrity and image clarity evaluation indexes as shown in Table 1, the importance of the indexes is divided into seven grades, from low to high for the sequential order: extremely unimportant, unimportant, slightly unimportant, general, slightly important, important, extremely important, respectively, with a score of 1-7 instead. The importance of each indicator is calculated as the mean/total of the scores obtained for that indicator (total score of 7).

Observation of Table 1 shows that the importance of the four indicators of category diversity, style diversity, contour completeness, and image clarity lies in the range of 0.8-0.9, indicating that the indicators selected for the experiment are representative in the evaluation of the experimental effect. Among them, the importance of image clarity is the highest, 0.9150, and the importance of category diversity is the lowest, 0.8332, indicating that the clarity of image synthesis in public service ads is more important in comparison, while comparing the image style, the importance of category indexes is slightly lower, which indicates that public service ads oriented to the population is more concerned about the style diversity presented by the ads.

The effect scores of PSA images synthesized by the GAN model are shown in Table 2, and the scores are 1-10 in descending order, and the experiment divides the scores into five grades, and the images whose scores are located in the range of 1-2 are defined to be very bad, 3-4 are bad, 5-6 are average, 7-8 are good, and 9-10 are very good.

The questionnaire results show that the mean values of the scores of the four evaluation indexes are, in descending order: category diversity (8.42), style diversity (8.53), contour completeness (8.94), and image clarity (9.17), which indicates that the categories of the product images synthesized by the GAN have an obvious degree of differentiation and the clarity of the synthesized images is better. The percentage of evaluated images with scores above 7 (i.e., evaluated as “good” and “very good”) in the above indicators are 80.75%, 80.82%, 86.13%, and 90.58%, respectively, indicating that the overall quality of the PSA images synthesized by the model is excellent.

2)

Evaluation of PSA Layout Effect

The questionnaire set up four groups of layout images, and the evaluator scored each group of PSAs according to their visual effects in terms of layout reasonableness, aesthetics, and typographic neatness. The rating scale is the same as that of the first part of the questionnaire, and the score of 1-2 is rated as very bad layout effect, 2-4 as bad, 5-6 as average, 7-8 as good, and 9-10 as very good effect. The breakdown of the scores for each group is shown in Tables 3 to 6.

The percentages of the four groups of images rated as “good” and “very good” in the three indicators of layout rationality, layout aesthetics and layout neatness were 86.36, 78.72 and 89.05, respectively. 86.54, 75.77, 78.72. Rationality, layout aesthetics and layout Neatness were 86.36, 78.72 and 89.05, respectively. 79.62, 76.68, 75.00. 80.71, 75.12, 89.15. The above values fluctuate above and below 75.04% in general, and the highest favorable rate can reach 89.15%, indicating that the layout of PSAs generated by the GAN model is excellent. Among them, the most reasonable layout is the second group of images, and the average value of the layout reasonableness score of this group is 8.49. The most beautiful layout of the first group of images, the average value of its score is 8.06. The highest degree of neatness of the layout of the fourth group of images, the average value of the neatness of the layout of 8.67. After observation and analysis, to summarize the rationality of the layout of the type of high score as follows: the product image is placed in the center, occupying an area ratio of about 0.5-0.8, the title and the text box is placed in the middle of the blank area and the area occupied by enough to place the text based on the basis of not too large. The area occupied by the title and text box should not be too large, as there should be enough space for the text to be placed. The type of layout that scores high in aesthetics is: the combination of elements should be flexible and changeable, while being reasonable and neat. Although placing elements in the center is not untidy and reasonable, this type of stereotypical layout will cause aesthetic fatigue. Typologies that scored high in typographic neatness are: the title and text components in the page should try to avoid covering over the image to avoid visual occlusion, in addition, the more parallelism contained between the element components, the neater the visual effect presented by the layout of the PSA.

In this paper, two metrics, Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM), are used to objectively evaluate the quality of the model-generated PSA print images. Among them, the smaller the value obtained for PSNR, the greater the image distortion, and vice versa, the smaller the distortion and the better the quality of the evaluated image. While SSIM takes the value range of 0-1, the larger the value, the more similar the structure of the two groups of images, that is, the better the quality of the evaluated image, when SSIM takes the value of 1, it indicates that the test image is as good as the quality of the original image.

The PSNR and SSIM values of the images are calculated separately using Python and the installed image processing libraries include Pillow, scikit-image.The original images in the dataset are fed into the evaluation model along with the GAN-generated images to obtain the PSNR and SSIM values of the generated images with respect to the original images. Three control groups have been set up for the experiment: the Fashion-MNIST image and the model synthesized image in this paper for quality comparison. Fashion-MNIST images are quality compared with CGAN model synthesized images. FL-GAN dataset images are quality compared with model generated images. The quantitative evaluation values of PSNR and SSIM for this paper’s model, CGAN model and FL-GAN model are shown in Table 7.

In the table, the PSNR value of the images generated by this model, CGAN model and FL-GAN model are all higher than 20, and the SSIM value is greater than 0.8, which indicates that under the evaluation of PSNR and SSIM indicators, the distortion of the images generated by the above three models is within the acceptable range, and is closer to the quality of the original image. Among them, the PSNR and SSIM values of this paper’s model are larger than CGAN and smaller than FL-GAN, indicating that this paper’s model is able to approach the quality of the original image to a greater extent under the circumstance of guaranteeing lower image distortion, and the overall quality of the generated image is superior.

KMO (Kaiser-Meyer-Olkin) is an index for comparing the correlation coefficient between variables.The value of KMO is between 0 and 1.The closer the value is to 1, the stronger the correlation between the variables, the more common the factors are, and the more suitable the variables are to be analyzed for factor analysis.On the contrary, the variables are not suitable to be analyzed for factor analysis. On the contrary, the variables are not suitable for factor analysis.Bartlett’s test of sphericity, the sig value is less than 0.05, the data are spherically distributed and suitable for factor analysis.

The KMO value of this study is 0.869 which shows that it is suitable for factor analysis.The Bartlett’s test of sphericity value is 2194.562 to reach significance and the sig. value is 0.000<0.05 which makes it suitable for factor analysis.

In this study, 53 variables from 124 valid questionnaires were subjected to the extraction of the common factors, including correlation matrix, total variance explained, fragmentation chart, and rotated component matrix. Using principal component analysis, 12 common factors were extracted with eigenvalues greater than 1. The total variance explained is shown in Table 8. Due to the oversized chart, only the portion of the information with an eigenvalue of 1 or more was intercepted.

As can be seen from Table 8, the public factors are arranged in the order of eigenvalues from the largest to the smallest, and the cumulative contribution of variance of the 12 public factors is 74.659%, which can explain most of the differences in the variables. Therefore, these 12 factors are the principal factors of the original variables.

In addition, the distribution of the public factors can be seen from the gravel plot shown in Figure 5. The eigenvalue of the first common factor is 7.247, which maximally contributes to explaining the original variables, and the subsequent ones show a decreasing trend one by one, and the eigenvalues of the factors after 12 are below 1, which contributes less to the explanation of the original variables and is not suitable for extraction.

Figure 5.

Scree plot

Using the method of rotating the component loadings matrix by maximum variance orthogonal rotation, which converged after 20 iterations, the results of the rotated component matrix were obtained as shown in Table 9. Where Q1~Q53 represent the 53 variables of the questionnaire.

Through the rotation, it can be clearly seen that there are several variables contained in each common factor. Analyzing the rotated component matrix table, it was found that some of the public factors contain too few items, some have the same or similar meanings, and some have low loading coefficients. In order to analyze more clearly and understandably, the male factors were summarized and organized, and seven principal component factors were finally identified.

The variables and loading coefficients specifically contained in the seven principal component factors identified are shown in Table 10. After summarizing and organizing, the factors for evaluating the perceived value of public service announcements are grouped into 7 main components, each of which in turn contains a number of variables, totaling 41 items. Based on the variables contained in the principal components, these 7 principal component factors are named for evaluating PSAs. 1)

Public factor 1 can be named as social factor. Social factor: public service announcements spread positive concepts to the public, have propaganda and educational effects on the society, and pass on the culture of the society, which can guide college students to the correct moral outlook, change the bad moral behavior habits of college students, and promote the formation of correct values of college students.