CN112417381B - Method and device for rapidly positioning infringement image applied to image copyright protection - Google Patents
Method and device for rapidly positioning infringement image applied to image copyright protection Download PDFInfo
- Publication number
- CN112417381B CN112417381B CN202011446862.9A CN202011446862A CN112417381B CN 112417381 B CN112417381 B CN 112417381B CN 202011446862 A CN202011446862 A CN 202011446862A CN 112417381 B CN112417381 B CN 112417381B
- Authority
- CN
- China
- Prior art keywords
- image
- retrieved
- data
- infringement
- retrieval
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 78
- 239000013598 vector Substances 0.000 claims abstract description 157
- 238000013136 deep learning model Methods 0.000 claims abstract description 33
- 238000001514 detection method Methods 0.000 claims abstract description 18
- 239000011159 matrix material Substances 0.000 claims description 24
- 238000012549 training Methods 0.000 claims description 22
- 238000007781 pre-processing Methods 0.000 claims description 13
- 230000008447 perception Effects 0.000 claims description 11
- 238000004422 calculation algorithm Methods 0.000 claims description 8
- 238000013527 convolutional neural network Methods 0.000 claims description 8
- 230000006870 function Effects 0.000 claims description 8
- 238000012163 sequencing technique Methods 0.000 claims description 7
- 230000008859 change Effects 0.000 claims description 4
- 238000004590 computer program Methods 0.000 claims description 4
- 238000011176 pooling Methods 0.000 claims description 4
- 230000002708 enhancing effect Effects 0.000 abstract 2
- 239000000284 extract Substances 0.000 abstract 1
- 238000004364 calculation method Methods 0.000 description 9
- 230000008569 process Effects 0.000 description 5
- 238000012216 screening Methods 0.000 description 4
- 238000005516 engineering process Methods 0.000 description 3
- 230000002776 aggregation Effects 0.000 description 2
- 238000004220 aggregation Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000011161 development Methods 0.000 description 1
- 239000012634 fragment Substances 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F21/00—Security arrangements for protecting computers, components thereof, programs or data against unauthorised activity
- G06F21/10—Protecting distributed programs or content, e.g. vending or licensing of copyrighted material ; Digital rights management [DRM]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/53—Querying
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/55—Clustering; Classification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Software Systems (AREA)
- Databases & Information Systems (AREA)
- Biophysics (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Technology Law (AREA)
- Computer Hardware Design (AREA)
- Computer Security & Cryptography (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
- Image Analysis (AREA)
Abstract
The invention relates to the field of image copyright protection, and discloses a method and a device for quickly positioning an infringement image for image copyright protection, wherein the method comprises the steps of respectively obtaining copyright image data and an image to be retrieved, and extracting a characteristic vector by using an image infringement detection deep learning model; calculating a perceptual hash value of each image; and establishing a large-scale data retrieval method by using the characteristic vector and the perceptual hash value, and performing positioning infringement retrieval on the image data to be retrieved by the large-scale data retrieval method to obtain an infringement retrieval result. The invention extracts the high-dimensional characteristic vector of the image by combining the CNN network and the R-MAC method, simultaneously calculates the perceptual hash value of the image, and identifies and compares the image by two layers of scales, thereby enhancing the identification of anti-attack means which possibly appears in an infringement image and enhancing the robustness of the model.
Description
Technical Field
The invention relates to the field of image copyright protection, in particular to a method and a device for quickly positioning an infringement image applied to image copyright protection.
Background
With the rapid development of multimedia and internet technologies, massive image data has been widely applied to various fields of society, and the problem of image copyright protection is becoming more severe. Generally, an image retrieval technology is used to search and compare an infringement image and an original image, which mainly includes two steps: how to express the characteristic information of the image and an efficient retrieval method. The difference in the way image content is described can be divided into two categories: text-based image retrieval and content-based image retrieval. The text-based image retrieval method mainly utilizes text annotation to describe the content in the image, but with the rapid increase of data magnitude, the manpower and financial resources consumed by manually annotating the text are not enough to support the huge data volume in reality. In recent decades, image retrieval technologies based on content have been rapidly developed, and in the early century, due to the excellent performance of SIFT features in the problem of image scale change, image retrieval methods based on local description operators have been widely researched, most local feature description methods encode features based on gradients to overcome semantic differences, such as a bag-of-words model, a Fisher vector or a vector local aggregation descriptor, and the like, but the encoding methods cannot capture high-level semantic information of images. In recent years, high-dimensional image features extracted by the image representation method based on the deep learning convolutional neural network show extremely high performance, and the method for extracting the features by image retrieval gradually changes from SIFT features to CNN features. With the continuous increase of data scale, the calculation cost for training the deep learning model is also continuously increased, so that the comparison and identification of images under large-scale data and the quick search and sequencing of infringement results need to be met by more reasonably utilizing high-dimensional image features and a more efficient retrieval method.
Disclosure of Invention
The invention provides a method and a device for rapidly positioning an infringement image for image copyright protection, thereby solving the problems in the prior art.
In a first aspect, the present invention provides a method for rapidly positioning an infringement image applied to image copyright protection, comprising the following steps:
s1) respectively acquiring copyright image data and an image to be retrieved, establishing a picture infringement detection deep learning model, and extracting a feature vector of each image in the copyright image data and the image to be retrieved by using the picture infringement detection deep learning model;
s2) calculating the perception hash value of each image in the copyright image data and the image to be retrieved;
s3) establishing a large-scale data retrieval method by utilizing the characteristic vector and the perception hash value, and rapidly positioning and infringing the retrieval of the image data to be retrieved by the large-scale data retrieval method to obtain an infringing retrieval result.
Further, in step S1), acquiring the copyright image data and the image to be retrieved, respectively, establishing a picture infringement detection deep learning model, and extracting a feature vector of each image in the copyright image data and the image to be retrieved by using the picture infringement detection deep learning model, including the following steps:
s11) carrying out first image preprocessing on the copyright image data and each image in the images to be retrieved, wherein the first image preprocessing comprises the step of scaling the copyright image data and each image in the images to be retrieved to a first preset size;
s12), establishing a CNN-based deep learning model, inputting each image preprocessed by a first image into the CNN-based deep learning model, and taking a feature map of the last convolutional layer of the CNN-based deep learning model as output;
s13) extracting a plurality of areas with different scales from each feature map by adopting an R-MAC method to obtain a plurality of area R-MAC features, summing and pooling the plurality of area R-MAC features, and splicing to obtain n-dimensional feature vectors corresponding to each image; the n-dimensional feature vector corresponding to each image includes the n-dimensional feature vector of each image in the copyrighted-image data and the n-dimensional feature vector of the image to be retrieved.
Further, in step S2), calculating a perceptual hash value of each image in the copyrighted-image data and the image to be retrieved, including the following steps:
s21) carrying out second image preprocessing on the copyright image data and each image in the images to be retrieved, wherein the second image preprocessing comprises the step of scaling the copyright image data and each image in the images to be retrieved to a second preset size;
s22) converting each image scaled to the second preset size into a gray image, respectively;
s23) calculating the discrete cosine transform of each image after being converted into the gray level image to obtain a discrete cosine transform coefficient matrix;
s24) extracting a low-frequency matrix with a preset size at the upper left corner of the discrete cosine transform coefficient matrix, calculating the element average value of the low-frequency matrix, setting elements which are larger than or equal to the element average value in the low-frequency matrix to be 1, setting elements which are smaller than the element average value in the low-frequency matrix to be 0, and obtaining the low-frequency matrix with the elements set to be 1 or 0;
s25) flattening the low-frequency matrix with the element set to be 1 or 0 into a one-dimensional vector to obtain copyright image data and a perception hash value of each image in the image to be retrieved.
Further, in step S3), the method for building a large-scale data retrieval by using the feature vector and the perceptual hash value includes constructing a retrieval index file and retrieving by using the retrieval index file, where constructing the retrieval index file includes the following steps:
s31) initializing a retrieval index file data structure, wherein the retrieval index file data structure comprises a reverse arrangement table, a code table, a reverse arrangement vector ID table and/or a reverse arrangement vector coding table;
s32) training data are obtained, and a clustering algorithm is trained by using the training data; the training data comprises a plurality of data points, and the data points are respectively n-dimensional characteristic vectors of each image in the copyright image data; establishing a reverse vector ID table, wherein the reverse vector ID table is used for storing a plurality of data points and IDs of the data points;
s33) determining the number of the centers of the clusters and the number range of elements in the cluster where each cluster center is located according to the data volume of the training data;
s34) randomly initializing all cluster centers and updating all cluster centers and code tables, comprising the following steps:
s3031) initializing all clustering centers;
s3032) calculating any one clustering center CqShortest distance d (C) to other cluster centersq，Cw)；d(Cq，Cw) Represents the clustering center CqAnd a distance clustering center CqNearest cluster center CwThe distance of (d);
s3033) obtaining a clustering center CqThe data point x in the cluster to the cluster center CqDistance d (C)qX), judgment 2d (C)q，x)≤d(Cq，Cw) If yes, the classification position of the data point x is unchanged; if not, entering step S3034);
s3034) calculating the distance from the data point x to other clustering centers, and classifying the data point x into the cluster where the clustering center closest to the other clustering centers is located;
s3035) repeating the steps S3032) to S3034), and sequentially obtaining the clustering center CqA classification location for each data point in the cluster;
s3036) repeating the steps S3032) to S3035), and sequentially obtaining the classification position of each data point in the cluster where each cluster center is located;
s3037) updating all the clustering centers, judging whether all the clustering centers change, and if so, returning to the step S3032); if not, finishing the cluster updating, obtaining all the updated cluster centers and all the data points in the cluster of each cluster center, wherein each data point corresponds to one ID, adding all the updated cluster centers into a code table, storing all the data points in the cluster of each cluster center and the IDs of the data points into corresponding inverted arrangement tables, each cluster center corresponds to one inverted arrangement table, inverted IDs and inverted code tables are stored in the inverted arrangement tables, the inverted IDs are used for storing the IDs of the data points, and the inverted code tables are used for storing all the data points in the cluster of the cluster center.
Further, the retrieval by using the retrieval index file comprises the following steps:
s35), acquiring the constructed retrieval index file, calculating the vector distance between the N-dimensional feature vector of the image to be retrieved and all the updated clustering centers, and acquiring N clustering center points closest to the N-dimensional feature vector of the image to be retrieved;
s36) obtaining a reverse arrangement table of the N clustering center points, traversing the reverse arrangement coding tables of the N clustering center points in parallel through OpenMP, and calculating the distance between the N-dimensional feature vector of the image to be retrieved and the feature vector in the reverse arrangement coding table of each clustering center point; and obtaining a plurality of feature vectors closest to the n-dimensional feature vector of the image to be retrieved.
Further, in step S3), performing fast positioning infringement search on the image data to be searched by a large-scale data search method to obtain an infringement search result, including the following steps:
s361) respectively calculating Euclidean distances between a plurality of feature vectors closest to the n-dimensional feature vector of the image to be retrieved and the n-dimensional feature vector of the image to be retrieved;
s371) obtaining a plurality of perceptual hash values of the copyright images corresponding to a plurality of feature vectors closest to the n-dimensional feature vector of the image to be retrieved, and calculating Hamming distances between the perceptual hash values of the image to be retrieved and the perceptual hash values of the copyright images respectively;
s381) establishing a first distance scoring function fi＝w1d1i+w2d2iI 1, 2, 1, m is the total number of the plurality of copyright images; w is a1Is a feature vector distance weight, w2To perceive the hash value distance weight, d1iIs the Euclidean distance between the ith feature vector nearest to the n-dimensional feature vector of the image to be retrieved and the n-dimensional feature vector of the image to be retrieved, d2iThe Hamming distance between the perceptual hash value of the ith copyright image and the perceptual hash value of the image to be retrieved is obtained; f. ofiThe value of credit of the ith copyright image; the ith copyright image corresponds to the ith characteristic vector closest to the n-dimensional characteristic vector of the image to be retrieved;
s391) respectively calculating the score values of a plurality of copyright images, sequencing the score values, setting a score threshold, judging whether the maximum score value is greater than the score threshold, if so, acquiring the copyright image corresponding to the maximum score value, and identifying the image to be retrieved as the copyright image infringing the right corresponding to the maximum score value; if not, the image to be retrieved is indicated to have no infringement image retrieved.
Further, in step S3), performing fast positioning infringement search on the image data to be searched by a large-scale data search method to obtain an infringement search result, including the following steps:
s362) respectively calculating Euclidean distances between a plurality of feature vectors closest to the n-dimensional feature vector of the image to be retrieved and the n-dimensional feature vector of the image to be retrieved;
s363) setting a Euclidean distance threshold, and selecting k feature vectors of which Euclidean distances from the n-dimensional feature vectors of the image to be retrieved are larger than the Euclidean distance threshold from a plurality of feature vectors closest to the n-dimensional feature vectors of the image to be retrieved;
s364), acquiring the perceptual hash values of k copyright images corresponding to k eigenvectors, and respectively calculating the Hamming distance between the perceptual hash value of the image to be retrieved and the perceptual hash values of the k copyright images;
s365) setting a Hamming distance threshold, and selecting z feature vectors of which the Hamming distance from the perceptual hash value of the image to be retrieved is larger than the Hamming distance threshold from the k feature vectors; the z feature vectors correspond to the z copyright images;
s366) establishing a second distance scoring function Fj＝W1D1j+W2D2j，j＝1、2、...、z，W1Is the second eigenvector distance weight, W2Is a second perceptual hash value distance weight, D1jIs Euclidean distance between the feature vector of the jth copyright image in the z copyright images and the n-dimensional feature vector of the image to be retrieved, D2jThe Hamming distance between the perceptual hash value of the jth copyright image in the z copyright images and the perceptual hash value of the image to be retrieved is obtained; fjThe value of credit of the j-th copyright image in the z copyright images;
s367) respectively calculating the score values of the z copyright images, sequencing the score values, setting a second score threshold value, judging whether the maximum score value in the score values of the z copyright images is larger than the second score threshold value, if so, acquiring the copyright image corresponding to the maximum score value in the score values of the z copyright images, and identifying the image to be retrieved as the copyright image corresponding to the maximum score value in the score values infringing the z copyright images; if not, the image to be retrieved is indicated to have no infringement image retrieved.
Further, the deep learning model based on CNN is to select VGG16 or AlexNet convolutional neural network pre-trained on ImageNet.
On the other hand, the invention provides a device for rapidly positioning an infringement image for image copyright protection, which comprises an image acquisition unit, a model establishing unit and an infringement image rapid positioning and retrieving unit;
the image acquisition unit is used for acquiring copyright image data and an image to be retrieved;
the model establishing unit is used for establishing a picture infringement detection deep learning model and extracting a feature vector containing copyright image data and global features and local features of each image in an image to be retrieved by using the picture infringement detection deep learning model; calculating the perceptual hash value of the copyright image data and each image in the image to be retrieved;
and the infringement image rapid positioning retrieval unit is used for establishing a large-scale data retrieval method by utilizing the characteristic vector and the perception hash value in the model establishing unit, and rapidly positioning and infringement retrieval is carried out on the image data to be retrieved acquired in the image acquisition unit by the large-scale data retrieval method to obtain an infringement retrieval result.
In a further aspect, the invention provides a computer-readable storage medium, in which a computer program is stored, which, when executed by a processor, causes the processor to carry out the steps of the above-mentioned method.
The invention has the beneficial effects that: firstly, high-dimensional feature vectors (namely global features and local features) of the image are extracted through the combination of a CNN (computer network) and an R-MAC (radio-media access control) method, the perceptual hash value of the image is calculated, the image is identified and compared in two-layer scales, especially, the identification of anti-attack means possibly appearing in an infringing image is enhanced, and the robustness of the model is greatly enhanced; in addition, the invention provides a retrieval method for quickly retrieving the source-tracing infringement images of billions of images, which fully utilizes computing resources to carry out high-efficiency clustering algorithm and inverted list calculation, is suitable for retrieval of single images and retrieval of batch data, can realize millisecond-level query of ten-million-level data under the single-machine effect (such as 48-core cpu 128g RAM), and greatly improves the retrieval efficiency under mass data; meanwhile, aiming at the characteristics of image copyright protection, the method is different from the screening of a single threshold, uses a mode of combining multiple indexes, provides a method for automatically tuning parameters aiming at the characteristics of data, and ensures the accuracy and recall rate of retrieval to the maximum extent.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments are briefly described below, and it is obvious that the drawings in the following description are some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings without creative efforts.
Fig. 1 is a flowchart illustrating a method for rapidly positioning an infringement image for image copyright protection according to an embodiment of the present invention.
Fig. 2 is a schematic flow chart of extracting feature vectors by using a picture infringement detection deep learning model according to the first embodiment.
Fig. 3 is a schematic flowchart of calculating a perceptual hash value of each image according to the first embodiment.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail below with reference to the accompanying drawings. It should be understood that the detailed description and specific examples, while indicating the invention, are intended for purposes of illustration only and are not intended to limit the scope of the invention. It should be noted that the terms "first," "second," and the like in the description and claims of the present invention and in the drawings described above are used for distinguishing between similar elements and not necessarily for describing a particular sequential or chronological order, and it should be understood that the terms so used are interchangeable under appropriate circumstances and are merely used to describe the distinguishing manner in which the embodiments of the present invention distinguish between similar elements. Furthermore, the terms "comprises," "comprising," and "having," and any variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, system, article, or apparatus that comprises a list of elements is not necessarily limited to those elements, but may include other elements not expressly listed or inherent to such process, method, article, or apparatus.
In a first aspect, the present invention provides a method for rapidly locating an infringement image for image copyright protection, as shown in fig. 1, including the following steps:
s1) respectively obtaining copyright image data and an image to be retrieved, establishing a picture infringement detection deep learning model, and extracting feature vectors including global features and local features of each image in the copyright image data and the image to be retrieved by using the picture infringement detection deep learning model, as shown in fig. 2, including the following steps:
s11) carrying out first image preprocessing on the copyright image data and each image in the images to be retrieved, wherein the first image preprocessing comprises the step of scaling the copyright image data and each image in the images to be retrieved to a first preset size;
s12), establishing a CNN-based deep learning model, inputting each image preprocessed by the first image into the CNN-based deep learning model respectively, and taking the feature map of the last convolutional layer of the CNN-based deep learning model as output;
s13) extracting a plurality of areas with different scales from each feature map by adopting an R-MAC method to obtain a plurality of area R-MAC features, summing and pooling the plurality of area R-MAC features, and splicing to obtain n-dimensional feature vectors corresponding to each image; the n-dimensional feature vector corresponding to each image includes the n-dimensional feature vector of each image in the copyrighted-image data and the n-dimensional feature vector of the image to be retrieved.
Step S1), extracting the regional characteristics of the characteristic graph obtained by calculating the last convolution layer of the CNN-based deep learning model through an R-MAC method, generating a series of regional vectors for the local region specified in the center of the image by using the R-MAC method, then summing and pooling for vector aggregation, outputting the vectors as 512-dimensional characteristic vectors, and storing all vector results in a database to be used in the retrieval process. The CNN-based deep learning model uses ImageNet pre-training model parameters, and can be selected to be finely adjusted on specific data, so that the requirements of data with different orders of magnitude are met. The deep learning model based on CNN is to select VGG16 or AlexNet convolutional neural network pre-trained on ImageNet.
S2) calculating the perceptual hash value of each image in the copyrighted-image data and the image to be retrieved, as shown in fig. 3, including the steps of:
s21) carrying out second image preprocessing on the copyright image data and each image in the images to be retrieved, wherein the second image preprocessing comprises the step of scaling the copyright image data and each image in the images to be retrieved to a second preset size;
s22) converting each image scaled to the second preset size into a gray image, respectively;
s23) calculating the discrete cosine transform of each image after being converted into the gray level image to obtain a discrete cosine transform coefficient matrix;
s24) extracting a low-frequency matrix with a preset size at the upper left corner of the discrete cosine transform coefficient matrix, calculating the element average value of the low-frequency matrix, setting elements which are larger than or equal to the element average value in the low-frequency matrix to be 1, setting elements which are smaller than the element average value in the low-frequency matrix to be 0, and obtaining the low-frequency matrix with the elements set to be 1 or 0;
s25) flattening the low-frequency matrix with the element set to be 1 or 0 into a one-dimensional vector to obtain copyright image data and a perception hash value of each image in the image to be retrieved.
In order to enhance the attack resisting capability of the model for an infringement image, on the basis of deep learning characteristics, perceptual hash is selected as a posteriori measurement index, firstly, the image is scaled to the size of 32x32 (namely, copyright image data and each image in the image to be retrieved are subjected to second image preprocessing), image redundant information is reduced, meanwhile, the calculation of discrete cosine transform is accelerated, the image is converted into a 64-degree gray image, the calculation amount is further simplified, and the discrete cosine transform of the image is calculated.
S3) establishing a large-scale data retrieval method by utilizing the characteristic vector and the perception hash value, and rapidly positioning and infringing the retrieval of the image data to be retrieved by the large-scale data retrieval method to obtain an infringing retrieval result.
In step S3), a large-scale data retrieval method is established using the eigenvectors and the perceptual hash values, including constructing a retrieval index file and retrieving using the retrieval index file, the constructing of the retrieval index file including the steps of:
s31) initializing a retrieval index file data structure, wherein the retrieval index file data structure comprises a reverse arrangement table, a code table, a reverse arrangement vector ID table and/or a reverse arrangement vector coding table;
s32) training data are obtained, and a clustering algorithm is trained by using the training data; the training data comprises a plurality of data points, and the data points are respectively n-dimensional characteristic vectors of each image in the copyright image data; establishing a reverse vector ID table, wherein the reverse vector ID table is used for storing a plurality of data points and IDs of the data points;
s33) determining the number of the centers of the clusters and the number range of elements in the cluster where each cluster center is located according to the data volume of the training data;
s34) randomly initializing all cluster centers and updating all cluster centers and code tables, comprising the following steps:
s3031) initializing all clustering centers;
s3032) calculating any one clustering center CqShortest distance d (C) to other cluster centersq，Cw)；d(Cq，Cw) Represents the clustering center CqAnd a distance clustering center CqNearest cluster center CwThe distance of (d);
s3033) obtaining a clustering center CqThe data point x in the cluster to the cluster center CqDistance d (C)qX), judgment 2d (C)q，x)≤d(Cq，Cw) If yes, the classification position of the data point x is unchanged; if not, entering step S3034);
s3034) calculating the distance from the data point x to other clustering centers, and classifying the data point x into the cluster where the clustering center closest to the other clustering centers is located;
s3035) repeating the steps S3032) to S3034), and sequentially obtaining the clustering center CqA classification location for each data point in the cluster;
s3036) repeating the steps S3032) to S3035), and sequentially obtaining the classification position of each data point in the cluster where each cluster center is located;
s3037) updating all the clustering centers, judging whether all the clustering centers change, and if so, returning to the step S3032); if not, finishing the cluster updating, obtaining all the updated cluster centers and all the data points in the cluster of each cluster center, wherein each data point corresponds to one ID, adding all the updated cluster centers into a code table, storing all the data points in the cluster of each cluster center and the IDs of the data points into corresponding inverted arrangement tables, each cluster center corresponds to one inverted arrangement table, inverted IDs and inverted code tables are stored in the inverted arrangement tables, the inverted IDs are used for storing the IDs of the data points, and the inverted code tables are used for storing all the data points in the cluster of the cluster center.
The retrieval is carried out by utilizing the retrieval index file, and the method comprises the following steps:
s35), acquiring the constructed retrieval index file, calculating the vector distance between the N-dimensional feature vector of the image to be retrieved and all the updated clustering centers, and acquiring N clustering center points closest to the N-dimensional feature vector of the image to be retrieved;
s36) obtaining a reverse arrangement table of the N clustering center points, traversing the reverse arrangement coding tables of the N clustering center points in parallel through OpenMP, and calculating the distance between the N-dimensional feature vector of the image to be retrieved and the feature vector in the reverse arrangement coding table of each clustering center point; and obtaining a plurality of feature vectors closest to the n-dimensional feature vector of the image to be retrieved.
Inputting a training data training clustering algorithm, fixing the central number of clustering according to the data volume, firstly checking the number of elements participating in training clustering and the element number range in a preset clustering central point, and judging whether random sampling is carried out or not; randomly initializing a clustering center, adding the clustering center into a code table, and performing clustering iteration for specified times; finding out the nearest clustering center and distance of each element to be clustered, updating the clustering center, replacing the clustering center point before the mean value of each central element obtained by parallel calculation is replaced, finding out the splitting center meeting the conditions aiming at the clustering center with empty elements, modifying the splitting center point value through small symmetrical disturbance, taking the splitting center point and the original center point mean value as a substitute, updating the code table, and continuing iteration until the iteration is finished.
The invention reduces unnecessary distance calculation by using the triangle inequality and accelerates the clustering algorithm. When the clustering algorithm is trained, the element number range of each clustering center is fixed, so that the clustering centers are prevented from having no representativeness due to too few elements, and meanwhile, excessive data elements are prevented from training the model, and the training time is prevented from being increased.
When the nearest cluster center and distance of each element are calculated, excessive time is consumed when the data size is large because all elements need to be traversed, and the calculation process is optimized by adopting the BLAS linear algebra library and the SIMD AVX2 instruction set, so that the calculation time is greatly saved. And when a retrieval index file is created, respectively calculating the distances from all the training data to all the clustering centers, adding the distances to the clustering centers closest to the training data to the clustering centers, and storing the characteristic vectors into an inverted coding list so as to construct the retrieval index file.
The method comprises the steps of loading a constructed retrieval index file, inputting a feature vector of an image to be retrieved (the feature vector of the image to be retrieved is the feature vector of a single image to be queried or the feature vectors of a plurality of images), firstly querying N clustering centers closest to the feature vector of the image to be retrieved, obtaining an inverted arrangement table of the current N clustering centers, traversing the inverted arrangement coding table in parallel through OpenMP, calculating the distance between the vectors and querying, putting the result into a returned maximum heap, circularly accumulating the vectors returned by querying, sequencing the maximum heap, rapidly positioning fragment data by coarsely querying the N clustering centers to reduce the data volume needing to be calculated in the process, and fully utilizing the performance of a multi-core processor through OpenMP to accelerate the retrieval speed.
Step S3), rapidly positioning and infringing the search for the image data to be searched by a large-scale data search method to obtain an infringing search result, comprising the following steps:
s361) respectively calculating Euclidean distances between a plurality of feature vectors closest to the n-dimensional feature vector of the image to be retrieved and the n-dimensional feature vector of the image to be retrieved;
s371) obtaining a plurality of perceptual hash values of the copyright images corresponding to a plurality of feature vectors closest to the n-dimensional feature vector of the image to be retrieved, and calculating Hamming distances between the perceptual hash values of the image to be retrieved and the perceptual hash values of the copyright images respectively;
s381) constructionEstablishing a first distance scoring function fi＝w1d1i+w2d2iI 1, 2, 1, m is the total number of the plurality of copyright images; w is a1Is a feature vector distance weight, w2To perceive the hash value distance weight, d1iIs the Euclidean distance between the ith feature vector nearest to the n-dimensional feature vector of the image to be retrieved and the n-dimensional feature vector of the image to be retrieved, d2iThe Hamming distance between the perceptual hash value of the ith copyright image and the perceptual hash value of the image to be retrieved is obtained; f. ofiThe value of credit of the ith copyright image; the ith copyright image corresponds to the ith characteristic vector closest to the n-dimensional characteristic vector of the image to be retrieved;
s391) respectively calculating the score values of a plurality of copyright images, sequencing the score values, setting a score threshold, judging whether the maximum score value is greater than the score threshold, if so, acquiring the copyright image corresponding to the maximum score value, and identifying the image to be retrieved as the copyright image infringing the right corresponding to the maximum score value; if not, the image to be retrieved is indicated to have no infringement image retrieved.
In this embodiment, a result (an euclidean distance) returned by the constructed retrieval index file is used, a plurality of perceptual hash values of the copyrighted images corresponding to a plurality of feature vectors closest to the n-dimensional feature vector of the image to be retrieved are found through the inverted vector ID table, hamming distances between the perceptual hash values of the image to be retrieved and the perceptual hash values of the copyrighted images are calculated and calculated, a distance scoring function is used for screening and sorting, a scoring function and two weighting threshold screening criteria are constructed based on two distance indexes according to infringement data characteristics, and the result is finally scored and sorted.
On the other hand, the embodiment of the invention provides a device for quickly positioning an infringement image for image copyright protection, which comprises an image acquisition unit, a model establishing unit and an infringement image quick positioning and retrieving unit;
the image acquisition unit is used for acquiring copyright image data and an image to be retrieved;
the model establishing unit is used for establishing a picture infringement detection deep learning model and extracting a feature vector containing copyright image data and global features and local features of each image in an image to be retrieved by using the picture infringement detection deep learning model; calculating the perceptual hash value of the copyright image data and each image in the image to be retrieved;
and the infringement image rapid positioning retrieval unit is used for establishing a large-scale data retrieval method by utilizing the characteristic vector and the perception hash value in the model establishing unit, and rapidly positioning and infringement retrieval is carried out on the image data to be retrieved acquired in the image acquisition unit by the large-scale data retrieval method to obtain an infringement retrieval result.
In yet another aspect, the present invention provides a computer-readable storage medium, which stores a computer program, and when the computer program is executed by a processor, the processor is caused to execute the steps of the above method.
In step S3), performing fast positioning infringement search on image data to be searched by a large-scale data search method to obtain an infringement search result, where the method is applied to fast positioning infringement images for image copyright protection, and includes the following steps:
s362) respectively calculating Euclidean distances between a plurality of feature vectors closest to the n-dimensional feature vector of the image to be retrieved and the n-dimensional feature vector of the image to be retrieved;
s363) setting a Euclidean distance threshold, and selecting k feature vectors of which Euclidean distances from the n-dimensional feature vectors of the image to be retrieved are larger than the Euclidean distance threshold from a plurality of feature vectors closest to the n-dimensional feature vectors of the image to be retrieved;
s364), acquiring the perceptual hash values of k copyright images corresponding to k eigenvectors, and respectively calculating the Hamming distance between the perceptual hash value of the image to be retrieved and the perceptual hash values of the k copyright images;
s365) setting a Hamming distance threshold, and selecting z feature vectors of which the Hamming distance from the perceptual hash value of the image to be retrieved is larger than the Hamming distance threshold from the k feature vectors; the z feature vectors correspond to the z copyright images;
s366) establishing a second distance scoring function Fj＝W1D1j+W2D2j，j＝1、2、...、z，W1Is the second eigenvector distance weight, W2Is a second perceptual hash value distance weight, D1jIs Euclidean distance between the feature vector of the jth copyright image in the z copyright images and the n-dimensional feature vector of the image to be retrieved, D2jThe Hamming distance between the perceptual hash value of the jth copyright image in the z copyright images and the perceptual hash value of the image to be retrieved is obtained; fjThe value of credit of the j-th copyright image in the z copyright images;
s367) respectively calculating the score values of the z copyright images, sequencing the score values, setting a second score threshold value, judging whether the maximum score value in the score values of the z copyright images is larger than the second score threshold value, if so, acquiring the copyright image corresponding to the maximum score value in the score values of the z copyright images, and identifying the image to be retrieved as the copyright image corresponding to the maximum score value in the score values infringing the z copyright images; if not, the image to be retrieved is indicated to have no infringement image retrieved.
The other contents of the second embodiment are the same as those of the first embodiment, and are not described herein again.
By adopting the technical scheme disclosed by the invention, the following beneficial effects are obtained:
firstly, high-dimensional feature vectors (global features and local features) of the image are extracted by combining a CNN network and an R-MAC method, a perceptual hash value of the image is calculated, the image is identified and compared by two layers of scales, and particularly, the identification of anti-attack means possibly appearing in an infringing image is enhanced, so that the robustness of the model is greatly enhanced; in addition, the invention provides a retrieval method for quickly retrieving the source-tracing infringement images of hundred million-level images, which fully utilizes computing resources to carry out high-efficiency clustering algorithm and inverted list calculation, is suitable for retrieval of single images and retrieval of batch data, can realize millisecond-level query of million-level data under the single-machine effect, and greatly improves the retrieval efficiency under mass data; meanwhile, aiming at the characteristics of image copyright protection, the method is different from the screening of a single threshold, uses a mode of combining multiple indexes, provides a method for automatically tuning parameters aiming at the characteristics of data, and ensures the accuracy and recall rate of retrieval to the maximum extent.
The foregoing is only a preferred embodiment of the present invention, and it should be noted that it will be apparent to those skilled in the art that various modifications and improvements can be made without departing from the principle of the present invention, and such modifications and improvements should also be considered within the scope of the present invention.
Claims (9)
1. The method for rapidly positioning the infringement image applied to the image copyright protection is characterized by comprising the following steps of:
s1) respectively acquiring copyright image data and an image to be retrieved, establishing a picture infringement detection deep learning model, and extracting a feature vector of each image in the copyright image data and the image to be retrieved by using the picture infringement detection deep learning model;
s2) calculating the perception hash value of the copyright image data and each image in the image to be retrieved;
s3) establishing a large-scale data retrieval method by using the feature vector and the perceptual hash value, and performing rapid positioning infringement retrieval on the image data to be retrieved by the large-scale data retrieval method to obtain an infringement retrieval result;
in step S3), a large-scale data retrieval method is established using the feature vector and the perceptual hash value, including constructing a retrieval index file and retrieving using the retrieval index file, where constructing the retrieval index file includes the following steps:
s31), initializing a retrieval index file data structure, wherein the retrieval index file data structure comprises a reverse arrangement table, a code table, a reverse arrangement vector ID table and/or a reverse arrangement vector coding table;
s32) obtaining training data, and training a clustering algorithm by using the training data; the training data comprises a plurality of data points, and the data points are respectively n-dimensional feature vectors of each image in the copyright image data; establishing a reverse vector ID table, wherein the reverse vector ID table is used for storing a plurality of data points and IDs of the data points;
s33) determining the number of the centers of the clusters and the number range of elements in the cluster where each cluster center is located according to the data volume of the training data;
s34) randomly initializing all cluster centers and updating all cluster centers and code tables, comprising the following steps:
s3031) initializing all clustering centers;
s3032) calculating any one clustering center CqShortest distance d (C) to other cluster centersq，Cw)，d(Cq，Cw) Represents the clustering center CqAnd is distant from the cluster center CqNearest cluster center CwThe distance of (d);
s3033) obtaining a clustering center CqThe data point x in the cluster to the cluster center CqDistance d (C)qX), judgment 2d (C)q，x)≤d(Cq，Cw) If yes, the classification position of the data point x is unchanged; if not, entering step S3034);
s3034) calculating the distance from the data point x to other clustering centers, and classifying the data point x into the cluster where the clustering center closest to the other clustering centers is located;
s3035) repeating the steps S3032) to S3034), and sequentially obtaining the clustering center CqA classification location for each data point in the cluster;
s3036) repeating the steps S3032) to S3035), and sequentially obtaining the classification position of each data point in the cluster where each cluster center is located;
s3037) updating all the clustering centers, judging whether all the clustering centers change, and if so, returning to the step S3032); if not, finishing the cluster updating, obtaining all the updated cluster centers and all the data points in the cluster of each cluster center, wherein each data point corresponds to one ID, adding all the updated cluster centers into a code table, storing all the data points in the cluster of each cluster center and the IDs of the data points into corresponding inverted arrangement tables, each cluster center corresponds to one inverted arrangement table, inverted IDs and inverted code tables are stored in the inverted arrangement tables, the inverted IDs are used for storing the IDs of the data points, and the inverted code tables are used for storing all the data points in the cluster of the cluster center.
2. The method for rapidly positioning infringement image for image copyright protection according to claim 1, wherein in step S1), separately acquiring copyright image data and image to be retrieved, establishing a deep learning model for picture infringement detection, and extracting feature vectors of each image in the copyright image data and the image to be retrieved by using the deep learning model for picture infringement detection, the method comprises the following steps:
s11) performing a first image preprocessing on the copyrighted-image data and each of the images to be retrieved, the first image preprocessing including scaling the copyrighted-image data and each of the images to be retrieved to a first preset size;
s12), establishing a CNN-based deep learning model, inputting each image preprocessed by a first image into the CNN-based deep learning model, and taking a feature map of the last convolutional layer of the CNN-based deep learning model as output;
s13) extracting a plurality of areas with different scales from each feature map by adopting an R-MAC method to obtain a plurality of area R-MAC features, summing and pooling the plurality of area R-MAC features, and splicing to obtain n-dimensional feature vectors corresponding to each image; the n-dimensional feature vector corresponding to each image comprises the n-dimensional feature vector of each image in the copyright image data and the n-dimensional feature vector of the image to be retrieved.
3. The method for rapidly positioning infringed image for image copyright protection according to claim 1 or 2, wherein in step S2), the step of calculating the perceptual hash value of each image in the copyrighted-image data and the image to be retrieved comprises the following steps:
s21) performing second image preprocessing on the copyrighted-image data and each of the images to be retrieved, the second image preprocessing including scaling the copyrighted-image data and each of the images to be retrieved to a second preset size;
s22) converting each image scaled to the second preset size into a gray image, respectively;
s23) calculating the discrete cosine transform of each image after being converted into the gray level image to obtain a discrete cosine transform coefficient matrix;
s24) extracting a low-frequency matrix with a preset size at the upper left corner of the discrete cosine transform coefficient matrix, calculating the element average value of the low-frequency matrix, setting elements which are larger than or equal to the element average value in the low-frequency matrix to be 1, setting elements which are smaller than the element average value in the low-frequency matrix to be 0, and obtaining the low-frequency matrix with the elements set to be 1 or 0;
s25) flattening the low-frequency matrix with the element set to be 1 or 0 into a one-dimensional vector to obtain the copyright image data and the perception hash value of each image in the image to be retrieved.
4. The method for rapidly positioning infringed image for image copyright protection according to claim 1, wherein the retrieval using the retrieval index file comprises the following steps:
s35), acquiring the constructed retrieval index file, calculating the vector distance between the N-dimensional feature vector of the image to be retrieved and all updated clustering centers, and acquiring N clustering center points closest to the N-dimensional feature vector of the image to be retrieved;
s36) obtaining the inverted arrangement table of the N clustering center points, traversing the inverted arrangement code table of the N clustering center points in parallel through OpenMP, and calculating the distance between the N-dimensional feature vector of the image to be retrieved and the feature vector in the inverted arrangement code table of each clustering center point; and obtaining a plurality of feature vectors closest to the n-dimensional feature vector of the image to be retrieved.
5. The method for rapidly positioning infringement image for image copyright protection according to claim 4, wherein in step S3), the rapid positioning infringement retrieval is performed on the image data to be retrieved by the large-scale data retrieval method to obtain an infringement retrieval result, and the method comprises the following steps:
s361) respectively calculating Euclidean distances between a plurality of feature vectors closest to the n-dimensional feature vector of the image to be retrieved and the n-dimensional feature vector of the image to be retrieved;
s371) obtaining a plurality of perceptual hash values of the copyright images corresponding to a plurality of feature vectors closest to the n-dimensional feature vector of the image to be retrieved, and calculating Hamming distances between the perceptual hash values of the image to be retrieved and the perceptual hash values of the plurality of copyright images respectively;
s381) establishing a first distance scoring function fi＝w1d1i+w2d2iI 1, 2, and m, wherein m is the total number of the plurality of copyright images; w is a1Is a feature vector distance weight, w2To perceive the hash value distance weight, d1iIs the Euclidean distance between the ith feature vector nearest to the n-dimensional feature vector of the image to be retrieved and the n-dimensional feature vector of the image to be retrieved, d2iThe Hamming distance between the perceptual hash value of the ith copyright image and the perceptual hash value of the image to be retrieved is obtained; f. ofiThe value of credit of the ith copyright image; the ith copyright image corresponds to the ith characteristic vector closest to the n-dimensional characteristic vector of the image to be retrieved;
s391) respectively calculating the score values of the plurality of copyright images, sequencing the score values, setting a score threshold, judging whether the maximum score value is larger than the score threshold, if so, acquiring the copyright image corresponding to the maximum score value, and determining the image to be retrieved as the copyright image corresponding to the maximum score value infringed; if not, the image to be retrieved is indicated to have no infringement image retrieved.
6. The method for rapidly positioning infringement image for image copyright protection according to claim 4, wherein in step S3), the rapid positioning infringement retrieval is performed on the image data to be retrieved by the large-scale data retrieval method to obtain an infringement retrieval result, and the method comprises the following steps:
s362) respectively calculating Euclidean distances between a plurality of feature vectors nearest to the n-dimensional feature vector of the image to be retrieved and the n-dimensional feature vector of the image to be retrieved;
s363) setting a Euclidean distance threshold, and selecting k eigenvectors of which the Euclidean distance from the n-dimensional eigenvector of the image to be retrieved is larger than the Euclidean distance threshold from a plurality of eigenvectors closest to the n-dimensional eigenvector of the image to be retrieved;
s364), acquiring the perceptual hash values of k copyright images corresponding to the k characteristic vectors, and respectively calculating the Hamming distance between the perceptual hash value of the image to be retrieved and the perceptual hash values of the k copyright images;
s365) setting a Hamming distance threshold, and selecting z feature vectors of which the Hamming distance from the k feature vectors to the perceptual hash value of the image to be retrieved is larger than the Hamming distance threshold; the z feature vectors correspond to z copyrighted images;
s366) establishing a second distance scoring function Fj＝W1D1j+W2D2j，j＝1、2、...、z，W1Is the second eigenvector distance weight, W2Is a second perceptual hash value distance weight, D1jIs Euclidean distance between the feature vector of the jth copyright image in the z copyright images and the n-dimensional feature vector of the image to be retrieved, D2jThe Hamming distance between the perceptual hash value of the jth copyright image in the z copyright images and the perceptual hash value of the image to be retrieved is obtained; fjThe value of credit of the j-th copyright image in the z copyright images;
s367) calculating the score values of the z copyrighted images, sorting the score values, setting a second score threshold, determining whether a maximum score value of the score values of the z copyrighted images is greater than the second score threshold, if so, obtaining the copyrighted image corresponding to the maximum score value of the score values of the z copyrighted images, and regarding the image to be retrieved as the copyrighted image corresponding to the maximum score value of the score values infringing the z copyrighted images; if not, the image to be retrieved is indicated to have no infringement image retrieved.
7. The method for rapidly positioning infringement images for image copyright protection according to claim 2, wherein the deep learning model based on CNN is a convolutional neural network selected from VGG16 or AlexNet pre-trained on ImageNet.
8. The device for rapidly positioning infringed image for image copyright protection is suitable for the method for rapidly positioning infringed image for image copyright protection according to any one of claims 1 to 7, and is characterized by comprising an image acquisition unit, a model building unit and an infringed image rapid positioning retrieval unit;
the image acquisition unit is used for acquiring copyright image data and an image to be retrieved;
the model establishing unit is used for establishing a picture infringement detection deep learning model, and extracting feature vectors containing the copyright image data and global features and local features of each image in the image to be retrieved by using the picture infringement detection deep learning model; calculating the perception hash value of the copyright image data and each image in the image to be retrieved;
the infringement image rapid positioning retrieval unit is used for establishing a large-scale data retrieval method by utilizing the characteristic vector and the perception hash value in the model establishing unit, and rapidly positioning and infringement retrieval is carried out on the image data to be retrieved acquired in the image acquisition unit by the large-scale data retrieval method to obtain an infringement retrieval result.
9. A computer-readable storage medium, storing a computer program which, when executed by a processor, causes the processor to carry out the steps of the method according to any one of claims 1 to 7.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011446862.9A CN112417381B (en) | 2020-12-11 | 2020-12-11 | Method and device for rapidly positioning infringement image applied to image copyright protection |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011446862.9A CN112417381B (en) | 2020-12-11 | 2020-12-11 | Method and device for rapidly positioning infringement image applied to image copyright protection |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112417381A CN112417381A (en) | 2021-02-26 |
| CN112417381B true CN112417381B (en) | 2021-06-22 |
Family
ID=74776245
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011446862.9A Active CN112417381B (en) | 2020-12-11 | 2020-12-11 | Method and device for rapidly positioning infringement image applied to image copyright protection |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112417381B (en) |
Families Citing this family (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113011342B (en) * | 2021-03-22 | 2022-06-14 | 支付宝(杭州)信息技术有限公司 | Feature extraction and processing method and system for copyright infringement retrieval |
| CN112990792B (en) * | 2021-05-11 | 2021-08-31 | 北京智源人工智能研究院 | Method and device for automatically detecting infringement risk and electronic equipment |
| CN113326388A (en) * | 2021-05-20 | 2021-08-31 | 上海云从汇临人工智能科技有限公司 | Data retrieval method, system, medium and device based on inverted list |
| CN113239226A (en) * | 2021-06-01 | 2021-08-10 | 平安科技（深圳）有限公司 | Image retrieval method, device, equipment and storage medium |
| CN113435485A (en) * | 2021-06-15 | 2021-09-24 | 北京百度网讯科技有限公司 | Picture detection method and device, electronic equipment and storage medium |
| CN114359590A (en) * | 2021-12-06 | 2022-04-15 | 支付宝(杭州)信息技术有限公司 | NFT image work infringement detection method and device and computer storage medium |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111753735A (en) * | 2020-06-24 | 2020-10-09 | 北京奇艺世纪科技有限公司 | Video clip detection method and device, electronic equipment and storage medium |
Family Cites Families (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11074495B2 (en) * | 2013-02-28 | 2021-07-27 | Z Advanced Computing, Inc. (Zac) | System and method for extremely efficient image and pattern recognition and artificial intelligence platform |
| CN105095435A (en) * | 2015-07-23 | 2015-11-25 | 北京京东尚科信息技术有限公司 | Similarity comparison method and device for high-dimensional image features |
| CN107256262B (en) * | 2017-06-13 | 2020-04-14 | 西安电子科技大学 | Image retrieval method based on object detection |
| CN108171264A (en) * | 2017-12-26 | 2018-06-15 | 北京非斗数据科技发展有限公司 | A kind of extraction identification technology encoded using deep learning combination Hash to picture infringing content |
| CN110033018B (en) * | 2019-03-06 | 2023-10-31 | 平安科技（深圳）有限公司 | Graph similarity judging method and device and computer readable storage medium |
| CN110968721A (en) * | 2019-11-28 | 2020-04-07 | 上海冠勇信息科技有限公司 | Method and system for searching infringement of mass images and computer readable storage medium thereof |
| CN111198959B (en) * | 2019-12-30 | 2023-03-28 | 郑州轻工业大学 | Two-stage image retrieval method based on convolutional neural network |
| CN111310129A (en) * | 2020-02-20 | 2020-06-19 | 北京海益同展信息科技有限公司 | Method, device and storage medium for maintaining right of image |
-
2020
- 2020-12-11 CN CN202011446862.9A patent/CN112417381B/en active Active
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111753735A (en) * | 2020-06-24 | 2020-10-09 | 北京奇艺世纪科技有限公司 | Video clip detection method and device, electronic equipment and storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112417381A (en) | 2021-02-26 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112417381B (en) | Method and device for rapidly positioning infringement image applied to image copyright protection | |
| CN111753060B (en) | Information retrieval method, apparatus, device and computer readable storage medium | |
| CN110717534B (en) | Target classification and positioning method based on network supervision | |
| CN102693299B (en) | System and method for parallel video copy detection | |
| Zheng et al. | $\mathcal {L} _p $-Norm IDF for Scalable Image Retrieval | |
| CN113918753B (en) | Image retrieval method based on artificial intelligence and related equipment | |
| US20120243789A1 (en) | Fast image classification by vocabulary tree based image retrieval | |
| CN109960738B (en) | Large-scale remote sensing image content retrieval method based on depth countermeasure hash learning | |
| US20160019671A1 (en) | Identifying multimedia objects based on multimedia fingerprint | |
| CN112395457B (en) | Video to-be-retrieved positioning method applied to video copyright protection | |
| CN102144231A (en) | Adaptive visual similarity for text-based image search results re-ranking | |
| CN106033426A (en) | A latent semantic min-Hash-based image retrieval method | |
| CN112948601B (en) | Cross-modal hash retrieval method based on controlled semantic embedding | |
| CN108491430A (en) | It is a kind of based on the unsupervised Hash search method clustered to characteristic direction | |
| CN114358188A (en) | Feature extraction model processing method, feature extraction model processing device, sample retrieval method, sample retrieval device and computer equipment | |
| Wang et al. | Beauty product image retrieval based on multi-feature fusion and feature aggregation | |
| CN110442749B (en) | Video frame processing method and device | |
| CN114491115B (en) | Multi-model fusion integrated image retrieval method based on deep hash | |
| CN116975340A (en) | Information retrieval method, apparatus, device, program product, and storage medium | |
| CN106570196B (en) | Video program searching method and device | |
| Furuta et al. | Efficient and interactive spatial-semantic image retrieval | |
| CN105760875A (en) | Binary image feature similarity discrimination method based on random forest algorithm | |
| CN105183845A (en) | ERVQ image indexing and retrieval method in combination with semantic features | |
| CN112632406B (en) | Query method, query device, electronic equipment and storage medium | |
| CN112579783B (en) | Short text clustering method based on Laplace atlas |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |