KR20110073238A - Apparatus and method for blocking the objectionable multimedia based on multimodal and multiscale features - Google Patents

Abstract

PURPOSE: According to the kind and category of a video, a harmful video blocking apparatus and method based on multimodal and multiscale features are provided to perform multi-stage harmfulness filtering suitable for each scale. CONSTITUTION: A multiscale feature analysis unit(110) creates multiscale features. A harmfulness classification model generator(120) creates multilevel harmfulness classification models. A harmfulness decision unit(130) decides the harmfulness of a video with comparison between at least one of the multiscale feature of the input video and at least one of the multilevel harmfulness classification model. In case the video is harmful, a harmful video blocking unit(140) secludes the input video.

Description

Apparatus and method for blocking the objectionable multimedia based on multimodal and multiscale features}

The present invention relates to a multimodal and multiscale feature based harmful image blocking device and method, and more particularly, from the known harmful and harmless image learning data, color, texture, shape, skin color, face, edge, MPEG-7 descriptor, Analyze and characterize multimodal information such as objects, object meanings, and object relationships on a multi-scale basis, and use them to create multilevel hazard classification models with multiple levels of complexity for hazard classification and generate hazard classification models. The present invention relates to a method and apparatus for blocking harmful images by judging the harmfulness of newly introduced images.

The Internet is not only rich in information, so called as a sea of information, but also easy to use, making it part of the daily lives of many people in modern times, providing positive aspects in social, economic, and academic terms. However, in contrast to this positive aspect, the reckless dissemination of harmful information by using the Internet's openness, interconnection and anonymity has become a serious social problem. In particular, teenagers who can access the Internet at any time are exposed to harmful information more frequently than before. This environment can attract emotionally and mentally harming young people who have poor value judgment and weak self-control. Therefore, there is a need for a method of blocking harmful information so that the socially disadvantaged youth or unwanted people are not exposed to harmful information.

Conventionally, harmful image blocking methods include metadata and text information based blocking techniques, hash and DB based blocking techniques, and content based blocking techniques. The blocking technique based on metadata and text information is a method of determining the harmfulness of an image by analyzing the harmfulness of the text included in the title, file name, and description of the image, and has a high over-blocking rate and an error-blocking rate. Hash and DB based blocking method calculates database of hash value of known harmful image and calculates hash value of image and compares it with previously constructed database in order to judge the harmfulness of new image which is input. This is a method of determining the harmfulness of the video. This method not only increases the size of the hash value database as the harmful image increases, but also increases the amount of computation for determining the harmfulness of the image. Also, even known harmful images may not be blocked if the hash value changes through slight modification.

Recently, the content-based blocking technique has been analyzed to extract a feature by analyzing the content of the harmful image, generate a hazard classification model from the feature, and determine the harmfulness of the input image based on the generated hazard classification model. This method solves the high over-blocking rate and error-blocking rate of metadata and text-based blocking, and solves the database size and computational problems of hash and DB-based blocking.

However, most content-based blocking techniques use low-level features such as color, texture, and shape as the features of harmful images, or MPEG-7 descriptors, which are mainly used for image retrieval. use. This information does not properly reflect the characteristics of the harmful image, which results in low blocking rate and high error rate. In order to solve this problem, a recent new approach detects skin color in pixels and utilizes skin and non-skin color ratios in the image as a characteristic of hazard determination. However, the approach using these features is also insufficient to semantically accurately describe and abbreviate the actual harmful images, so that the hazard classification model generated using these features still shows low performance. In addition, the complexity of generating harmful features of an image is equally applied to all images, and it takes a long time to generate a high level of harmful features, and blocks harmful images by processing images having different complexity in the same manner. It leads to a decrease in the overall performance of the system.

Therefore, in order to reduce the overblocking rate and error rate of the harmful image blocking method, and to improve the processing performance and speed, the multi-scale unit that can apply the harmful classification model suitable for the complexity level of the image can be applied. There is a need for a harmful image blocking method with multi-level harmful image filtering.

The technical problem to be achieved by the present invention is characterized by analyzing multi-modal information such as color, texture, shape, skin color, face, edge, MPEG-7 descriptor, object, meaning from multi-scale unit from image learning data, Provides an apparatus and method for generating harmful classification models with multi-level complexity through machine learning and judging the harmfulness of newly input images using the generated multi-level harmful classification model to block harmful images. There is.

Multimodal and multiscale feature-based harmful image blocking device according to an aspect of the present invention, multi-scale feature analysis unit for generating multi-scale harmful and harmless features by analyzing the multi-modal information extracted from the image learning data; A hazard classification model generator for generating a multilevel hazard classification model by statistically processing the generated multiscale harmful and harmless features and performing machine learning; Analyze multi-modal information extracted from input image data input for hazard determination, extract at least one of the multiscale features of the input image, and compare the extracted feature with at least one of the multilevel hazard classification model. Hazard determination unit for determining the harmfulness of; And a harmful image blocking unit which blocks the input image when the determination result is harmful.

Multimodal and multiscale feature-based harmful image blocking method according to another aspect of the present invention comprises the steps of analyzing the multi-modal information extracted from the image learning data to generate a multi-scale harmful and harmless features; Statistically processing the generated multiscale harmful and harmless features and applying machine learning to generate a multilevel hazard classification model; Analyzing at least one of the multi-scale features of the input image by analyzing multi-modal information of the input image data input for hazard determination; multi-leveling the at least one multi-scale feature extracted from the input image data Determining a hazard of the input image by comparing with at least one of a hazard classification model; And blocking the input image when the determination result is harmful.

According to the multi-modal and multi-scale feature based harmful image blocking device and method of the present invention, multi-scale feature extraction and multi-level harmful classification model generation are possible by utilizing the multi-modal information included in the image to determine the harmfulness of the image. It was. As a result, by performing multi-level harmful filtering for each scale according to the type and category of the image, not only can the over-blocking rate and the error-blocking rate caused by the harmful image blocking be lowered, but also the processing performance of the harmful image blocking is improved. Reduce costs In addition, the apparatus and method of the harmful image blocking apparatus of the present invention can be applied to the multi-level of the harmful classification model of the multi-level, it is possible to freely control the depth of the image analysis and the complexity of the harmful image blocking according to the application environment.

1 is a block diagram showing the configuration of a multimodal and multiscale feature based harmful image blocking device according to an embodiment of the present invention;
FIG. 2A is a block diagram illustrating a detailed configuration of the multiscale feature analyzer 110 illustrated in FIG. 1, and FIGS. 2B to 2D are coarse grain unit feature analyzer 1110 of FIG. 2A and middle grain unit feature analysis. A block diagram showing a detailed configuration of each of the unit 1120 and the fine grain unit feature analyzer 1130;
3 is a block diagram showing the detailed configuration of the hazard classification model generation unit 120 shown in FIG.
4 is a block diagram showing the detailed configuration of the hazard determination unit 130 shown in FIG.
5 is a flowchart illustrating a method for blocking harmful images based on multimodal and multiscale features according to an embodiment of the present invention.

DETAILED DESCRIPTION Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those skilled in the art may easily implement the present invention. The present invention may, however, be embodied in many different forms and should not be construed as limited to the embodiments set forth herein. In the drawings, parts irrelevant to the description are omitted in order to clearly describe the present invention, and like reference numerals designate like parts throughout the specification.

Throughout the specification, when a part is said to "include" a certain component, it means that it can further include other components, without excluding other components unless specifically stated otherwise. In addition, the terms “… unit”, “module”, and the like described in the specification mean a unit that processes at least one function or operation, which means that the unit may be implemented by hardware or software or a combination of hardware and software.

1 is a block diagram illustrating a configuration of an apparatus for blocking harmful images based on multimodal and multiscale features according to an embodiment of the present invention. As shown in FIG. 1, the multimodal and multiscale feature-based harmful image blocking device 100 includes a multiscale feature analyzer 110, a harmful classification model generator 120, a hazard determination unit 130, and a harmful object. And an image blocker 140. Image learning data consists of harmful images and harmless images, and is used as data to model the harmfulness of images.

The multiscale feature analyzer 110 extracts multimodal information including color, texture, shape, skin color, face, edge, MPEG-7 descriptor, object, and meaning from the image learning data, and uses the extracted multimodal information. Produces multiscale harmful and harmless features.

The hazard classification model generation unit 120 generates a multilevel hazard classification model by statistically processing multi-scale presence or absence features generated by the multiscale feature analysis unit 110 and performing machine learning. In one embodiment, the multilevel hazard classification model includes a low level, a medium level, and a high level hazard classification model, and is used as a reference model for determining the harmfulness of a later input image.

The hazard determination unit 130 analyzes the multi-modal information extracted from the input image data input for hazard determination, extracts the multiscale feature, and at least one of the multilevel hazard models generated by the hazard classification model generator 120. The harmfulness of the image is determined by comparing with.

The harmful image blocking unit 140 blocks the image determined to be harmful.

2A is a block diagram illustrating a detailed configuration of the multiscale feature analyzer 110 illustrated in FIG. 1. Referring to FIG. 2A, the multiscale feature analyzer 110 includes a coarse-grained granularity feature analyzer 1110, a medium grain unit feature analyzer 1120, and a fine-grained granularity. Including the feature analysis unit 1130, to generate the presence or absence of harmful features in the multi-scale unit, and provides it to the harmful classification model generation unit 120.

In one embodiment, the coarse grain unit feature analyzer 1110 generates a complexity based feature by analyzing color complexity, texture complexity, and shape complexity of the image training data.

The intermediate grain unit feature analyzer 1120 generates a single modal based low level feature by analyzing skin color, face, edge information, and MPEG-7 descriptor included in the image learning data. The single modal based low level feature refers to a feature generated based on color, texture, and shape information, and is referred to as a low level because the generated feature does not include information such as correlation between meaning and information.

The fine grain unit feature analysis unit 1130 detects the object from the image learning data, performs harmful semantic analysis and object relation analysis of the object to generate a multimodal based high level feature.

2B to 2D are block diagrams illustrating detailed structures of the coarse grain unit feature analyzer 1110, the intermediate grain unit feature analyzer 1120, and the fine grain unit feature analyzer 1130 of FIG. 2A.

Referring to FIG. 2B, the coarse grain unit feature analyzer 1110 may include a color complexity analyzer 1111 for analyzing color complexity of image training data, and a texture complexity analyzer 1112 for analyzing texture complexity of image training data. , A shape complexity analyzer 1113 for analyzing shape complexity of image learning data, and a complexity-based feature generator 1114 for generating a complexity-based feature according to the type and category of an image based on the analyzed color, texture, and shape complexity. ). In one embodiment, the complexity is evaluated by analyzing the type and distribution of color, the type and distribution of texture, the number of edges constituting the shape, the number and distribution of regions, and the like.

Referring to FIG. 2C, the intermediate grain unit feature analyzer 1120 may include a skin color detector 1121 for detecting skin color information from image training data, a face detector 1122 for detecting face information from image training data, and image training data. An edge detector 1123 for detecting edge information from the image, an MPEG-7 descriptor extractor 1124 for extracting an MPEG-7 descriptor from image learning data, and the skin color, face, edge information, and MPEG-7 descriptor for analyzing the image. And a single modal based low level feature generator 1125 for generating a single modal based low level feature according to type and category.

Referring to FIG. 2D, the fine grain unit feature analyzer 1130 may include an object detector 1131 which detects object information from image learning data, and harmful meanings of the detected object (chest exposure, pussy exposure, sexual activity, masturbation). Object meaning analysis unit 1132 for analyzing whether or not to include, etc., an object relationship analysis unit 1133 for analyzing the relationship between the detected objects (a body part such as face, chest, pussy, buttocks and human body) and the analysis And a multimodal based high level feature generation unit 1134 for generating a multimodal based high level feature according to the type and category of the image based on the object meaning and the relationship between the objects. Object detection is performed by a method widely used in the field of image processing, and object relationship analysis is performed using the position, size, number information of the detected object.

3 is a block diagram showing the detailed configuration of the hazard classification model generation unit 120 shown in FIG. Referring to FIG. 3, the harmful classification model generation unit 120 may perform statistics on color complexity, texture complexity, and shape complexity features generated by the coarse grain unit feature analysis unit 1110 of the multiscale feature analysis unit 110. Skin color detection information and face generated by the low level hazard classification model generator 1210 and the intermediate grain unit feature analyzer 1120 of the multiscale feature analyzer 110 that generate a low level hazard classification model through processing and machine learning. A medium level hazard classification model generator 1220 and a multiscale feature analyzer 110 that generate a medium level classification model through statistical processing and machine learning on features of detection information, edge detection information, and MPEG-7 descriptor information. High-level existence through statistical processing and machine learning on the features of the object detection information, semantic analysis information, and object relationship analysis information generated by the fine grain unit feature analysis unit 1130 And a high level hazard classification model generation unit 1230 for generating a solution classification model.

In an alternative embodiment, the hazard classification model generator 1210 not only generates the above-described low, medium, and high level hazard classification models, but also combines the hazard classification models of each level in series or in parallel to form a hazard classification model. You can also create

4 is a block diagram showing a detailed configuration of the hazard determination unit 130 shown in FIG. Referring to FIG. 4, the hazard determiner 130 may include a coarse grain unit feature extractor 1310, an intermediate grain unit feature extractor 1320, a fine grain unit feature extractor 1330, and an image hazard determiner 1340. It includes. The coarse grain unit feature extractor 1310 extracts a complexity-based feature of the input image data by analyzing color, texture, and shape complexity features of the image data input as a hazard determination target. The intermediate grain unit feature extractor 1320 extracts a single modal based low level feature of the input image data by analyzing at least one of skin color, face, edge information, and MPECG-7 descriptor included in the input image data. The fine grain unit feature extractor 1330 extracts a multi-modal based high level feature by detecting an object from the input image data and analyzing the meaning of the detected object and the relationship between the objects.

The coarse grain unit feature extractor 1310, the intermediate grain unit feature extractor 1320, and the fine grain unit feature extractor 1330 are coarse grain units included in the multiscale feature analyzer 110 illustrated in FIG. 2A. It may operate in the same or similar manner as the feature analyzer 1110, the intermediate grain unit feature analyzer 1120, and the fine grain unit feature analyzer 1130.

In one embodiment, the coarse grain unit feature extractor 1310, the intermediate grain unit feature extractor 1320, and the fine grain unit feature extractor 1330 of the hazard determination unit 130 may be configured according to the type and category of the input image data. Some or all of them may be selected and operated, and the characteristics of the input image generated from the selected extractor may be compared with at least one of the low, medium, and high level harmful classification models generated by the harmful classification model generator 120. By determining the harmfulness of the image.

5 is a flowchart illustrating a method for blocking harmful images based on multimodal and multiscale features according to an embodiment of the present invention. Referring to FIG. 5, first, multimodal information including color, texture, shape, skin color, face, edge, MPEG-7 descriptor, object, and object meaning are extracted from image learning data of which harmlessness is known in advance. Generate multi-scale harmful and harmless features (S510). In an embodiment, the multi-scale harmful and harmless feature generation step S510 may include generating a complexity-based feature by analyzing color complexity, texture complexity, and shape complexity of the image training data, skin color, and face included in the image training data. Generating a single modal based low level feature by analyzing edge information and MPEG-7 descriptors, and generating a multimodal based high level feature by detecting an object from image learning data and performing harmful semantic analysis and object relation analysis of the object. Steps.

Next, a multilevel hazard classification model including a low level, a medium level, and a high level hazardous classification model is generated according to the harmful features generated in step S510 (S520). Specifically, the multilevel hazard classification model generation step S520 may include generating a low level hazard classification model using a complexity-based feature; Generating a medium hazard classification model using a single modal based low level feature; And generating a high level hazard classification model using the multimodal based high level features. The multilevel hazard classification model is generated as a learning result through statistical processing and machine learning on multiscale harmful features extracted in step S500.

Next, at least one multiscale feature is extracted from the input image data input to determine whether the object is harmful (S530). In one example, the multiscale feature includes a complexity based feature, a single modal based low level feature, and a multimodal based high level feature, and at least one of these multiscale features will be extracted according to the type and category of the input image data.

Next, the harmfulness of the image is determined by comparing the at least one multiscale feature extracted in step S530 with at least one of the multilevel harmful classification models generated in step S520 (S540).

If it is determined in step S540 that the image is harmful, the corresponding image is blocked (S550).

A feature of the present invention is to analyze and characterize multimodal information such as color, texture, shape, skin color, face, edge, MPEG-7 descriptor, object, semantic analysis in multiscale units from image learning data, and utilize these features. Multi-level harmful classification model is generated through machine learning, and the harmful image is blocked by judging the harmfulness of the newly input image using the generated harmful classification model. By using the multi-level harmful image filtering based on the characteristics of the multi-scale unit using the multi-modal information as described above, the over-blocking rate and the error-blocking rate that occur in the harmful image blocking are significantly lowered, and the processing performance and speed are improved.

Embodiments of the invention described above may be implemented in any of a variety of ways. For example, embodiments may be implemented using hardware, software, or a combination thereof. If implemented in software, it may be implemented as software running on one or more processors using various operating systems or platforms. In addition, such software may be written using any of a number of suitable programming languages, and may also be compiled into executable machine code or intermediate code executing on a framework or virtual machine.

The invention also relates to a computer readable medium (e.g., computer memory, one or more) having recorded thereon one or more programs that perform a method of implementing the various embodiments of the invention discussed above when executed on one or more computers or other processors. Or a floppy disk, a compact disk, an optical disk, a magnetic tape, a flash memory, or the like.

The present invention can also be embodied as computer-readable codes on a computer-readable recording medium. The computer-readable recording medium includes all kinds of recording devices in which data that can be read by a computer system is stored. Examples of computer-readable recording media include ROM, RAM, CD_ROM, magnetic tape, floppy disks, and optical data storage, and may also include those implemented in the form of carrier waves (e.g., transmission over the Internet). . The computer readable recording medium can also be distributed over network coupled computer systems so that the computer readable code is stored and executed in a distributed fashion.

Multimodal and multiscale feature-based harmful image blocking device and method of the present invention can be applied to portable multimedia playback devices (MP3, PMP, etc.), mobile phones and PDAs.

So far I looked at the center of the preferred embodiment for the present invention. Those skilled in the art will appreciate that the present invention can be implemented in a modified form without departing from the essential features of the present invention. Therefore, the disclosed embodiments should be considered in descriptive sense only and not for purposes of limitation. The scope of the present invention is shown in the claims rather than the foregoing description, and all differences within the scope will be construed as being included in the present invention.

Claims (14)

A multiscale feature analysis unit configured to generate multiscale harmful and harmless features by analyzing multimodal information extracted from image learning data;
A hazard classification model generator for generating a multilevel hazard classification model by statistically processing the generated multiscale harmful and harmless features and performing machine learning;
Analyze multi-modal information extracted from input image data input for hazard determination, extract at least one of the multiscale features of the input image, and compare the extracted feature with at least one of the multilevel hazard classification model. Hazard determination unit for determining the harmfulness of; And
A harmful image blocking unit which blocks the input image when the determination result is harmful
Multimodal and multiscale feature-based harmful image blocking device comprising a.
The method of claim 1, wherein the multiscale feature analysis unit,
A coarse grain unit feature analyzer configured to generate a complexity-based feature by analyzing color complexity, texture complexity, and shape complexity of the image learning data;
An intermediate grain unit feature analyzer configured to generate a single modal based low level feature by analyzing skin color, face, edge information, and MPEG-7 descriptor included in the image learning data; And
Fine grain unit feature analysis unit for generating multi-modal based high-level features by detecting an object from the image learning data, and performing harmful meaning analysis and object relation analysis of the object
Multimodal and multiscale feature-based harmful image blocking device comprising a.
The method of claim 2, wherein the coarse grain unit feature analysis unit,
A color complexity analyzer for analyzing color complexity of the image learning data;
A texture complexity analyzer for analyzing texture complexity of the image learning data;
A shape complexity analyzer for analyzing shape complexity of the image learning data;
A complexity-based feature extractor for extracting a complexity-based feature according to the type and category of an image by referring to the analyzed color, texture, and shape complexity.
Multimodal and multiscale feature-based harmful image blocking device comprising a.
The method of claim 2, wherein the intermediate grain unit characteristic analysis unit,
A skin color detector for detecting skin color information from the image learning data;
A face detector detecting face information from the image learning data;
An edge detector detecting edge information from the image learning data;
An MPEG-7 descriptor extractor for extracting an MPEG-7 descriptor from the image learning data; And
A single modal based low level feature extractor which analyzes the skin color, face, edge information and MPEG-7 descriptor to extract a single modal based low level feature according to the type and category of the image.
Multimodal and multiscale feature-based harmful image blocking device comprising a.
The method of claim 2, wherein the fine grain unit feature analysis unit,
An object detector detecting object information from the image learning data;
An object semantic analyzer for analyzing harmful meanings of the detected objects;
An object relationship analyzer analyzing the relationship between the detected objects; And
A multi-modal based high level feature extraction unit for extracting multi-modal based high level features according to the type and category of the image based on the harmful meaning of the analyzed object and the relationship information between the objects.
Multimodal and multiscale feature-based harmful image blocking device comprising a.
According to claim 2, The harmful classification model generation unit,
A low level hazard classification model generator for generating a low level hazard classification model using the complexity-based feature generated by the coarse grain unit feature analyzer;
A medium hazard classification model generator for generating a medium hazard classification model using the single modal based low level features generated by the intermediate grain unit feature generator; And
A high level hazard classification model generator for generating a high level hazard classification model using the multimodal based high level features generated by the fine grain unit feature analysis unit.
Multimodal and multiscale feature-based harmful image blocking device comprising a.
According to claim 1, wherein the hazard determination unit,
A coarse grain unit feature extraction unit for extracting a complexity-based feature by analyzing color complexity, texture complexity, and shape complexity of the input image data;
An intermediate grain unit feature extraction unit for extracting a single modal based low level feature by analyzing skin color, face, edge information, and MPEG-7 descriptor included in the input image data;
A fine grain unit feature extraction unit for extracting a multimodal based high level feature by detecting an object included in the input image data and performing harmful semantic analysis and object relationship analysis of the object; And
Comparing the at least one multiscale feature extracted from at least one of the coarse grain unit feature extractor, the intermediate grain unit feature extractor, and the fine grain unit feature extractor with at least one of the multilevel hazard classification model Image Hazard Determination Unit
Multimodal and multiscale feature-based harmful image blocking device comprising a.
The method of claim 7, wherein some or all of the coarse grain unit feature extractor, the intermediate grain unit feature extractor, and the fine grain unit feature extractor are selected according to the type and category of the input image data. A multimodal and multiscale feature based harmful image blocking device for selectively extracting at least one of multiscale features.
The method of claim 7, wherein the multiscale feature extraction and hazard determination unit selects at least one of the low level hazard classification model, the medium level hazard classification model, and the high level hazard classification model according to the type and category of the input image data. Multimodal and multiscale feature based noisy image blocker to compare the characteristics of input image data.
Analyzing the multi-modal information extracted from the image training data to generate multiscale harmful and harmless features;
Statistically processing the generated multiscale harmful and harmless features and applying machine learning to generate a multilevel hazard classification model;
Extracting at least one of the multi-scale features of the input image by analyzing multimodal information of input image data inputted for harmfulness determination;
Determining a hazard of the input image by comparing the at least one multiscale feature extracted from the input image data with at least one of the multilevel hazard classification models; And
Blocking the input image when the determination result is harmful
Multimodal and multiscale feature-based harmful image blocking method comprising a.
The method of claim 10, wherein generating the multiscale harmful and harmless features comprises:
Generating a complexity-based feature by analyzing color complexity, texture complexity, and shape complexity of the image learning data;
Generating a single modal based low level feature by analyzing skin color, face, edge information and MPEG-7 descriptor included in the image learning data; And
Detecting an object from the image learning data, performing harmful meaning analysis and object relation analysis of the object to generate a multimodal based high level feature
Multimodal and multiscale feature-based harmful image blocking method comprising a.
12. The method of claim 11, wherein statistically processing the harmful and harmless features of the generated multiscale unit and applying machine learning to generate a multilevel hazard classification model,
Generating a low level hazard classification model using the complexity based feature;
Generating a midlevel harmful classification model using the single modal based low level feature; And
Generating a high level hazard classification model using the multimodal based high level features
Multimodal and multiscale feature-based harmful image blocking method comprising a.
The method of claim 10, wherein the extracting at least one of the multiscale features of the input image comprises:
Analyzing color complexity, texture complexity, and shape complexity of the input image data and extracting a complexity based feature based on the analyzed complexity;
Extracting skin color information, face information, edge information, and MPEG-7 descriptor information from the input image data, and extracting a single modal based low level feature based on the extracted information; And
Analyzing object information, semantic information, and relationship information between the objects of the input image data and extracting at least one multiscale feature by performing at least one of extracting a multimodal based high level feature based on the analysis result And multiscale feature based harmful image blocking method.
The method of claim 10, wherein the extracting at least one of the multiscale features of the input image comprises at least one of a complexity-based feature, a single-modal-based low-level feature, and a multi-modal-based high-level feature according to the type and category of the input image. Multimodal and multiscale feature based harmful image blocking method for extracting one.

Images