WO2021164515A1

WO2021164515A1 - Detection method and apparatus for tampered image

Info

Publication number: WO2021164515A1
Application number: PCT/CN2021/074001
Authority: WO
Inventors: 许庆堂; 周继恩; 陆堃彪; 张青清; 陈磊; 何运田
Original assignee: 中国银联股份有限公司
Priority date: 2020-02-17
Filing date: 2021-01-27
Publication date: 2021-08-26
Also published as: TWI759099B; TW202133038A; CN111325265B; CN111325265A

Abstract

Provided are a detection method and apparatus for a tampered image. The method comprises: determining a first feature description of a first image and a first image type to which the first image belongs, wherein the first image is an image to be reviewed; from feature descriptions of reviewed images in the first image type, determining a second feature description of which the feature similarity with the first feature description meets a first similarity requirement; for the second feature description of which the feature similarity meets the first similarity requirement, determining a second image corresponding to the second feature description; and if the image similarity between the first image and the second image meets a second similarity requirement, determining that the first image is a tampered image. By using the method, a tampered image can be accurately and rapidly detected on the basis of the global features of the image.

Description

Method and device for detecting tampered images

Cross-references to related applications

This application claims the priority of a Chinese patent application filed with the Chinese Patent Office on February 17, 2020, the application number is 202010097367.5, and the application name is "A method and device for detecting tampered images", the entire content of which is incorporated by reference In this application.

Technical field

This application relates to the field of artificial intelligence technology, and in particular to a method and device for detecting tampered images.

Background technique

At present, digital imaging equipment has become an indispensable item in modern life. With the rapid development of various editing software, it is becoming more and more convenient for people to modify and edit digital images. This will undoubtedly lead to incorrect image tampering. It is easy to mislead people when they consider important information, which can cause adverse effects on society. Therefore, image detection forensics technology is one of the current research hotspots.

Image tampering methods mainly include image mosaic tampering, image copy-paste tampering, and image object removal tampering. The existing image tampering detection methods mainly use the target detection model corresponding to the target detection method to extract the differentiated features existing at the boundary of the image tampering area and the non-tampering area. For example, the storefront image containing the owner's image is changed to another People separate the tampered area (the image of the shop owner) and the non-tampered area (the image of the storefront) based on the difference in the graphics angle, color, pixel, brightness, etc. generated at the junction of the image splicing, so as to determine whether the image is a tampered image. However, the difference caused by image splicing can be eliminated by re-shooting. For example, if the spliced image is re-shot, some cameras can perform secondary processing on the splicing boundary. The difference of the splicing boundary is eliminated. In this way, the target detection method is invalid and the difference of the image splicing boundary cannot be obtained.

Therefore, there is an urgent need for a method and device for detecting tampered images, which are used to accurately and quickly detect tampered images based on the global characteristics of the image.

Summary of the invention

The embodiments of the present invention provide a detection method and device for tampered images, which are used to accurately and quickly detect tampered images based on the global characteristics of the image.

In the first aspect, an embodiment of the present invention provides a method for detecting tampered images, the method including:

Determine the first feature description of the first image and the first image type to which the first image belongs; the first image is the image to be reviewed; from the feature description of the reviewed image in the first image type, determine the The feature similarity of the first feature description meets the second feature description of the first similarity requirement; for the second feature description whose feature similarity meets the first similarity requirement, determine the second image corresponding to the second feature description; If the image similarity between the first image and the second image meets the second similarity requirement, it is determined that the first image is a tampered image.

Using the above method, the first image type of the first image is used to confirm the audited image corresponding to the first image type, so as to obtain the feature description corresponding to the audited image. In this way, the feature description of the reviewed image of the same image type as the first image type can be obtained first. Then determine the feature similarity between the first description feature of the first image and the feature description of the reviewed image, so as to determine the second feature description that meets the first similarity requirement in the feature similarity, and confirm the second image corresponding to the second feature description . In this way, the second image with greater similarity to the first image can be preliminarily determined, that is, the reviewed image with greater similarity to the image to be reviewed can be preliminarily determined; these determined reviewed images can be basically judged to be suspected of being reviewed. The image to be tampered with. Further, the image similarity between the second image and the first image is confirmed. If the image similarity meets the second similarity requirement, the first image is an image tampered with based on the second image. That is to say, these determined audited images, that is, images that are suspected of being tampered with by the image to be audited, are further calculated for image similarity with the image to be audited. If there is an image similarity that meets the second similarity requirement, it shall meet the requirements of the second degree of similarity. The image that is suspected of being tampered with by the image to be audited corresponding to the image similarity required by the two similarity is confirmed as an image that has been tampered with by the image to be audited, that is, the first image is a tampered image. In this way, it is possible to accurately and quickly detect tampered images based on the global features of the image.

In a possible design, determining the image similarity between the first image and the second image in the following manner includes: determining a third feature description of the first image and a fourth feature description of the second image Feature description; wherein the dimensions of the third feature description and the fourth feature description are the same; the dimensions of the first feature description and the second feature description are the same; the number of dimensions of the third feature description is more than The number of dimensions of the first feature description; the similarity between the third feature description and the fourth feature description is calculated, so as to obtain the image similarity between the first image and the second image.

Using the above method, by determining the third feature description of the first image and the fourth feature description of the second image, the image similarity between the third feature description and the fourth feature description is calculated, because the third feature description and the fourth feature description The dimension of the feature description is the same, and the dimension is higher than the dimension of the first feature description and the second feature description, so the accuracy of the calculated image similarity is higher than the feature similarity. Therefore, the feature similarity between the image to be reviewed and the image to be reviewed can be calculated first through the low-dimensional feature description, and the reviewed image that is relatively similar to the image to be reviewed can be screened out to achieve preliminary screening, which can speed up the calculation. Further calculate the image similarity between the second image corresponding to the feature similarity that meets the first similarity and the first image, that is, calculate the similarity between the image to be reviewed and the reviewed image through the high-dimensional feature description, so as to make the accuracy of the image similarity Greatly improve. Realize the image-based global features, accurately and quickly detect tampered images.

In a possible design, determining the image similarity between the first image and the second image includes: determining the image of the first image and the second image through a two-channel image similarity calculation model Similarity; wherein, the dual-channel image similarity calculation model is obtained through training of historical sample pairs; each historical sample pair includes two images with similarity marks.

Using the above method, the dual-channel image similarity calculation model obtained through historical sample pair training, because each historical sample pair includes two images with similarity marks, these two images can be used as positive samples and negative samples to train dual-channel Image similarity calculation model. In this way, when the image similarity between the first image and the second image is calculated by the dual-channel image similarity calculation model, the effect of tampering operations such as image rotation, scale conversion, and image stitching on detection is effectively avoided, and the accuracy of image similarity is increased. Rate.

In a possible design, determining the second feature description whose similarity with the first feature description meets the requirements of the first similarity includes: determining the feature described with the first feature by a nearest neighbor search method The similarity meets the second feature description required by the first similarity.

Using the above method and the nearest neighbor search method, it is possible to accurately and quickly determine the second feature description whose similarity with the first feature description meets the requirements of the first similarity.

In a possible design, the description of each second feature in each image type is determined by the following methods, including: classifying the reviewed images to obtain the reviewed images in each image type; Feature extraction is performed on each audited image in the image types, the extracted features are reduced in dimensionality based on the product quantization method to obtain the feature description of the audited image, and the feature index of the feature description of the audited image is determined.

Using the above method, the audited images are classified, the features are extracted corresponding to the audited images in each image type, and the dimensionality reduction processing is performed on the extracted features based on the product quantization method to obtain the second feature description. In this way, the dimension of the feature description of each reviewed image and the calculation of the first feature description is lower, and the determination of the second feature description is accelerated. And by setting the feature index of the second feature description, the feature description of the reviewed image can be quickly obtained when calculating the feature similarity, and the speed of detecting tampered images can be accelerated.

In a possible design, performing feature extraction on each reviewed image in the image type includes: performing feature extraction on each reviewed image in the image type through a set feature extraction model; determining The third feature description of the first image and the fourth feature description of the second image include: performing feature extraction on the first image through the feature extraction model to obtain the third feature description; The feature extraction model performs feature extraction on the second image to obtain the fourth feature description.

Using the above method, the third feature description of the first image is extracted by setting the feature extraction model, and the fourth feature description of the second image is extracted. The high-dimensional third feature description and the fourth feature description can be extracted through the feature extraction model to realize the rapid and high-precision detection of tampered images based on the global features of the image.

In a second aspect, an embodiment of the present invention provides a device for detecting tampered images, the device including:

The determining module is used to determine the first feature description of the first image and the first image type to which the first image belongs; the first image is an image to be reviewed;

The calculation module is used to determine, from the feature description of the reviewed image in the first image type, a second feature description whose feature similarity with the first feature description meets the first similarity requirement; A second feature description with similarity requirements, determine the second image corresponding to the second feature description; if the image similarity between the first image and the second image meets the second similarity requirements, determine the The first image is a tampered image.

In a possible design, determining the image similarity between the first image and the second image in the following manner includes:

The determining module is further configured to determine a third feature description of the first image and a fourth feature description of the second image; wherein the dimensions of the third feature description and the fourth feature description are the same; The dimensions of the first feature description and the second feature description are the same; the number of dimensions of the third feature description is more than the number of dimensions of the first feature description;

The calculation module is specifically configured to calculate the similarity between the third feature description and the fourth feature description, so as to obtain the image similarity between the first image and the second image.

In a third aspect, an embodiment of the present application further provides a computing device, including: a memory, configured to store program instructions; a processor, configured to call the program instructions stored in the memory, and execute according to the obtained program as in the first aspect The methods described in the various possible designs.

In a fourth aspect, embodiments of the present application also provide a computer-readable non-volatile storage medium, including computer-readable instructions. When the computer reads and executes the computer-readable instructions, the computer executes the same as in the first aspect. The methods described in the various possible designs.

These implementation manners or other implementation manners of the present application will be more concise and understandable in the description of the following embodiments.

Description of the drawings

In order to explain the technical solutions in the embodiments of the present invention more clearly, the following will briefly introduce the drawings needed in the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained from these drawings without creative labor.

FIG. 1 is a schematic diagram of an architecture for detecting tampered images according to an embodiment of the present invention;

2 is a schematic flowchart of a method for detecting tampered images according to an embodiment of the present invention;

FIG. 3 is a schematic flowchart of a method for detecting tampered images according to an embodiment of the present invention;

FIG. 4 is a schematic diagram of a detection device for tampering images provided by an embodiment of the present invention.

Detailed ways

In order to make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative work shall fall within the protection scope of the present invention.

The embodiment of the present invention provides a schematic diagram of the architecture of a method for detecting tampered images, as shown in FIG. The feature library system 101 sets a retrieval index for the feature description of the reviewed image corresponding to each image type. After the image to be reviewed is input into the image detection system 102, the image detection system 102 determines the image type of the image to be reviewed, and performs feature extraction on the image to be reviewed, that is, extracts the first feature description; the image detection system 102 obtains the image type of the image to be reviewed Then, the retrieval index in the reviewed image feature library system 101 is called to obtain the feature description of the reviewed image of the same image type as the image type of the image to be reviewed; the image detection system 102 calculates the first feature description and the feature of the reviewed image The feature similarity of the description, the feature description whose feature similarity meets the requirements of the first similarity is the second feature description, and the audited image corresponding to the second feature description is obtained from the audited image feature database system 101 according to the second feature description; image detection The system 102 calculates the image similarity between the image to be reviewed and the image that has been reviewed. The tampering evaluation system 103 obtains the image similarity between the image to be reviewed and the image that has been reviewed, evaluates whether the image similarity meets the second similarity requirement, and if it does, determines that the image to be reviewed is a tampered image. Otherwise, the image to be reviewed is a non-tampered image.

Based on this, the embodiment of the present application provides a method for detecting tampered images, as shown in FIG. 2, including:

Step 201: Determine the first feature description of the first image and the first image type to which the first image belongs; the first image is an image to be reviewed;

Here, the first image is the image to be reviewed. For example, to receive the preferential rights and interests, you need to upload the business license and/or storefront image, and you need to upload the preferential rights and interests by uploading the falsified business license and/or the corresponding storefront image multiple times. The business license and/or storefront image are the images to be reviewed. The image types can be pharmacy storefront images, supermarket storefront images, gas station storefront images, ID photo images, and so on. The first feature description is the feature description data of the image to be reviewed; it can be in the form of a vector, for example, a feature vector is formed by the color, brightness, line angle, etc. of each point; it can also be a feature set, for example, the pixel set includes each The feature set formed by the pixel data of the point and the brightness set including the brightness of each point, etc.; the specific feature description method is not limited here. In other words, determine the image type and image feature description of the image to be reviewed.

Step 202: From the feature description of the reviewed image in the first image type, determine a second feature description whose feature similarity with the first feature description meets the first similarity requirement;

Here, the feature similarity is the similarity between the feature description of each reviewed image and the first feature description. The feature description of the reviewed image corresponds to the first feature description, and is a feature description that can calculate the similarity between the feature description of the reviewed image and the feature description of the first feature description through a corresponding algorithm. For example, the first feature description is the feature vector of the image to be reviewed, the feature description of the reviewed image is the feature vector of the reviewed image, and the cosine similarity between the feature vector of the image to be reviewed and the feature vector of the reviewed image is calculated as the feature similarity. ; Or calculate the inner product of the feature vector of the image to be audited and the feature vector of the audited image as the feature similarity; or calculate the Euclidean distance between the feature vector of the image to be audited and the feature vector of the audited image as the feature similarity, etc. Etc.; the feature similarity algorithm is not specifically limited. For another example, orthogonally decompose the feature vector of the first feature description and/or the feature vector of the reviewed image, and perform quantization on the decomposed low-dimensional orthogonal subspace to obtain a PQ (Product quantization, product quantization) code , Calculate the feature similarity through the nearest neighbor algorithm. In addition, the low-dimensional space is encoded with a smaller codebook, which can also reduce the data storage space. Therefore, the specific existence form of the feature description is not limited. The first similarity is a set threshold of feature similarity or a probability distribution of feature similarity. If the feature similarity is a value of 0-1, the threshold of feature similarity can be set to 0.8 according to industry experience. If the feature similarity exceeds 0.8, the audited image corresponding to the feature similarity is determined to be an image that is suspected of being tampered with; Among them, greater than the threshold 0.8 is the first similarity requirement; another example, feature similarity is the set of similarity of each feature, if the probability distribution of the set of similarity of each feature is the same as the set probability distribution, it is determined The reviewed image corresponding to the feature similarity is an image that is suspected of being tampered with; where the same probability distribution is set as the first similarity requirement; therefore, the specific existence form of the first similarity is not limited. That is to say, the feature description of the reviewed image is acquired according to the first image type, and the feature similarity between the first feature description and the feature description of each reviewed image is calculated. If the feature similarity meets the requirement of the first similarity, the second feature description corresponding to the feature similarity that meets the first similarity requirement is obtained.

Wherein, determining the second feature description whose feature similarity with the first feature description meets the first similarity requirement includes: determining that the feature similarity with the first feature description meets the first similarity through a nearest neighbor search method The second feature description of the degree requirement. That is to say, the feature similarity between the first feature description and the feature description of the reviewed image can be calculated by the nearest neighbor search algorithm, and then the second feature description corresponding to the feature similarity that meets the first similarity requirement can be determined.

Before calculating the feature similarity, the description of each second feature in each image type can be determined by the following methods, including: classifying the reviewed images to obtain the reviewed images in each image type; Feature extraction is performed on each audited image in the image types, the extracted features are reduced in dimensionality based on the product quantization method to obtain the feature description of the audited image, and the feature index of the feature description of the audited image is determined. In other words, before calculating the feature similarity, classify the reviewed images, for example, divide the images into ID photos, storefront images, etc. The ID photos can be divided into ID cards, business licenses, driving photos, etc., storefront images It can be divided into supermarket storefront images, gas station storefront images, and so on. After obtaining the audited images in each image type, extract the features of the audited images in each image type, and perform dimensionality reduction processing on the extracted features of each audited image by multiplication and quantification, to obtain the features of the audited image describe.

Wherein, performing feature extraction on each audited image in the image type includes: performing feature extraction on each audited image in the image type through a set feature extraction model; determining the value of the first image The third feature description and the fourth feature description of the second image include: performing feature extraction on the first image through the feature extraction model to obtain the third feature description; Perform feature extraction on the second image to obtain the fourth feature description.

Here, feature extraction is performed on each of the reviewed images in the image types, through a set feature extraction model, for example, for more than 10,000 merchant door head photos. Label different types of images, such as supermarket images marked as "1", gas station images marked as "2", training on the basis of the VGG16 network model; the fully connected layer feature parameters of the model are obtained, which can be regarded as audited The fourth feature description of the image; the image to be audited is trained on the basis of the VGG16 network model; the fully connected layer feature parameters of the model are obtained, which can be used as the third feature description of the image to be audited; the middle feature layer is subsequently taken as the 512-dimensional feature output, That is, the dimensionality reduction process is performed on the third feature description to obtain the first feature description; the dimensionality reduction process is performed on the fourth feature description to obtain the feature description of the reviewed image. Or, directly take the value of each feature of each point of the audited image to form a high-dimensional vector, that is, the fourth feature description; directly take the value of each feature of each point of the image to be audited to form a high-dimensional vector, That is the third feature description. The manner of determining the third feature description and the fourth feature description here is not limited.

Step 203: For a second feature description whose feature similarity meets the first similarity requirement, determine a second image corresponding to the second feature description;

Here, for the feature description of the reviewed image whose feature similarity meets the first similarity requirement, the feature description of the reviewed image that meets the first similarity requirement is the second feature description, and the second feature description is determined based on the second feature description The corresponding reviewed image.

Step 204: If the image similarity between the first image and the second image meets the second similarity requirement, determine that the first image is a tampered image.

Here, the image similarity is the similarity whose accuracy is higher than the feature similarity. The image similarity is obtained by calculating the similarity between the image to be audited and the audited image that meets the first similarity requirement through the corresponding algorithm. For example, the feature vector of the image to be reviewed and the feature vector of the reviewed image that meets the first similarity requirement is calculated to calculate the cosine similarity, or the feature vector of the image to be reviewed and the audited image that meets the first similarity requirement Calculate the inner product of the feature vector to calculate the image similarity; or calculate the Euclidean distance between the feature vector of the image to be reviewed and the feature vector of the reviewed image that meets the requirements of the first similarity; do image similarity, etc., image similarity The algorithm is not specifically limited. The second similarity is the set threshold of image similarity or the probability distribution of image similarity. If the image similarity is a value of 0-1, the threshold of image similarity can be set to 0.8 according to industry experience. If the image similarity exceeds 0.8, the audited image corresponding to the image similarity is determined to be an image that has been tampered with; Among them, greater than the threshold 0.8 is the second similarity requirement; for another example, the image similarity is the set of similarity of each feature. If the probability distribution of the set of similarity of each feature is the same as the set probability distribution, it is determined The audited image corresponding to the image similarity is an image that has been tampered with; where the same probability distribution is set as the second similarity requirement; therefore, the specific existence form of the second similarity is not limited. That is, if the image similarity between the image to be reviewed and the image that has been reviewed meets the second similarity requirement, it is determined that the image to be reviewed is a tampered image.

Wherein, determining the image similarity between the first image and the second image in the following manner includes: determining a third feature description of the first image and a fourth feature description of the second image; wherein The third feature description and the fourth feature description have the same dimensions; the first feature description and the second feature description have the same dimensions; the third feature description has more dimensions than the first feature description The number of dimensions; the similarity between the third feature description and the fourth feature description is calculated to obtain the image similarity between the first image and the second image.

Here, the third feature description of the image to be reviewed and the fourth description feature of the reviewed image can be obtained through the VGG16 network model, or directly by determining the value of each feature of each point. The third feature description and the fourth feature description stored in step 203 can be directly called, or the third feature description and the fourth feature description can be recalculated. Here, the method of determining the third feature description and the fourth feature description is not limited. The above-mentioned third feature description and the fourth feature description have the same dimensions, the first feature description and the second feature description have the same dimensions, and the third feature description has a higher dimension than the first feature description. That is to say, the feature description dimension on which the calculation of image similarity is based is higher than the feature description dimension on which the calculation of feature similarity is based.

Wherein, determining the image similarity between the first image and the second image includes: determining the image similarity between the first image and the second image through a two-channel image similarity calculation model; The two-channel image similarity calculation model is obtained by training on historical sample pairs; each historical sample pair includes two images with similarity marks. In other words, historical sample pairs, that is, historical sample pairs containing positive sample images and negative sample images, can be used to train a two-channel image similarity calculation model, so that the audited images and images to be audited that meet the requirements of the first similarity are input The two-channel image similarity calculation model can obtain more accurate image similarity. Here you can also directly calculate the corresponding third descriptive feature of the image to be reviewed and the fourth descriptive feature of the reviewed image, such as the inner product of high-dimensional feature vectors, cosine similarity, etc. The specific calculation method of image similarity is not limited .

The embodiment of the present application also provides a flow of a method for detecting tampered images, as shown in FIG. 3, including:

Step 301: Create a database of reviewed images.

Step 302: Classify the reviewed images in the reviewed image database.

Step 303: Extract the features of each reviewed image in each type of reviewed image for each type of reviewed image, and determine the fourth feature description of each reviewed image.

Step 304: Perform dimensionality reduction processing on the fourth feature description of each reviewed image to obtain the feature description of each reviewed image. Through Faiss (a framework for efficient similarity search and clustering for dense vectors), HASH (a function that compresses messages of any length to a fixed-length message summary) and other methods, the characteristics of each type of audited image Describe setting search index.

Step 305: Obtain an image to be reviewed;

Step 306: Determine the image type and the first feature description of the image to be reviewed;

Step 307: According to the image type of the image to be reviewed, the feature description of the reviewed image of the same image type is determined through the index.

Step 308: Calculate the feature similarity between the first feature description of the image to be reviewed and the feature description of the reviewed image.

Step 309: Determine the feature similarity that meets the first similarity requirement, determine the second feature description corresponding to the feature similarity that meets the first similarity requirement, and obtain the reviewed image corresponding to the second feature description according to the second feature description.

Step 310: Determine the fourth feature description of the reviewed image corresponding to the second feature description, and determine the third feature description of the image to be reviewed.

Step 311: Calculate image similarity through the third feature description and the fourth feature description.

Step 312: Determine that there is an image similarity that meets the second similarity requirement.

Step 313: Determine that the image to be reviewed is a tampered image.

Based on the same concept, an embodiment of the present invention provides a detection device for a tampered image. FIG. 4 is a schematic diagram of a detection device for a tampered image provided by an embodiment of the application, as shown in FIG. 4, including:

The determining module 401 is configured to determine the first feature description of the first image and the first image type to which the first image belongs; the first image is an image to be reviewed;

The calculation module 402 is used to determine, from the feature description of the reviewed image in the first image type, a second feature description whose feature similarity with the first feature description meets the first similarity requirement; The second feature description required by the first similarity is determined, and the second image corresponding to the second feature description is determined; if the image similarity between the first image and the second image meets the second similarity requirement, the second image is determined The first image is a tampered image.

In a possible design, determining the image similarity between the first image and the second image in the following manner includes: the determining module 401 is further configured to: determine the third feature description of the first image And the fourth feature description of the second image; wherein the dimensions of the third feature description and the fourth feature description are the same; the dimensions of the first feature description and the second feature description are the same; The number of dimensions of the third feature description is more than the number of dimensions of the first feature description; the calculation module 402 is specifically configured to: calculate the similarity between the third feature description and the fourth feature description, so as to obtain the The image similarity between the first image and the second image.

In a possible design, the calculation module 402 is specifically configured to: determine the image similarity between the first image and the second image through a dual-channel image similarity calculation model; wherein, the dual-channel image The similarity calculation model is obtained by training on historical sample pairs; each historical sample pair includes two images with similarity marks.

The calculation module 402 is specifically configured to determine a second feature description whose feature similarity with the first feature description meets the first similarity requirement through a nearest neighbor search method.

The determining module 401 is specifically configured to: classify the reviewed images to obtain the reviewed images in each image type; for each image type, perform feature extraction on each of the reviewed images in the image type, based on the product The quantification method performs dimensionality reduction processing on the extracted features to obtain the feature description of the reviewed image, and determines the feature index of the feature description of the reviewed image.

The calculation module 402 is further configured to: perform feature extraction on each of the reviewed images in the image type through the set feature extraction model; determine the third feature description of the first image and the description of the second image The fourth feature description includes: performing feature extraction on the first image through the feature extraction model to obtain the third feature description; performing feature extraction on the second image through the feature extraction model to obtain the The fourth feature description.

Those skilled in the art should understand that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, this application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, this application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

This application is described with reference to flowcharts and/or block diagrams of methods, equipment (systems), and computer program products according to this application. It should be understood that each process and/or block in the flowchart and/or block diagram, and the combination of processes and/or blocks in the flowchart and/or block diagram can be realized by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing equipment to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing equipment can be used to generate It is a device that realizes the functions specified in one process or multiple processes in the flowchart and/or one block or multiple blocks in the block diagram.

These computer program instructions can also be stored in a computer-readable memory that can guide a computer or other programmable data processing equipment to work in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture including the instruction device. The device implements the functions specified in one process or multiple processes in the flowchart and/or one block or multiple blocks in the block diagram.

These computer program instructions can also be loaded on a computer or other programmable data processing equipment, so that a series of operation steps are executed on the computer or other programmable equipment to produce computer-implemented processing, so as to execute on the computer or other programmable equipment. The instructions provide steps for implementing the functions specified in one process or multiple processes in the flowchart and/or one block or multiple blocks in the block diagram.

Obviously, those skilled in the art can make various changes and modifications to the application without departing from the spirit and scope of the application. In this way, if these modifications and variations of this application fall within the scope of the claims of this application and their equivalent technologies, then this application is also intended to include these modifications and variations.

Claims

A method for detecting tampered images, characterized in that, the method includes:

Determine the first feature description of the first image and the first image type to which the first image belongs; the first image is an image to be reviewed;

From the feature description of the reviewed image in the first image type, determine the second feature description whose feature similarity with the first feature description meets the first similarity requirement; for those whose feature similarity meets the first similarity requirement The second feature description determines the second image corresponding to the second feature description; if the image similarity between the first image and the second image meets the second similarity requirement, then the first image is determined to be tampered with image.
The method according to claim 1, wherein determining the image similarity between the first image and the second image in the following manner comprises:

Determine the third feature description of the first image and the fourth feature description of the second image; wherein the dimensions of the third feature description and the fourth feature description are the same; the first feature description and the fourth feature description The dimensions of the second feature description are the same; the number of dimensions of the third feature description is more than the number of dimensions of the first feature description;

The similarity between the third feature description and the fourth feature description is calculated, so as to obtain the image similarity between the first image and the second image.
The method according to claim 1, wherein determining the image similarity between the first image and the second image comprises:

Determine the image similarity between the first image and the second image through a dual-channel image similarity calculation model; wherein, the dual-channel image similarity calculation model is obtained by training on historical sample pairs; each historical sample The alignment includes two images with similarity marks.
The method according to any one of claims 1 to 3, wherein determining a second feature description whose similarity with the first feature description meets the first similarity requirement comprises:

Through the nearest neighbor search method, the second feature description whose similarity with the first feature description meets the first similarity requirement is determined.
The method of claim 4, wherein:

Determine each second feature description in each image type by the following methods, including:

Categorize the reviewed images to obtain the reviewed images in each image type;

For each image type, perform feature extraction on each reviewed image in the image type, perform dimensionality reduction processing on the extracted features based on the product quantization method to obtain the feature description of the reviewed image, and determine the characteristics of the reviewed image The feature index of the feature description.
The method according to claim 5, characterized in that, performing feature extraction on each reviewed image in the image type comprises:

Perform feature extraction on each reviewed image in the image type through the set feature extraction model;

Determining the third feature description of the first image and the fourth feature description of the second image includes:

Performing feature extraction on the first image by using the feature extraction model to obtain the third feature description;

Perform feature extraction on the second image through the feature extraction model to obtain the fourth feature description.
A detection device for tampering images, characterized in that the device comprises:

The determining module is used to determine the first feature description of the first image and the first image type to which the first image belongs; the first image is an image to be reviewed;

The calculation module is used to determine, from the feature description of the reviewed image in the first image type, a second feature description whose feature similarity to the first feature description meets the first similarity requirement; A second feature description with similarity requirements, determine the second image corresponding to the second feature description; if the image similarity between the first image and the second image meets the second similarity requirements, determine the The first image is a tampered image.
8. The device according to claim 7, wherein determining the image similarity between the first image and the second image in the following manner comprises:

The determining module is also used to determine the third feature description of the first image and the fourth feature description of the second image; wherein the dimensions of the third feature description and the fourth feature description are the same; The dimensions of the first feature description and the second feature description are the same; the number of dimensions of the third feature description is more than the number of dimensions of the first feature description;

The calculation module is specifically configured to calculate the similarity between the third feature description and the fourth feature description, so as to obtain the image similarity between the first image and the second image.
A computing device, characterized in that it comprises:

Memory, used to store program instructions;

The processor is configured to call the program instructions stored in the memory, and execute the method according to any one of claims 1 to 6 according to the obtained program.
A computer-readable non-volatile storage medium, characterized by comprising computer-readable instructions, when the computer reads and executes the computer-readable instructions, the computer is caused to execute any one of claims 1 to 6 Methods.