AI Automatic Image Screening Method for Advertising Promotion Based on Online Network Platform

Abstract

The present invention discloses an AI image automatic screening method for advertising promotion based on an online network platform, belonging to the technical field of image processing. First, the images on the online network platform are converted to the YUV space to obtain the Y, U, and V channel images. Then, various Laplacian operator operations of 4-neighborhood and 8-neighborhood are respectively performed on each channel image to construct the first and second channel detail enhancement images. Next, the edge sharpness of each region in these two types of images is calculated respectively to construct the first and second edge sharpness matrices. Then, the above matrices corresponding to the Y, U, and V channel images are processed by a picture quality evaluation model to obtain a picture quality score. Finally, when the picture quality score is greater than the score threshold, the picture is marked as a recommended state. The present invention solves the problem of low accuracy of picture quality evaluation existing in the prior art.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and particularly to an AI image automatic screening method for advertising promotion based on an online network platform. Background Art

[0002] With the rapid development of the Internet advertising market, the impact of picture quality on advertising effects has become increasingly significant. Currently, most advertising platforms still rely on manual review and simple picture quality screening technologies, which mainly focus on the basic attributes of pictures, such as resolution, file size, and simple color histogram analysis. However, this traditional method lacks in-depth analysis of picture details and edge sharpness, and cannot accurately evaluate the quality of pictures. Existing technologies usually cannot effectively distinguish pictures with similar resolutions but significantly different visual qualities, resulting in low-quality pictures frequently entering the advertising promotion system. This not only reduces the visual appeal of advertisements but also may affect the click-through rate and conversion rate of users. Summary of the Invention

[0003] Aiming at the above deficiencies in the prior art, an AI image automatic screening method for advertising promotion based on an online network platform provided by the present invention solves the problem of low accuracy in picture quality evaluation existing in the prior art.

[0004] To achieve the above invention objective, the technical solution adopted by the present invention is: an AI image automatic screening method for advertising promotion based on an online network platform, comprising the following steps:

[0005] S1. Convert the pictures on the online network platform to the YUV space to obtain multiple channel images, where the multiple channel images include: Y channel image, U channel image, and V channel image;

[0006] S2. Perform Laplacian operator operations of multiple 4-neighborhoods and 8-neighborhoods on each channel image respectively to construct a first channel detail enhancement image and a second channel detail enhancement image;

[0007] S3. Calculate the edge sharpness of each region in the first channel detail enhancement image and the second channel detail enhancement image respectively to construct a first edge sharpness matrix and a second edge sharpness matrix;

[0008] S4. Use a picture quality evaluation model to process the first edge sharpness matrix and the second edge sharpness matrix corresponding to the Y, U, and V channel images to obtain a picture quality score;

[0009] S5. When the picture quality score is greater than the score threshold, mark the picture as a recommended status.

[0010] Further, the 4-neighborhood Laplacian operator in S2 includes: standard 4-neighborhood Laplacian operator , the modified 4-neighborhood Laplacian operator ; The 8-neighborhood Laplacian operator includes: the standard 8-neighborhood Laplacian operator , the gradient 8-neighborhood Laplacian operator .

[0011] Further, S2 includes the following sub-steps:

[0012] S21. Perform various 4-neighborhood Laplacian operator operations on each channel image respectively, and fuse the operation results to obtain the first channel detail enhanced image;

[0013] S22. Perform various 8-neighborhood Laplacian operator operations on each channel image respectively, and fuse the operation results to obtain the second channel detail enhanced image.

[0014] Further, S21 includes the following sub-steps:

[0015] S211. Perform the standard 4-neighborhood Laplacian operator and the modified 4-neighborhood Laplacian operator operations on each channel image respectively to obtain the first detail image and the second detail image;

[0016] S212. Take the absolute value of each pixel value in the first detail image, and take the absolute value of each pixel value in the second detail image;

[0017] S213. Add the pixel values of the first detail image after taking the absolute value and the second detail image after taking the absolute value at the same pixel points to obtain the first channel detail enhanced image. When the channel image in S211 is the Y channel image, the first channel detail enhanced image is the first Y channel detail enhanced image. When the channel image in S211 is the U channel image, the first channel detail enhanced image is the first U channel detail enhanced image. When the channel image in S211 is the V channel image, the first channel detail enhanced image is the first V channel detail enhanced image.

[0018] Further, S22 includes the following sub-steps:

[0019] S221. Perform the standard 8-neighborhood Laplacian operator and the modified 8-neighborhood Laplacian operator operations on each channel image respectively to obtain the third detail image and the fourth detail image;

[0020] S222. Take the absolute value of each pixel value in the third detail image, and take the absolute value of each pixel value in the fourth detail image;

[0021] S223. Add the pixel values of the third detail image after taking the absolute value and the fourth detail image after taking the absolute value at the same pixel points to obtain a second-channel detail enhancement image. When the channel image in S221 is the Y-channel image, the second-channel detail enhancement image is the second Y-channel detail enhancement image; when the channel image in S221 is the U-channel image, the second-channel detail enhancement image is the second U-channel detail enhancement image; when the channel image in S221 is the V-channel image, the second-channel detail enhancement image is the second V-channel detail enhancement image.

[0022] Further, S3 includes the following sub-steps:

[0023] S31. Calculate the edge sharpness for each region in the first-channel detail enhancement image and construct a first edge sharpness matrix.

[0024] S32. Calculate the edge sharpness for each region in the second-channel detail enhancement image and construct a second edge sharpness matrix.

[0025] Further, the steps of constructing the edge sharpness matrix in both S31 and S32 include the following:

[0026] A1. Mark the pixel points with pixel values greater than the pixel threshold in the channel detail enhancement image as edge points.

[0027] A2. Divide the channel detail enhancement image into multiple regions.

[0028] A3. In each region, connect the adjacent edge points horizontally to form multiple horizontal edge lines.

[0029] A4. In each region, connect the adjacent edge points vertically to form multiple vertical edge lines.

[0030] A5. Calculate the edge sharpness of the region based on the multiple horizontal edge lines and multiple vertical edge lines.

[0031] A6. Use the edge sharpness of the region as an element to construct an edge sharpness matrix.

[0032] Further, the formula for calculating the edge sharpness of the region in A5 is: , where θ is the edge sharpness of the region, N level,i is the number of edge points on the i-th horizontal edge line, N ver,i is the number of edge points on the i-th vertical edge line, i is a positive integer, L level is the number of horizontal edge lines in a region, L ver is the number of vertical edge lines in a region.

[0033] Further, the image quality evaluation model in S4 includes: a Y-channel sharpness feature fusion unit, a U-channel sharpness feature fusion unit, a V-channel sharpness feature fusion unit, a first convolutional block, a second convolutional block, a third convolutional block, a fourth convolutional block, a fifth convolutional block, a sixth convolutional block, a first fully connected layer, a second fully connected layer, and a third fully connected layer;

[0034] The input end of the Y-channel sharpness feature fusion unit is used to input the first edge sharpness matrix and the second edge sharpness matrix corresponding to the Y-channel image;

[0035] The input end of the U-channel sharpness feature fusion unit is used to input the first edge sharpness matrix and the second edge sharpness matrix corresponding to the U-channel image;

[0036] The input end of the V-channel sharpness feature fusion unit is used to input the first edge sharpness matrix and the second edge sharpness matrix corresponding to the V-channel image;

[0037] The input end of the first convolutional block is connected to the output end of the Y-channel sharpness feature fusion unit, and its output end is respectively connected to the input end of the fourth convolutional block and the input end of the first fully connected layer;

[0038] The input end of the second convolutional block is connected to the output end of the U-channel sharpness feature fusion unit, and its output end is respectively connected to the input end of the fifth convolutional block and the input end of the first fully connected layer;

[0039] The input end of the third convolutional block is connected to the output end of the V-channel sharpness feature fusion unit, and its output end is respectively connected to the input end of the sixth convolutional block and the input end of the first fully connected layer;

[0040] The input end of the second fully connected layer is respectively connected to the output ends of the fourth convolutional block, the fifth convolutional block, and the sixth convolutional block;

[0041] The input end of the third fully connected layer is respectively connected to the output end of the first fully connected layer and the output end of the second fully connected layer, and its output end serves as the output end of the image quality evaluation model.

[0042] Further, the expressions of the Y-channel sharpness feature fusion unit, the U-channel sharpness feature fusion unit, and the V-channel sharpness feature fusion unit are: , where R is the output of the Y-channel sharpness feature fusion unit, the U-channel sharpness feature fusion unit, or the V-channel sharpness feature fusion unit, G 1 is the first edge sharpness matrix, G 2 is the second edge sharpness matrix, is element-wise addition, and Conv is convolution operation.

[0043] The beneficial effects of the present invention are:

[0044] 1. The present invention converts an image into the YUV space, separates luminance and color, forms a Y-channel image through luminance Y, a U-channel image through blue chrominance U, and a V-channel image through red chrominance V. By analyzing the edge sharpness of the picture from three aspects of luminance Y, blue chrominance U, and red chrominance V, it can capture the edge features of the image from multiple angles and comprehensively, avoiding the limitations of single-channel analysis.

[0045] 2. The present invention performs various 4-neighborhood Laplacian operator operations on each channel image and various 8-neighborhood Laplacian operator operations on each channel image. The 4-neighborhood Laplacian operator mainly focuses on the relationship between a pixel point and its 4 adjacent pixels above, below, left, and right, and can highlight the details and edges at relatively small scales in the image, having a better enhancement effect on some fine lines, textures, and other details in the image. The 8-neighborhood Laplacian operator, on the other hand, considers the adjacent pixels in 8 directions around the pixel point, can capture larger-scale image features and edge information, and has a better outlining effect on the contours of objects and large-area edges in the image. By performing these two operations on each channel image simultaneously, it can comprehensively capture the detailed information of the image from different scales, enabling the enhanced image to retain rich fine textures while highlighting obvious edges and contours, improving the overall sharpness and detail expressiveness of the image.

[0046] 3. The present invention calculates the edge sharpness of the first-channel detail-enhanced image and the second-channel detail-enhanced image respectively. After the detail enhancement of different channel images, their edge sharpness reflects the edge features of the image under different attributes of luminance and chrominance. Then, the first edge sharpness matrix and the second edge sharpness matrix corresponding to the Y, U, and V channel images are processed using a picture quality assessment model to obtain a picture quality score, obtaining the picture quality score from multiple dimensions and improving the assessment accuracy of the picture quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is a flowchart of an AI picture automatic screening method for advertising promotion based on an online network platform;

[0048] Figure 2 is a schematic structural diagram of a picture quality assessment model. DETAILED DESCRIPTION OF THE INVENTION

[0049] The following describes the specific embodiments of the present invention to facilitate those skilled in the art of this technology to understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art of this technology, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions created using the concept of the present invention are within the scope of protection.

[0050] As Figure 1 shown, an automatic screening method for AI images in advertising promotion based on an online network platform includes the following steps:

[0051] S1. Convert the images on the online network platform to the YUV space to obtain multiple channel images, where the multiple channel images include: a Y-channel image, a U-channel image, and a V-channel image;

[0052] S2. Perform Laplacian operator operations on each channel image in multiple 4-neighborhood and 8-neighborhood manners to construct a first-channel detail enhancement image and a second-channel detail enhancement image;

[0053] S3. Calculate the edge sharpness for each region in the first-channel detail enhancement image and the second-channel detail enhancement image respectively to construct a first edge sharpness matrix and a second edge sharpness matrix;

[0054] S4. Use an image quality assessment model to process the first edge sharpness matrix and the second edge sharpness matrix corresponding to the Y, U, and V channel images to obtain an image quality score;

[0055] S5. When the image quality score is greater than the score threshold, mark the image as the recommended status.

[0056] In this embodiment, the score threshold is specifically set according to experiments or experience.

[0057] In this embodiment, the Laplacian operator in the 4-neighborhood in S2 includes: a standard 4-neighborhood Laplacian operator , a modified 4-neighborhood Laplacian operator ; the Laplacian operator in the 8-neighborhood includes: a standard 8-neighborhood Laplacian operator , a gradient 8-neighborhood Laplacian operator .

[0058] The Laplacian operator is essentially a second-order derivative operator used to detect the mutation of pixel values in an image, that is, the edge. The standard 4-neighborhood Laplacian operator only considers the adjacent pixels in the four directions of up, down, left, and right, is more sensitive to edges in other directions, can highlight the simple lines and texture details in the image, and enhance the edge features at small scales. The modified 4-neighborhood Laplacian operator increases the weight in the diagonal direction, can detect edges more comprehensively, especially has a better capture effect on oblique details and edges, and helps to enrich the detail information of the image.

[0059] The Laplacian operator of the 8-neighborhood considers pixel information in more directions compared to the 4-neighborhood, and can detect the edges and features of an image on a larger scale. The standard 8-neighborhood Laplacian operator has a better effect on outlining the overall contour of objects and large-area edges in the image; the gradient 8-neighborhood Laplacian operator is more sensitive to regions with large pixel value changes through weight setting, which helps to detect edges with high contrast in the image and achieve multi-scale image analysis, mining the structural information of the image from different levels.

[0060] In this embodiment, S2 includes the following sub-steps:

[0061] S21. Perform Laplacian operator operations of multiple 4-neighborhoods on each channel image respectively, and fuse the operation results to obtain a first-channel detail enhancement image;

[0062] S22. Perform Laplacian operator operations of multiple 8-neighborhoods on each channel image respectively, and fuse the operation results to obtain a second-channel detail enhancement image.

[0063] In this embodiment, S21 includes the following sub-steps:

[0064] S211. Perform standard 4-neighborhood Laplacian operator and modified 4-neighborhood Laplacian operator operations on each channel image respectively to obtain a first detail image and a second detail image;

[0065] In S211, when performing the standard 4-neighborhood Laplacian operator operation on the channel image, replace the original pixel value in the channel image with the operation result to obtain a first detail image; when performing the modified 4-neighborhood Laplacian operator operation on the channel image, replace the original pixel value in the channel image with the operation result to obtain a second detail image;

[0066] S212. Take the absolute value of each pixel value in the first detail image and take the absolute value of each pixel value in the second detail image;

[0067] S213. Add the pixel values of the first detail image after taking the absolute value and the second detail image after taking the absolute value at the same pixel points to obtain a first-channel detail enhancement image. When the channel image in S211 is the Y-channel image, the first-channel detail enhancement image is the first Y-channel detail enhancement image; when the channel image in S211 is the U-channel image, the first-channel detail enhancement image is the first U-channel detail enhancement image; when the channel image in S211 is the V-channel image, the first-channel detail enhancement image is the first V-channel detail enhancement image.

[0068] The present invention respectively uses standard 4-neighborhood and modified 4-neighborhood Laplacian operator operations. The former is sensitive to edges in the horizontal and vertical directions, and the latter can take into account diagonal details. The combination of the two can capture detail information in the channel image from different angles, so that the generated first and second detail images contain rich image features. For example, in an advertising image, it can better display product texture and contour details. The present invention takes the absolute value of the pixel values of the first and second detail images to avoid losing detail information due to the cancellation of positive and negative values in the Laplacian operation results, ensuring the stability and integrity of the detail features and making the enhancement effect of the image details more reliable. The present invention adds the corresponding pixel values of the two images after taking the absolute value, fuses the detail information obtained by the two operator operations, further enhances the detail effect, and makes the generated channel detail-enhanced image more prominent in details in all directions.

[0069] In this embodiment, S22 includes the following sub-steps:

[0070] S221. Respectively perform standard 8-neighborhood Laplacian operator and modified 8-neighborhood Laplacian operator operations on each channel image to obtain a third detail image and a fourth detail image;

[0071] S222. Take the absolute value of each pixel value in the third detail image and take the absolute value of each pixel value in the fourth detail image;

[0072] S223. Add the pixel values of the third detail image after taking the absolute value and the fourth detail image after taking the absolute value at the same pixel points to obtain a second channel detail-enhanced image. When the channel image in S221 is the Y channel image, the second channel detail-enhanced image is the second Y channel detail-enhanced image. When the channel image in S221 is the U channel image, the second channel detail-enhanced image is the second U channel detail-enhanced image. When the channel image in S221 is the V channel image, the second channel detail-enhanced image is the second V channel detail-enhanced image.

[0073] The standard 8-neighborhood and modified 8-neighborhood Laplacian operators consider the information in 8 directions around the pixel and can detect edges and details in all directions more comprehensively. The combination of the two can extract rich and complete edge and detail information in the channel image. For example, in a landscape advertising picture, it can better restore the contour and texture details of scenery such as mountains and rivers. Taking the absolute value of the pixel values of the third and fourth detail images prevents the problem of losing detail information due to the cancellation of positive and negative values in the Laplacian operation results. It ensures that the detail features in different regions of the image can be effectively retained, improves the reliability of detail enhancement, and enables the fine details in the image to be clearly presented. The present invention adds the corresponding pixel values of the two images after taking the absolute value, fuses the detail information obtained by the two 8-neighborhood operator operations, and further strengthens the image details. The generated second channel detail-enhanced image highlights the image contour.

[0074] In this embodiment, S3 includes the following sub-steps:

[0075] S31. Calculate the edge sharpness for each region in the first-channel detail-enhanced image and construct a first edge sharpness matrix;

[0076] S32. Calculate the edge sharpness for each region in the second-channel detail-enhanced image and construct a second edge sharpness matrix.

[0077] In this embodiment, the first-channel detail-enhanced image and the second-channel detail-enhanced image can be divided into regions. Horizontally divide the first-channel detail-enhanced image and the second-channel detail-enhanced image equally at intervals M - 1 times, and vertically divide the first-channel detail-enhanced image and the second-channel detail-enhanced image equally at intervals L - 1 times. The sizes of the constructed first edge sharpness matrix and second edge sharpness matrix are: .

[0078] In this embodiment, the steps of constructing the edge sharpness matrix in both S31 and S32 include the following steps:

[0079] A1. Mark the pixel points in the channel detail-enhanced image with pixel values greater than the pixel threshold as edge points;

[0080] A2. Divide the channel detail-enhanced image into multiple regions;

[0081] A3. In each region, connect the adjacent edge points in the horizontal direction to form multiple horizontal edge lines;

[0082] A4. In each region, connect the adjacent edge points in the vertical direction to form multiple vertical edge lines;

[0083] A5. Calculate the edge sharpness of the region according to the multiple horizontal edge lines and multiple vertical edge lines;

[0084] A6. Use the edge sharpness of the region as an element to construct an edge sharpness matrix.

[0085] The pixel threshold is specifically set according to experiments or experience. In this embodiment, the pixel threshold can be set as the pixel mean value in the channel detail-enhanced image.

[0086] In this embodiment, the formula for calculating the edge sharpness of the region in A5 is: , where θ is the edge sharpness of the region, N level,i is the number of edge points on the i-th horizontal edge line, N ver,i is the number of edge points on the i-th vertical edge line, i is a positive integer, L levelis the number of horizontal edge lines in a region, L ver is the number of vertical edge lines in a region.

[0087] In the present invention, pixel points with pixel values greater than the pixel threshold are marked as edge points to mark the edge positions of the image. Steps A2 - A4 form horizontal and vertical edge lines by dividing regions and connecting adjacent edge points, further clearly measuring the widths of the edge lines in the horizontal and vertical directions. The present invention considers the number of edge lines in the horizontal and vertical directions and the number of edge points on each line, calculating the edge sharpness of the region from multiple dimensions. This can comprehensively reflect the texture distribution of the image in a region, the refinement of the texture. The greater the edge sharpness, the finer the texture and the better the quality of the picture.

[0088] As Figure 2 shown, the picture quality evaluation model in S4 includes: Y-channel sharpness feature fusion unit, U-channel sharpness feature fusion unit, V-channel sharpness feature fusion unit, first convolutional block, second convolutional block, third convolutional block, fourth convolutional block, fifth convolutional block, sixth convolutional block, first fully connected layer, second fully connected layer, and third fully connected layer;

[0089] The input end of the Y-channel sharpness feature fusion unit is used to input the first edge sharpness matrix and the second edge sharpness matrix corresponding to the Y-channel image;

[0090] The input end of the U-channel sharpness feature fusion unit is used to input the first edge sharpness matrix and the second edge sharpness matrix corresponding to the U-channel image;

[0091] The input end of the V-channel sharpness feature fusion unit is used to input the first edge sharpness matrix and the second edge sharpness matrix corresponding to the V-channel image;

[0092] The input end of the first convolutional block is connected to the output end of the Y-channel sharpness feature fusion unit, and its output end is respectively connected to the input end of the fourth convolutional block and the input end of the first fully connected layer;

[0093] The input end of the second convolutional block is connected to the output end of the U-channel sharpness feature fusion unit, and its output end is respectively connected to the input end of the fifth convolutional block and the input end of the first fully connected layer;

[0094] The input end of the third convolutional block is connected to the output end of the V-channel sharpness feature fusion unit, and its output end is respectively connected to the input end of the sixth convolutional block and the input end of the first fully connected layer;

[0095] The input end of the second fully connected layer is respectively connected to the output ends of the fourth convolutional block, the output end of the fifth convolutional block, and the output end of the sixth convolutional block;

[0096] The input end of the third fully connected layer is respectively connected to the output ends of the first fully connected layer and the second fully connected layer, and its output end serves as the output end of the picture quality evaluation model.

[0097] In the YUV color space, the Y channel represents luminance, and the U and V channels represent chrominance. Each channel sharpness feature fusion unit processes the features of the corresponding channel, fuses the corresponding features of the first edge sharpness matrix and the second edge sharpness matrix of the affiliated channel, then uses the convolutional blocks of the first layer: the first convolutional block, the second convolutional block, and the third convolutional block for shallow feature extraction, and then uses the second convolutional block: the fourth convolutional block, the fifth convolutional block, and the sixth convolutional block for deep feature extraction. The first fully connected layer synthesizes all shallow features, and the second fully connected layer synthesizes all deep features, realizing the synthesis of shallow and deep features, obtaining the picture quality score, and improving the accuracy of picture quality evaluation.

[0098] In this embodiment, the expressions of the Y channel sharpness feature fusion unit, the U channel sharpness feature fusion unit, and the V channel sharpness feature fusion unit are: , where R is the output of the Y channel sharpness feature fusion unit, the U channel sharpness feature fusion unit, or the V channel sharpness feature fusion unit, and G 1 is the first edge sharpness matrix, and G 2 is the second edge sharpness matrix, is element-wise addition, and Conv is the convolution operation.

[0099] In this embodiment, the convolutional block includes: a convolutional layer, a batch normalization layer, and an activation function layer.

[0100] In this embodiment, the third fully connected layer is used to output the picture quality score.

[0101] The first edge sharpness matrix and the second edge sharpness matrix respectively reflect the edge sharpness information of the channel image from different angles. The features in the matrix are extracted through the convolution operation (Conv). Then, through the element-wise addition operation, the features after convolution of the two matrices are fused, realizing the complementarity of the features.

[0102] The present invention converts the image into the YUV space, separates luminance and color, forms the Y channel image through the luminance Y, forms the U channel image through the blue chrominance U, and forms the V channel image through the red chrominance V. By analyzing the edge sharpness of the picture from three aspects: luminance Y, blue chrominance U, and red chrominance V, it can capture the edge features of the image from multiple angles and comprehensively, avoiding the limitations of single-channel analysis.

[0103] The present invention performs a variety of Laplacian operator operations on the 4-neighborhood for each channel image and a variety of Laplacian operator operations on the 8-neighborhood for each channel image. The 4-neighborhood Laplacian operator mainly focuses on the relationship between a pixel point and its 4 adjacent pixels above, below, left, and right, and can highlight the details and edges at relatively small scales in the image, having a better enhancement effect on some fine lines, textures, and other details in the image. The 8-neighborhood Laplacian operator, on the other hand, considers the adjacent pixels in 8 directions around the pixel point, can capture image features and edge information at a larger scale, and has a better outlining effect on the contours of objects and large-area edges in the image. By simultaneously performing these two operations on each channel image, the detailed information of the image can be comprehensively captured from different scales, enabling the enhanced image to retain rich fine textures while highlighting obvious edges and contours, thereby improving the overall clarity and detail expressiveness of the image.

[0104] The present invention calculates the edge sharpness for the first-channel detail-enhanced image and the second-channel detail-enhanced image respectively. After the detail enhancement of different channel images, their edge sharpness reflects the edge features of the image under different attributes of brightness and chromaticity. Then, the first-edge sharpness matrix and the second-edge sharpness matrix corresponding to the Y, U, and V channel images are processed using a picture quality assessment model to obtain a picture quality score, obtaining the picture quality score from multiple dimensions and improving the assessment accuracy of the picture quality.

[0105] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. An automatic screening method for advertising promotion AI pictures based on an online network platform, characterized in that: The following steps are involved: S1. Convert the image on the online network platform into YUV space to obtain multiple channel images, wherein the multiple channel images include: Y channel image, U channel image and V channel image; S2, performing a variety of 4-neighborhood and 8-neighborhood Laplacian operator operations on each channel image, respectively, to construct a first channel detail enhanced image and a second channel detail enhanced image; The S2 comprises the following sub-steps: S21, performing a plurality of 4-neighborhood Laplacian operator operations on each channel image respectively, fusing the images after the operations to obtain a first channel detail enhanced image; S22, performing multiple 8-neighborhood Laplacian operator operations on each channel image respectively, and fusing the images after the operations to obtain a second channel detail enhanced image; The S21 comprises the following sub-steps: S211, performing standard 4-neighborhood Laplacian operator and modified 4-neighborhood Laplacian operator operations on each channel image respectively to obtain a first detail image and a second detail image; S212, taking the absolute value of each pixel value in the first detail image, and taking the absolute value of each pixel value in the second detail image; S213, adding the pixel values ​​of the first detail image after taking the absolute value and the second detail image after taking the absolute value according to the same pixel points to obtain a first channel detail enhanced image, when the channel image in S211 is a Y channel image, the first channel detail enhanced image is a first Y channel detail enhanced image, when the channel image in S211 is a U channel image, the first channel detail enhanced image is a first U channel detail enhanced image, when the channel image in S211 is a V channel image, the first channel detail enhanced image is a first V channel detail enhanced image; S3, respectively calculating the edge clarity of each region in the first channel detail enhanced image and the second channel detail enhanced image, and constructing a first edge clarity matrix and a second edge clarity matrix; S4, using the picture quality assessment model to process the first edge clarity matrix and the second edge clarity matrix corresponding to the Y, U and V channel images to obtain a picture quality score; S5. When the image quality score is greater than the score threshold, mark the image as recommended.

2. The method for automatically screening advertising promotion AI pictures based on an online network platform according to claim 1 is characterized in that: The 4-neighborhood Laplacian operator in S2 includes: standard 4-neighborhood Laplacian operator , the modified 4-neighborhood Laplacian ; 8-neighborhood Laplacian operators include: Standard 8-neighborhood Laplacian operator , gradient 8-neighborhood Laplacian .

3. The method for automatically screening advertising promotion AI pictures based on an online network platform according to claim 1 is characterized in that: The S22 comprises the following sub-steps: S221, performing standard 8-neighborhood Laplacian operator and modified 8-neighborhood Laplacian operator operations on each channel image respectively to obtain a third detail image and a fourth detail image; S222, taking the absolute value of each pixel value in the third detail image, and taking the absolute value of each pixel value in the fourth detail image; S223. Add the pixel values ​​of the third detail image after taking the absolute value and the fourth detail image after taking the absolute value at the same pixel points to obtain a second channel detail enhanced image. When the channel image in S221 is a Y channel image, the second channel detail enhanced image is a second Y channel detail enhanced image. When the channel image in S221 is a U channel image, the second channel detail enhanced image is a second U channel detail enhanced image. When the channel image in S221 is a V channel image, the second channel detail enhanced image is a second V channel detail enhanced image.

4. The method for automatically screening advertising promotion AI pictures based on an online network platform according to claim 1 is characterized in that: The S3 comprises the following sub-steps: S31, calculating edge clarity of each region in the first channel detail enhanced image, and constructing a first edge clarity matrix; S32, calculating edge clarity for each region in the second channel detail enhanced image, and constructing a second edge clarity matrix.

5. The method for automatically screening advertising promotion AI pictures based on an online network platform according to claim 4 is characterized in that: The edge definition matrix constructed in S31 and S32 includes the following steps: A1. Pixels whose pixel values ​​in the channel detail enhanced image are greater than the pixel threshold are marked as edge points; A2, dividing the channel detail enhanced image into multiple regions; A3. In each region, connect each adjacent edge point in the horizontal direction to form multiple horizontal edge lines; A4. In each region, connect each adjacent edge point in the vertical direction to form multiple vertical edge lines; A5. Calculate edge definition of the region according to the plurality of horizontal edge lines and the plurality of vertical edge lines; A6. Use the edge clarity of the region as an element to construct an edge clarity matrix.

6. The method for automatically screening advertising promotion AI pictures based on an online network platform according to claim 5 is characterized in that: The formula for calculating the edge definition of the area in A5 is: , where θ is the edge clarity of the region, N level,i is the number of edge points on the i-th horizontal edge line, N ver,i is the number of edge points on the i-th vertical edge line, i is a positive integer, L level is the number of horizontal edge lines in a region, L ver is the number of vertical edge lines in a region.

7. The method for automatically screening advertising promotion AI pictures based on an online network platform according to claim 1, characterized in that: The picture quality assessment model in S4 includes: a Y channel clarity feature fusion unit, a U channel clarity feature fusion unit, a V channel clarity feature fusion unit, a first convolution block, a second convolution block, a third convolution block, a fourth convolution block, a fifth convolution block, a sixth convolution block, a first fully connected layer, a second fully connected layer and a third fully connected layer; The input end of the Y channel definition feature fusion unit is used to input the first edge definition matrix and the second edge definition matrix corresponding to the Y channel image; The input end of the U channel definition feature fusion unit is used to input the first edge definition matrix and the second edge definition matrix corresponding to the U channel image; The input end of the V channel definition feature fusion unit is used to input a first edge definition matrix and a second edge definition matrix corresponding to the V channel image; The input end of the first convolution block is connected to the output end of the Y channel clarity feature fusion unit, and its output end is connected to the input end of the fourth convolution block and the input end of the first fully connected layer respectively; The input end of the second convolution block is connected to the output end of the U channel clarity feature fusion unit, and its output end is connected to the input end of the fifth convolution block and the input end of the first fully connected layer respectively; The input end of the third convolution block is connected to the output end of the V channel clarity feature fusion unit, and its output end is connected to the input end of the sixth convolution block and the input end of the first fully connected layer respectively; The input end of the second fully connected layer is respectively connected to the output end of the fourth convolution block, the output end of the fifth convolution block, and the output end of the sixth convolution block; The input end of the third fully connected layer is connected to the output end of the first fully connected layer and the output end of the second fully connected layer respectively, and the output end thereof serves as the output end of the picture quality assessment model.

8. The method for automatically screening advertising promotion AI pictures based on an online network platform according to claim 7 is characterized in that: The expressions of the Y channel definition feature fusion unit, the U channel definition feature fusion unit and the V channel definition feature fusion unit are: , where R is the output of the Y channel definition feature fusion unit, the U channel definition feature fusion unit or the V channel definition feature fusion unit, G1 is the first edge definition matrix, G2 is the second edge definition matrix, is element-wise addition, and Conv is a convolution operation.

Images