An online advertisement content security monitoring system based on image recognition

Abstract

The present application relates to the technical field of image recognition, and particularly relates to an online advertisement content security monitoring system based on image recognition, comprising: an image determination module, which is used to pre-process collected online advertisement images to obtain standardized advertisement images, and determine the standardized advertisement images as regular images or synthetic images according to local noise residual difference amounts; a text extraction module, which is used to extract advertisement content in the regular images to obtain text content; an analysis and deconstruction module, which is used to analyze the synthetic images according to distribution characteristics of pixel change amounts to output classification labels, and identify the classification labels to deconstruct corresponding synthetic regions to obtain text content; and a risk processing module, which is used to judge the text content to obtain identification results, and process illegal content. The present application improves the monitoring accuracy of online advertisement illegal content.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and in particular to an online advertising content security monitoring system based on image recognition. Background Technology

[0002] With the rapid development of the internet, digital media has gradually become an important channel for information dissemination, and the advertising industry has undergone profound changes. Against the backdrop of the gradual decline of traditional advertising models, real-time monitoring systems have emerged, becoming a major innovation in the advertising industry. These systems utilize big data, artificial intelligence, and other technologies to achieve real-time tracking and analysis of advertising effectiveness, providing advertisers and advertising companies with precise decision-making support for advertising placement. Real-time monitoring systems have made significant progress in data collection, processing, and analysis. The application of advanced technologies such as big data, artificial intelligence, and cloud computing has enabled the system to achieve higher accuracy and efficiency in processing massive amounts of data, mining user behavior, and predicting advertising effectiveness. However, existing advertising content security detection systems cannot address the problem of identification errors caused by illegal information being mixed with normal content through layer overlay.

[0003] Chinese Patent Publication No. CN120656176A discloses a method and system for monitoring the content security of billboards on highways. The method includes: S1, acquiring the location information of billboards set up on the highway; periodically taking pictures of billboards set up on the highway using deployed photovoltaic-powered cameras, wherein the cameras communicate with a cloud recognition platform via a 5G network and upload the location information of the corresponding cameras and the billboard images taken each time; S2, the cloud recognition platform receives the billboard images uploaded by the cameras, extracts image features from the billboard images using image recognition technology, performs content compliance analysis on the billboard images, identifies the set of violations corresponding to the billboard images, and binds the identified violations with the corresponding billboard images, feeding them back to the display end; S3, acquiring the shooting environment information and remaining power value of the photovoltaic-powered cameras deployed on the highway at the current time; assessing the battery life crisis coefficient of each photovoltaic-powered camera deployed on the highway based on the current time's weather forecast information; and calculating the feedback demand assessment value of each deployed billboard based on the set of violations corresponding to the images taken by each deployed billboard in historical data. Therefore, it can be seen that the method and system for monitoring the security of billboard content on highways has the problem that, because it only analyzes the advertising images using conventional image content recognition technology, it lacks analysis of the authenticity of the advertising images, which makes it impossible to determine whether the advertising images have been tampered with, and thus makes it impossible to identify hidden information in the advertising content. Summary of the Invention

[0004] To address this issue, the present invention provides an online advertising content security monitoring system based on image recognition, which overcomes the problem in the prior art that, due to the lack of analysis on the authenticity of advertising images, it is impossible to determine whether there are multi-layer composite areas in the advertising content, thus failing to identify hidden information in the advertising content.

[0005] To achieve the above objectives, the present invention provides an online advertising content security monitoring system based on image recognition, comprising: The image determination module is used to preprocess the acquired online advertising images to obtain standardized advertising images, calculate the local noise residual difference of the standardized advertising images, and determine whether the standardized advertising images are regular images or synthetic images based on the local noise residual difference. The text extraction module is used to extract the advertising content from the regular image to obtain the text content; The analysis and deconstruction module is used to obtain the multi-layer composite region in the composite image, calculate the pixel change of the multi-layer composite region, analyze the composite image according to the distribution characteristics of the pixel change to output classification labels, and identify the classification labels to deconstruct the corresponding composite region to obtain the text content. The risk processing module is used to assess the risk of the text content to obtain identification results and to process the illegal content.

[0006] Furthermore, the image determination module is used to determine that the standardized advertising image is a regular image when the local noise residual difference is less than or equal to a preset difference.

[0007] Furthermore, the image determination module is used to determine that the standardized advertising image is a synthetic image when the local noise residual difference is greater than the preset difference amount.

[0008] Furthermore, the text extraction module is used to sequentially perform edge recognition, segmentation, and content extraction on the conventional image to obtain text content.

[0009] Furthermore, the analysis and deconstruction module is used to obtain the pixel values ​​and position coordinates of several pixels within the multi-layer synthesis area of ​​the synthesized image; The analysis and deconstruction module is used to map the position coordinates to the corresponding background area in order to extract the pixel values ​​of the corresponding background area pixels; The analysis and deconstruction module is used to determine the pixel change in the multi-layer composite area based on the absolute difference between the pixel values ​​of several pixels in the multi-layer composite area and the pixel values ​​of the corresponding background area pixels.

[0010] Furthermore, the analysis and deconstruction module is used to determine the layer composition type as a blended layer composition when the pixel change is continuously distributed, and output a blending classification label.

[0011] Furthermore, the analysis and deconstruction module is used to determine that the layer composition type is an embedded layer composition when the pixel change is concentrated, and outputs an embedded classification label.

[0012] Furthermore, the analysis and deconstruction module is used to reconstruct the text and background in the corresponding synthetic area based on the text pixel gradient when the classification label is a fused classification label in order to extract the text content.

[0013] Furthermore, the analysis and deconstruction module is used to identify the embedded region to extract text content when the classification label is an embedded classification label.

[0014] Furthermore, the risk processing module is used to perform content parsing and semantic analysis on the text content to obtain text features, and to determine the risk status of the text content based on the text features to output the recognition result.

[0015] Compared with existing technologies, the beneficial effects of this invention are as follows: The system of this invention, by setting up an image judgment module, identifies whether there are synthetic images in standardized advertising images, thereby determining the target image that may hide illegal information, effectively narrowing the scope of subsequent in-depth analysis, and providing a theoretical basis for determining the existence of hidden content. By setting up an analysis and deconstruction module, it identifies the type of synthetic image, reveals the information carrying capacity and spatial distribution of hidden information in different layer synthetic areas, improves the extraction efficiency of hidden content in synthetic images, identifies layer synthetic type labels and performs targeted deconstruction on different layer synthetic types, extracts hidden information in synthetic images, ensures that the synthetic text content can be completely restored and accurately captured, enhances the reliability of advertising content monitoring, avoids the problem of missing the detection of hidden illegal information in advertising images when only analyzing advertising images through conventional image content recognition technology, and improves the accuracy of monitoring illegal content in online advertisements.

[0016] Furthermore, the system of the present invention determines whether there are multi-layer composite regions in standardized advertising images by setting a preset difference amount. Since hidden illegal information in advertising images may be loaded into the original image through local layer overlay, transparent blending, or region embedding, the pixel differences between the composite and non-composite regions exhibit localized characteristics, resulting in a large number of invalid images entering the deep analysis process. By determining whether there are multi-layer composite regions in standardized advertising images, images that need deep analysis can be quickly screened out, and computing resources can be concentrated on images that need deep analysis to reduce subsequent computing resource waste, thereby reducing invalid calculations in the advertising monitoring process and further improving the accuracy of monitoring illegal advertising content.

[0017] Furthermore, the system described in this invention determines the layer composition type based on the distribution characteristics of pixel changes. Since hidden violation information is usually not directly displayed, but rather mixed with or embedded in the original image, the newly added content is difficult to obtain through conventional text recognition extraction methods. This causes the pixel change characteristics generated by different composition methods to be confused with each other, making it impossible to accurately distinguish the layer composition type. By determining the layer composition type of the composite image, incorrect deconstruction paths caused by type misjudgment can be avoided, solving the problem that the original image and hidden information in the advertising content are difficult to distinguish and cannot be identified, and further improving the accuracy of monitoring advertising violation content.

[0018] Furthermore, the system described in this invention extracts hidden content by identifying the image synthesis type. Since the combination relationship between the hidden content in the advertising image and the original image is affected by the synthesis method, using a uniform extraction method for content extraction can easily lead to the loss of some hidden information, resulting in the inability to fully recover hidden text areas or implicit violations. By identifying the image synthesis type and performing targeted deconstruction, the hidden content in the synthesized image can be extracted, which can effectively improve the completeness of the restoration of hidden information and ensure that violations under various hidden tampering methods can be effectively restored. This is beneficial for the detection and processing of real violations and further improves the monitoring accuracy of advertising violations. Attached Figure Description

[0019] Figure 1 This is an overall structural block diagram of the online advertising content security monitoring system based on image recognition, as described in an embodiment of the present invention. Figure 2 This is a flowchart illustrating the process by which the online advertising content security monitoring system based on image recognition determines whether a standardized advertising image is a regular image or a composite image, according to an embodiment of the present invention. Figure 3This is a flowchart illustrating the process by which the image recognition-based online advertising content security monitoring system analyzes synthetic images and outputs classification tags, as described in an embodiment of the present invention. Figure 4 This is a flowchart illustrating the process of identifying classification tags and performing deconstruction of corresponding synthetic regions in an online advertising content security monitoring system based on image recognition, as described in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0021] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0022] Please see Figure 1 The diagram shown is an overall structural block diagram of an online advertising content security monitoring system based on image recognition, according to an embodiment of the present invention. The system described in this embodiment includes: The image determination module is used to preprocess the acquired online advertising images to obtain standardized advertising images, calculate the local noise residual difference of the standardized advertising images, and determine whether the standardized advertising images are regular images or synthetic images based on the local noise residual difference. A text extraction module, which is connected to the image determination module, is used to extract the advertising content from the regular image to obtain the text content; The analysis and deconstruction module, which is connected to the image determination module, is used to obtain the multi-layer composite region in the composite image, calculate the pixel change of the multi-layer composite region, analyze the composite image according to the distribution characteristics of the pixel change to output a classification label, and identify the classification label to deconstruct the corresponding composite region to obtain the text content. The risk processing module is connected to the text extraction module and the analysis and deconstruction module respectively, and is used to perform risk assessment on the text content to obtain identification results and process the illegal content.

[0023] Specifically, online advertising images include service promotion ads, product ads, and event marketing ads.

[0024] Specifically, preprocessing includes cleaning, denoising, image enhancement, and standardization of online advertising images.

[0025] Specifically, standardized advertising images include service promotion advertisements after cleaning, product advertisements after noise reduction, and event marketing advertisements after image enhancement.

[0026] Specifically, the identification results include security status results, content risk results, and hidden risk results.

[0027] Specifically, the safety status results include normal safety status and violation status.

[0028] Specifically, a normal safety status is when the text content and image structure in the advertisement image do not match the risk features in the risk type feature library; The violation status is that the text content and image structure in the advertisement image are detected to match the risk features in the risk type feature library.

[0029] Specifically, the risk features include keyword features, semantic association features, and image structure features.

[0030] Specifically, the risk type feature library includes risk features of false advertising, fraudulent inducement, and vulgar and harmful content.

[0031] Specifically, content risk outcomes include false advertising, fraudulent inducement, and vulgar or inappropriate content.

[0032] Specifically, hidden risk outcomes include no hidden risk, hidden fraud risk, and hidden inducement risk.

[0033] Specifically, the process of risk assessment of text content to obtain recognition results involves acquiring text features, extracting keywords, contextual semantic relationships, and behavioral expression information from the text features, comparing the extracted information with a risk type feature database, locating the area where the text is located, and generating recognition results.

[0034] Specifically, the process for handling violations is as follows: Obtain the content from the identification results, determine the location of the violation content based on the identification results, and determine the processing target based on the source of the violation content; For violations in regular images, the corresponding areas are marked and removed or corrected. For violations in the composite image, layer clearing is performed based on the location of the violation content to obtain the processed advertisement image.

[0035] In this embodiment, the system uses an image determination module to identify whether a composite image exists in a standardized advertising image, thereby identifying target images that may hide illegal information. This effectively narrows the scope of subsequent in-depth analysis and provides a theoretical basis for determining the existence of hidden content. An analysis and deconstruction module identifies the type of composite image, revealing the information carrying capacity and spatial distribution of hidden information in different layer composite areas, improving the extraction efficiency of hidden content in composite images. The system also identifies layer composite type labels and performs targeted deconstruction on different layer composite types to extract hidden information from composite images, ensuring that the composite text content can be completely restored and accurately captured. This enhances the reliability of advertising content monitoring and avoids the problem of overlooking the detection of hidden illegal information in advertising images when only conventional image content recognition technology is used to analyze them, thus improving the accuracy of monitoring illegal content in online advertising.

[0036] Please see Figure 2 As shown, it is a logical flowchart of the process of determining whether a standardized advertising image is a conventional image or a synthetic image in an embodiment of the present invention.

[0037] Specifically, the image determination module is used to determine that the standardized advertising image is a regular image when the local noise residual difference is less than or equal to a preset difference.

[0038] Specifically, the image determination module is used to determine that the standardized advertising image is a synthetic image when the local noise residual difference is greater than the preset difference amount.

[0039] Specifically, the local noise residual difference is the difference between the statistical value of the local noise residual and the statistical value of the neighboring noise residual.

[0040] Specifically, the noise residual statistics are the standard deviations of the differences between the pixel values ​​before and after denoising in the image.

[0041] Understandably, the process of obtaining the preset difference amount involves dividing the compliant advertising image into regions to calculate the difference between the local noise residual and the noise residual of adjacent regions, calculating the standard deviation of the difference, and finding the maximum value of the standard deviation as the preset difference amount. Specifically, when the local noise residual difference amount is greater than the preset difference amount, the noise distribution of the region image shows a significant abrupt change, indicating the presence of multi-layer compositing. When the local noise residual difference amount is less than or equal to the preset difference amount, the noise distribution of the region image is relatively uniform, indicating that it represents the original layer content of the image and that multi-layer compositing does not exist.

[0042] Specifically, compliant advertising images are those whose text content meets security requirements after preprocessing and do not have a multi-layer composite structure.

[0043] In this embodiment, the system determines whether a standardized advertising image has a multi-layer composite region by setting a preset difference amount. Since hidden illegal information in the advertising image may be loaded into the original image through local layer overlay, transparent blending, or region embedding, the pixel difference between the composite region and the non-composite region exhibits localized characteristics, resulting in a large number of invalid images entering the deep analysis process. By determining whether a standardized advertising image has a multi-layer composite region, images that need deep analysis can be quickly screened out, and computing resources can be concentrated on the images that need deep analysis to reduce the waste of subsequent computing resources. This reduces invalid calculations in the advertising monitoring process and further improves the accuracy of monitoring illegal advertising content.

[0044] Please see Figure 3 As shown, it is a logical flowchart of the process of analyzing the synthesized image and outputting classification labels in an embodiment of the present invention.

[0045] Specifically, the analysis and deconstruction module is used to obtain the multi-layer composite region in the composite image and calculate the pixel change of each pixel in the multi-layer composite region relative to the corresponding background region. Based on the distribution characteristics of the pixel change, the layer composite type is determined and a classification label is output.

[0046] Specifically, the analysis and deconstruction module is used to determine the layer composition type as a blended layer composition when the pixel change is continuously distributed, and output a blending classification label.

[0047] Specifically, the analysis and deconstruction module is used to determine that the layer composition type is an embedded layer composition when the pixel change is concentrated, and outputs an embedded classification label.

[0048] Specifically, the process of obtaining the multi-layer composite region in the composite image involves determining the abnormal region based on the amount of local residual difference.

[0049] Specifically, a continuous distribution of pixel changes means that the pixel change values ​​continuously cover each area within the multi-layer composite area, and none of the pixel change values ​​are greater than the change value threshold.

[0050] Specifically, the pixel change is concentrated in the multi-layer composite area where the pixel change value is greater than the change value threshold and is concentrated in the image character structure area, while the pixel change value in other areas is not greater than the change value threshold.

[0051] Specifically, the image character structure region is the area occupied by the characters in the composite image, determined by their outlines, stroke connections, and arrangement.

[0052] Specifically, the process for determining the threshold value of the change value is as follows: Calculate the pixel changes of natural images in a set of natural images; The distribution of pixel changes in each image is statistically analyzed, and the maximum value of pixel changes in each image is used as the change value threshold.

[0053] In this embodiment, the natural image set can be at least one of the CIFAR-10 set, CIFAR-100 set, AFHQ set, and FFHQ set.

[0054] In this embodiment, the system determines the layer composition type based on the distribution characteristics of pixel changes. Since hidden violation information is usually not directly displayed, but rather mixed with or embedded in the original image, the new content is difficult to obtain through conventional text recognition extraction methods. This causes the pixel change characteristics generated by different composition methods to be confused with each other, making it impossible to accurately distinguish the layer composition type. By determining the layer composition type of the composite image, the system can avoid incorrect deconstruction paths caused by type misjudgment, solve the problem of difficulty in distinguishing and identifying the original image and hidden information in the advertising content, and further improve the accuracy of monitoring advertising violation content.

[0055] Please see Figure 4 As shown, it is a logical flowchart of the process of identifying classification tags to perform the deconstruction of the corresponding synthetic region in an embodiment of the present invention.

[0056] Specifically, the analysis and deconstruction module is used to reconstruct the text and background in the corresponding synthetic area based on the text pixel gradient when the classification label is a fused classification label in order to extract the text content.

[0057] In this embodiment, the specific process of reconstructing the text and background within the corresponding synthetic area based on the text pixel gradient to extract the text content is as follows: When a fusion classification label is identified, the gradient direction distribution information in the adjacent region of the pixel is calculated based on the gradient characteristics of the text pixel in the corresponding abnormal region. By analyzing the continuity and consistency of the gradient direction between adjacent pixels, the foreground text layer and the background layer are distinguished. The foreground text region, due to its continuous character edge structure, exhibits high consistency with small gradient direction changes between adjacent pixels. In contrast, the background region, influenced by its texture structure, shows greater gradient direction changes between different pixels, exhibiting strong randomness. The pixels belonging to the foreground text layer are reconstructed according to their original spatial coordinates to generate the foreground text image, while the pixels belonging to the background layer are reconstructed to generate the background image. Further contrast enhancement processing is performed on the reconstructed foreground text image to improve the pixel feature difference between the semi-transparent text and the background, making the hidden text features more prominent; Finally, the enhanced foreground text image is input into the text recognition model for recognition, and the corresponding text sequence is obtained.

[0058] Specifically, the analysis and deconstruction module is used to identify the embedded region to extract text content when the classification label is an embedded classification label.

[0059] Specifically, the process of identifying the embedded region to extract text content is as follows: when the embedded classification label is identified, image recognition technology is used to perform routine recognition on the synthetic image and obtain routine results. The embedded region of the synthetic image is identified to obtain dense results. The dense results are merged with the routine results, duplicates are removed, and the text sequence is obtained.

[0060] In this embodiment, the system extracts hidden content by identifying the image synthesis type. Since the combination relationship between the hidden content in the advertising image and the original image is affected by the synthesis method, using a uniform extraction method for content extraction can easily lead to the loss of some hidden information, resulting in the inability to fully recover hidden text areas or implicit violations. By identifying the image synthesis type and performing targeted deconstruction, the hidden content in the synthesized image can be extracted, which can effectively improve the completeness of the restoration of hidden information and ensure that violations under various hidden tampering methods can be effectively restored. This is beneficial for the detection and processing of real violations and further improves the monitoring accuracy of advertising violations.

[0061] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. An online advertising content security monitoring system based on image recognition, characterized in that, include: The image determination module is used to preprocess the acquired online advertising images to obtain standardized advertising images, calculate the local noise residual difference of the standardized advertising images, and determine whether the standardized advertising images are regular images or synthetic images based on the local noise residual difference. The text extraction module is used to extract the advertising content from the regular image to obtain the text content; The analysis and deconstruction module is used to obtain the multi-layer composite region in the composite image, calculate the pixel change of the multi-layer composite region, analyze the composite image according to the distribution characteristics of the pixel change to output classification labels, and identify the classification labels to deconstruct the corresponding composite region to obtain the text content. The risk processing module is used to assess the risk of the text content to obtain identification results and to process the illegal content.

2. The online advertising content security monitoring system based on image recognition according to claim 1, characterized in that, The image determination module is used to determine that the standardized advertising image is a regular image when the local noise residual difference is less than or equal to a preset difference.

3. The online advertising content security monitoring system based on image recognition according to claim 2, characterized in that, The image determination module is used to determine that the standardized advertising image is a synthetic image when the local noise residual difference is greater than the preset difference amount.

4. The online advertising content security monitoring system based on image recognition according to claim 3, characterized in that, The text extraction module is used to sequentially perform edge recognition, segmentation, and content extraction on the conventional image to obtain the text content.

5. The online advertising content security monitoring system based on image recognition according to claim 3, characterized in that, The analysis and deconstruction module is used to obtain the pixel values ​​and position coordinates of several pixels within the multi-layer composite area of ​​the synthesized image; The analysis and deconstruction module is used to map the position coordinates to the corresponding background area in order to extract the pixel values ​​of the corresponding background area pixels; The analysis and deconstruction module is used to determine the pixel change in the multi-layer composite area based on the absolute difference between the pixel values ​​of several pixels in the multi-layer composite area and the pixel values ​​of the corresponding background area pixels.

6. The online advertising content security monitoring system based on image recognition according to claim 5, characterized in that, The analysis and deconstruction module is used to determine the layer composition type as a blended layer composition when the pixel change is continuously distributed, and output a blending classification label.

7. The online advertising content security monitoring system based on image recognition according to claim 6, characterized in that, The analysis and deconstruction module is used to determine that the layer composition type is an embedded layer composition when the pixel change is concentrated, and outputs an embedded classification label.

8. The online advertising content security monitoring system based on image recognition according to claim 7, characterized in that, The analysis and deconstruction module is used to reconstruct the text and background in the corresponding synthetic area based on the text pixel gradient when the classification label is a fused classification label, so as to extract the text content.

9. The online advertising content security monitoring system based on image recognition according to claim 8, characterized in that, The analysis and deconstruction module is used to identify the embedded region to extract text content when the classification label is an embedded classification label.

10. The online advertising content security monitoring system based on image recognition according to claim 9, characterized in that, The risk processing module is used to perform content parsing and semantic analysis on the text content to obtain text features, and to determine the risk status of the text content based on the text features to output the recognition result.

Images