Methods, systems, devices, and media for detecting suggestive content in images and videos.

Abstract

This invention discloses a method, system, device, and medium for detecting suggestive content in images and videos, relating to the field of video image recognition technology. The key technical points include the following steps: acquiring a complete video file to be processed; segmenting the complete video file to obtain at least one video slice; transforming the video slice using a preset analysis and transformation model to obtain a first video feature vector corresponding to each video slice; and constructing slice segmentation lines between every two adjacent video slices. This invention enables style analysis of video content, identification of embedded suggestive segments from normal videos, and filtering of suggestive segments within the video. It solves the problems of traditional methods being time-consuming and labor-intensive, and misidentifying parts of normal video content, thus affecting the efficiency and accuracy of filtering.

Description

Technical Field

[0001] This invention relates to the field of video image recognition technology, and more specifically, to methods, systems, devices, and media for detecting suggestive content in images and videos. Background Technology

[0002] With the rapid development of computer technology and the internet, artificial intelligence has become a crucial direction for technological development worldwide. In recent years, deep learning technology has achieved remarkable results in the field of artificial intelligence, providing strong support for various application scenarios. However, existing deep learning methods still have certain limitations when handling tasks such as filtering suggestive segments in videos. Specifically, the following approaches are used:

[0003] I. Identification using induced fragment markers

[0004] Some lead-in segments are labeled when they are generated, and we can use the labeled information to filter lead-in segments; however, in practice, there are relatively few videos with explicitly labeled segments.

[0005] II. Manually filtering suggestive segments from videos

[0006] Manually labeling suggestive segments in videos and then filtering them using the labeled information is a time-consuming and labor-intensive method.

[0007] III. Filtering through the inducible fragment itself

[0008] Using general-purpose suggestive segment recognition tools to directly filter videos; however, these tools do not refer to normal video content during algorithm design and model training, which may lead to misidentification of normal video content, thus affecting the efficiency and accuracy of filtering.

[0009] Therefore, the present invention proposes a method, system, device and medium for detecting suggestive content in images and videos, in order to solve the related problems existing in the prior art mentioned above. Summary of the Invention

[0010] The purpose of this invention is to provide a method, system, device, and medium for detecting suggestive content in images and videos, in order to solve the related problems existing in the prior art.

[0011] The above-mentioned technical objective of the present invention is achieved through the following technical solution: a method for detecting suggestive content in images and videos, the suggestive detection method comprising the following steps:

[0012] Obtain the complete video file to be processed;

[0013] The complete video file is segmented to obtain at least one video slice. The video slice is then transformed using a preset analysis and transformation model to obtain the first video feature vector corresponding to each video slice.

[0014] Construct the slice segmentation line between every two adjacent video slices;

[0015] Compare the first video feature vectors of at least one video slice on both sides of the slice segmentation line to obtain at least one first similarity value. Average the at least one first similarity value to obtain a second similarity value. Use the magnitude of the second similarity value to determine whether the video slices on both sides of the segmentation line are normal video slices or induced video slices.

[0016] The video slices are transformed using a preset analysis and transformation model, and a corresponding first video style vector is constructed for each video slice;

[0017] All the first video style vectors are aggregated into two style categories, and the two style categories are respectively used to construct the corresponding first style category vectors;

[0018] Determine whether two first style category vectors are similar to obtain a third similarity value. Use the magnitude of the third similarity value to determine whether the two style categories contain suggestive video slices.

[0019] The present invention is further configured such that: the step of segmenting the complete video file to obtain at least one video slice, and analyzing and transforming the video slice to obtain the first video feature vector corresponding to each video slice, specifically includes the following steps:

[0020] The complete video file was modeled using a 3D-CNN network;

[0021] The complete video file is divided according to the preset time interval to obtain at least one video slice;

[0022] The video slices are transformed using a preset analysis and transformation model to obtain the first video feature vector corresponding to each video slice, and the first video feature vector represents the content of the video slice.

[0023] The present invention is further configured as follows: comparing the first video feature vectors of at least one video slice on both sides of the slice segmentation line to obtain at least one first similarity value, averaging the at least one first similarity value to obtain a second similarity value, and using the magnitude of the second similarity value to determine whether the video slices on both sides of the segmentation line are normal video slices or inducement video slices, specifically including the following steps:

[0024] The sum of the first video feature vectors of the N adjacent video slices on both sides of the slice separator line is compared, where N is a positive integer, to obtain a first similarity value A. n , n = 1, 2, 3 ... n, the sum of the first video feature vectors is performed using a multilayer perceptron;

[0025] The obtained multiple first similarity values ​​A n The average value is then used to obtain the second similarity value.

[0026] If the second similarity value is large, then the two sides of the slice dividing line are a normal video slice and a induced video slice, respectively; if the second similarity value is small, then the two sides of the slice dividing line are either a normal video slice or an induced video slice.

[0027] The present invention is further configured such that: the first video style vector represents the style of the video slice, including video style, image quality and color tone.

[0028] The present invention is further configured such that the construction steps of the analysis and transformation model are as follows:

[0029] Obtain a complete video file with multiple known content and style features;

[0030] The content and style features of the complete video file are labeled to construct the standard training set.

[0031] The initial analysis and transformation model is constructed using a deep learning-based convolutional neural network.

[0032] The analysis and transformation model is trained using the standard training set to obtain the trained analysis and transformation model.

[0033] The present invention also provides a system for detecting suggestive content in images and videos, the suggestive detection system comprising:

[0034] Video file acquisition module: used to acquire the complete video file to be processed;

[0035] Video feature vector construction module: used to segment the complete video file to obtain at least one video slice, and use a preset analysis and transformation model to transform the video slice to obtain the first video feature vector corresponding to each video slice;

[0036] Segmentation Marker Construction Module: Used to construct the segmentation markers between every two adjacent video slices;

[0037] First judgment module: used to compare the first video feature vectors of at least one video slice on both sides of the slice segmentation line to obtain at least one first similarity value, average the at least one first similarity value to obtain a second similarity value, and use the magnitude of the second similarity value to determine whether the video slices on both sides of the segmentation line are normal video slices or inducement video slices.

[0038] Video style vector construction module: used to transform the video slices using a preset analysis and transformation model, and construct a corresponding first video style vector for each video slice;

[0039] The style category vector construction module is used to aggregate all the first video style vectors into two style categories, and the two style categories respectively construct the corresponding first style category vectors;

[0040] The second judgment module is used to determine whether the two first style category vectors are similar, obtain a third similarity value, and use the magnitude of the third similarity value to determine whether the two style categories contain misleading video slices.

[0041] The present invention also provides a computer device, the computer device including a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform the steps of the above-described method for detecting suggestive content in images and videos.

[0042] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the above-described method for detecting suggestive content in images and videos.

[0043] In summary, the present invention has the following beneficial effects: In the present invention, the style of video content can be analyzed, embedded suggestive segments can be identified from normal videos, and suggestive segments within the video can be filtered out. This solves the problems of traditional methods being time-consuming and labor-intensive, as well as misidentifying parts of normal video content, which affects the efficiency and accuracy of filtering. Attached Figure Description

[0044] Figure 1 This is a flowchart of a method for detecting suggestive content in images and videos according to an embodiment of the present invention;

[0045] Figure 2 This is a schematic diagram of similarity comparison in the image and video inducement content detection method in an embodiment of the present invention. Detailed Implementation

[0046] The following is in conjunction with the appendix Figure 1-2 The present invention will be described in further detail below.

[0047] Example 1

[0048] Methods for detecting suggestive content in images and videos, such as Figure 1 As shown, the inducibility detection method includes the following steps:

[0049] S1: Obtain the complete video file to be processed;

[0050] S2: Segment the complete video file to obtain at least one video slice. Then, use a pre-defined analysis and transformation model to transform the video slice, obtaining the first video feature vector corresponding to each video slice. Specifically, this includes the following steps:

[0051] A 3D-CNN network is used to model the complete video file;

[0052] The complete video file is divided according to a preset time interval to obtain at least one video slice. It should be noted that the preset time interval in this embodiment is 3 seconds.

[0053] The video slices are transformed using a pre-defined analysis and transformation model to obtain the first video feature vector corresponding to each video slice. The first video feature vector represents the content of the video slice.

[0054] S3: Construct slice division lines between every two adjacent video slices;

[0055] S4: Compare the first video feature vectors of at least one video slice on both sides of the slice segmentation line to obtain at least one first similarity value. Average the at least one first similarity value to obtain a second similarity value. Use the magnitude of the second similarity value to determine whether the video slices on both sides of the segmentation line are normal video slices or induced video slices. Specifically, this includes the following steps:

[0056] The sum of the first video feature vectors of the N adjacent video slices on both sides of the slice separator line is compared, where N is a positive integer, to obtain a first similarity value A. n n = 1, 2, 3...n, the sum of the first video feature vectors is obtained by multilayer perceptron summation;

[0057] The obtained multiple first similarity values ​​A n The average value is then used to obtain the second similarity value.

[0058] like Figure 2 As shown, this embodiment uses a third-order similarity comparison, as detailed below:

[0059] Compare one video slice 100 on each side of a slice segmentation line 200 to obtain a first similarity value A1;

[0060] By comparing two adjacent video slices 100 on the left and right sides of the same slice segmentation line 200, a first similarity value A2 is obtained;

[0061] By comparing the three adjacent video slices 100 on both sides of the same slice segmentation line 200, a first similarity value A3 is obtained;

[0062] A1, A2, and A3 are summed using a multiple perceptron to obtain the second similarity value.

[0063] If the second similarity value is large, then the slice dividing line is divided into a normal video slice and a induced video slice on both sides; if the second similarity value is small, then the slice dividing line is divided into a normal video slice or an induced video slice on both sides.

[0064] S5: Use a preset analysis and transformation model to transform the video slices and construct a corresponding first video style vector for each video slice; the first video style vector represents the style of the video slice, including video style, image quality and color tone.

[0065] S6: Aggregate all first video style vectors into two style categories, and construct corresponding first style category vectors for each of the two style categories;

[0066] S7: Determine whether two first style category vectors are similar to obtain a third similarity value. Use the magnitude of the third similarity value to determine whether the two style categories contain suggestive video slices.

[0067] The steps for constructing the analysis and transformation model are as follows:

[0068] Obtain a complete video file with multiple known content and style features;

[0069] The content and style features of the complete video file are labeled to construct a standard training set.

[0070] An initial analysis and transformation model was constructed using a deep learning-based convolutional neural network.

[0071] The analysis and transformation model is trained using a standard training set to obtain the trained analysis and transformation model.

[0072] Example 2

[0073] Based on the same concept as in Embodiment 1, the present invention also provides a system for detecting suggestive content in images and videos, the suggestive detection system comprising:

[0074] Video file acquisition module: used to acquire the complete video file to be processed;

[0075] Video feature vector construction module: used to segment the complete video file to obtain at least one video slice, and use a preset analysis and transformation model to transform the video slice to obtain the first video feature vector corresponding to each video slice;

[0076] Segmentation Marker Construction Module: Used to construct segmentation marks between every two adjacent video slices;

[0077] First judgment module: used to compare the first video feature vectors of at least one video slice on both sides of the slice segmentation line to obtain at least one first similarity value, average the at least one first similarity value to obtain a second similarity value, and use the magnitude of the second similarity value to determine whether the video slices on both sides of the segmentation line are normal video slices or induced video slices.

[0078] Video style vector construction module: used to transform video slices using a preset analysis and transformation model, and construct a corresponding first video style vector for each video slice;

[0079] The style category vector construction module is used to aggregate all first video style vectors into two style categories, and the two style categories respectively construct the corresponding first style category vectors;

[0080] The second judgment module is used to determine whether two first style category vectors are similar, obtain a third similarity value, and use the magnitude of the third similarity value to determine whether the two style categories contain induced video slices.

[0081] The present invention also provides a computer device, the computer device including a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform the steps of the above-described method for detecting suggestive content in images and videos.

[0082] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the above-described method for detecting suggestive content in images and videos.

[0083] This specific embodiment is merely an explanation of the present invention and is not intended to limit the invention. After reading this specification, those skilled in the art can make modifications to this embodiment without contributing any inventive step, but such modifications are protected by patent law as long as they are within the scope of the claims of the present invention.

Claims

1. A method for detecting suggestive content in images and videos, characterized in that, The inducibility detection method includes the following steps: Obtain the complete video file to be processed; The complete video file is segmented to obtain at least one video slice. The video slice is then transformed using a preset analysis and transformation model to obtain a first video feature vector corresponding to each video slice. Construct the slice segmentation line between every two adjacent video slices; Compare the first video feature vectors of at least one video slice on both sides of the slice segmentation line to obtain at least one first similarity value. Average the at least one first similarity value to obtain a second similarity value. Use the magnitude of the second similarity value to determine whether the video slices on both sides of the segmentation line are normal video slices or induced video slices. The video slices are transformed using a preset analysis and transformation model, and a corresponding first video style vector is constructed for each video slice; All the first video style vectors are aggregated into two style categories, and the two style categories are used to construct corresponding first style category vectors. Determine whether two first style category vectors are similar to obtain a third similarity value. Use the magnitude of the third similarity value to determine whether the two style categories contain suggestive video slices. The process of comparing the first video feature vectors of at least one video slice on both sides of the slice segmentation line to obtain at least one first similarity value, averaging the at least one first similarity value to obtain a second similarity value, and using the magnitude of the second similarity value to determine whether the video slices on both sides of the segmentation line are normal video slices or misleading video slices, specifically includes the following steps: The sum of the first video feature vectors of the N adjacent video slices on both sides of the slice separator line is compared, where N is a positive integer, to obtain a first similarity value A. n n = 1, 2, 3...n, and the sum of the first video feature vectors is performed using a multilayer perceptron. The obtained multiple first similarity values ​​A n The average value is then used to obtain the second similarity value. If the second similarity value is large, then the two sides of the slice dividing line are a normal video slice and a induced video slice, respectively; if the second similarity value is small, then the two sides of the slice dividing line are either a normal video slice or an induced video slice.

2. The method for detecting suggestive content in images and videos according to claim 1, characterized in that, The process of segmenting the complete video file to obtain at least one video slice, and then analyzing and transforming the video slices to obtain the first video feature vector corresponding to each video slice, specifically includes the following steps: The complete video file was modeled using a 3D-CNN network; The complete video file is divided according to a preset time interval to obtain at least one video slice; The video slices are transformed using a preset analysis and transformation model to obtain the first video feature vector corresponding to each video slice, and the first video feature vector represents the content of the video slice.

3. The method for detecting suggestive content in images and videos according to claim 1, characterized in that, The first video style vector represents the style of the video slice, including video style, image quality, and color tone.

4. The method for detecting suggestive content in images and videos according to claim 2, characterized in that, The steps for constructing the analysis and transformation model are as follows: Obtain a complete video file with multiple known content and style features; The content and style features of the complete video file are labeled to construct a standard training set. The initial analysis and transformation model is constructed using a deep learning-based convolutional neural network. The analysis and transformation model is trained using the standard training set to obtain the trained analysis and transformation model.

5. A detection system for use in the image and video suggestive content detection method according to any one of claims 1-4, characterized in that, Inductive detection systems include: Video file acquisition module: used to acquire the complete video file to be processed; Video feature vector construction module: used to segment the complete video file to obtain at least one video slice, and use a preset analysis and transformation model to transform the video slice to obtain the first video feature vector corresponding to each video slice; Segmentation Marker Construction Module: Used to construct the segmentation markers between every two adjacent video slices; First judgment module: used to compare the first video feature vectors of at least one video slice on both sides of the slice segmentation mark to obtain at least one first similarity value, average the at least one first similarity value to obtain a second similarity value, and use the magnitude of the second similarity value to determine whether the video slices on both sides of the segmentation mark are normal video slices or induced video slices. Video style vector construction module: used to transform the video slices using a preset analysis and transformation model, and construct a corresponding first video style vector for each video slice; The style category vector construction module is used to aggregate all the first video style vectors into two style categories, and the two style categories respectively construct the corresponding first style category vectors; The second judgment module is used to determine whether the two first style category vectors are similar, obtain a third similarity value, and use the magnitude of the third similarity value to determine whether the two style categories contain misleading video slices.

6. A computer device, characterized in that, The computer device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the method for detecting suggestive content in images and videos as described in any one of claims 1-4.

7. A computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method for detecting suggestive content in images and videos as described in any one of claims 1-4.

Images