CN113537206B - Push data detection method, push data detection device, computer equipment and storage medium - Google Patents
Push data detection method, push data detection device, computer equipment and storage medium Download PDFInfo
- Publication number
- CN113537206B CN113537206B CN202010756055.0A CN202010756055A CN113537206B CN 113537206 B CN113537206 B CN 113537206B CN 202010756055 A CN202010756055 A CN 202010756055A CN 113537206 B CN113537206 B CN 113537206B
- Authority
- CN
- China
- Prior art keywords
- target object
- detected
- text
- push
- influence
- 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
- 238000001514 detection method Methods 0.000 title claims abstract description 224
- 238000000034 method Methods 0.000 claims abstract description 44
- 238000013145 classification model Methods 0.000 claims abstract description 32
- 238000012549 training Methods 0.000 claims description 154
- 230000002441 reversible effect Effects 0.000 claims description 153
- 238000004422 calculation algorithm Methods 0.000 claims description 41
- 238000004364 calculation method Methods 0.000 claims description 19
- 238000004590 computer program Methods 0.000 claims description 16
- 238000013527 convolutional neural network Methods 0.000 claims description 15
- 230000011218 segmentation Effects 0.000 claims description 14
- 238000009795 derivation Methods 0.000 claims description 4
- 238000000605 extraction Methods 0.000 claims description 3
- 238000010586 diagram Methods 0.000 description 22
- 230000006870 function Effects 0.000 description 17
- 238000013528 artificial neural network Methods 0.000 description 14
- 238000011176 pooling Methods 0.000 description 10
- 238000012545 processing Methods 0.000 description 10
- 238000010801 machine learning Methods 0.000 description 9
- 238000012015 optical character recognition Methods 0.000 description 9
- 239000000284 extract Substances 0.000 description 6
- 238000013473 artificial intelligence Methods 0.000 description 5
- 230000004927 fusion Effects 0.000 description 5
- 230000007246 mechanism Effects 0.000 description 5
- 230000004913 activation Effects 0.000 description 4
- 238000012550 audit Methods 0.000 description 4
- 238000005516 engineering process Methods 0.000 description 4
- 238000007726 management method Methods 0.000 description 4
- 238000003058 natural language processing Methods 0.000 description 4
- 238000007781 pre-processing Methods 0.000 description 4
- 238000011160 research Methods 0.000 description 3
- 235000013334 alcoholic beverage Nutrition 0.000 description 2
- 230000006399 behavior Effects 0.000 description 2
- 239000008280 blood Substances 0.000 description 2
- 210000004369 blood Anatomy 0.000 description 2
- 238000007405 data analysis Methods 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 238000013136 deep learning model Methods 0.000 description 2
- 230000000670 limiting effect Effects 0.000 description 2
- 238000012417 linear regression Methods 0.000 description 2
- 230000008569 process Effects 0.000 description 2
- 238000007637 random forest analysis Methods 0.000 description 2
- 230000003068 static effect Effects 0.000 description 2
- 102100032202 Cornulin Human genes 0.000 description 1
- LFQSCWFLJHTTHZ-UHFFFAOYSA-N Ethanol Chemical compound CCO LFQSCWFLJHTTHZ-UHFFFAOYSA-N 0.000 description 1
- 101000920981 Homo sapiens Cornulin Proteins 0.000 description 1
- 241001465754 Metazoa Species 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000003190 augmentative effect Effects 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 125000004122 cyclic group Chemical group 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 238000003709 image segmentation Methods 0.000 description 1
- 230000006698 induction Effects 0.000 description 1
- 238000007689 inspection Methods 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 238000002372 labelling Methods 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 230000036961 partial effect Effects 0.000 description 1
- 230000002829 reductive effect Effects 0.000 description 1
- 230000002787 reinforcement Effects 0.000 description 1
- 230000001360 synchronised effect Effects 0.000 description 1
- 238000013526 transfer learning Methods 0.000 description 1
- 238000013519 translation Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/35—Clustering; Classification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/205—Parsing
- G06F40/216—Parsing using statistical methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/279—Recognition of textual entities
- G06F40/289—Phrasal analysis, e.g. finite state techniques or chunking
-
- 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
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- 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
-
- 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
- G06N3/047—Probabilistic or stochastic networks
-
- 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
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L67/00—Network arrangements or protocols for supporting network services or applications
- H04L67/50—Network services
- H04L67/55—Push-based network services
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Probability & Statistics with Applications (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Computer Networks & Wireless Communication (AREA)
- Signal Processing (AREA)
- Databases & Information Systems (AREA)
- Image Analysis (AREA)
Abstract
The application relates to an information push detection method, an information push detection device, computer equipment and a storage medium. The method comprises the following steps: acquiring information to be detected, wherein the information to be detected comprises an image to be detected and a text to be detected; extracting a target object area in an image to be detected, identifying a target object category in the target object area, and determining a first pushing influence characteristic corresponding to information to be detected according to the target object category; extracting an image text in the image to be detected, and matching the image text through a preset keyword library to obtain a second pushing influence characteristic corresponding to the information to be detected; inputting the text to be detected into a text classification model to be detected for text classification to obtain a text category to be detected, and obtaining a third pushing influence characteristic corresponding to the information to be detected according to the text category to be detected; and determining a push detection result corresponding to the information to be detected based on the first push influence feature, the second push influence feature and the third push influence feature. By adopting the method, the accuracy of information push detection can be improved.
Description
Technical Field
The present application relates to the field of internet technologies, and in particular, to a method and apparatus for detecting push data, a computer device, and a storage medium.
Background
With the development of internet technology, people begin to learn information through the internet. Currently, information is pushed through an internet platform conveniently and quickly, for example, video is pushed through a video website, live video is pushed through a live broadcast website, news is pushed through a news website, and friend circle information is sent through an instant messaging application. At present, information pushing is carried out through the Internet, the Internet platform can detect pushed information, the information which is forbidden to be pushed is prevented from being pushed, and information safety is guaranteed. At present, the detection of pushed information is usually performed according to expert opinions, and the push is performed by the detection, however, the detection efficiency and the accuracy of the detection method by the expert opinion are low.
Disclosure of Invention
In view of the foregoing, it is desirable to provide an information push detection method, apparatus, computer device, and storage medium that can improve the efficiency and accuracy of information push detection.
An information push detection method, the method comprising:
acquiring information to be detected, wherein the information to be detected comprises an image to be detected and a text to be detected;
extracting a target object area in an image to be detected, identifying the category of a target object in the target object area, and determining a first pushing influence characteristic corresponding to information to be detected according to the category of the target object;
extracting an image text in the image to be detected, and matching the image text through a preset keyword library to obtain a second pushing influence characteristic corresponding to the information to be detected;
inputting the text to be detected into a text classification model to be detected for text classification, obtaining a text category to be detected, and obtaining a third pushing influence characteristic corresponding to the information to be detected according to the text category to be detected;
and determining a push detection result corresponding to the information to be detected based on the first push influence feature, the second push influence feature and the third push influence feature.
An information push detection device, the device comprising:
the information acquisition module is used for acquiring information to be detected, wherein the information to be detected comprises an image to be detected and a text to be detected;
the first feature determining module is used for extracting a target object area in the image to be detected, identifying the category of a target object in the target object area, and determining a first pushing influence feature corresponding to the information to be detected according to the category of the target object;
The second feature obtaining module is used for extracting image texts in the images to be detected, and matching the image texts through a preset keyword library to obtain second pushing influence features corresponding to the information to be detected;
the third characteristic obtaining module is used for inputting the text to be detected into the text classification model to be detected for text classification, obtaining the text category to be detected, and obtaining a third pushing influence characteristic corresponding to the information to be detected according to the text category to be detected;
the result determining module is used for determining a push detection result corresponding to the information to be detected based on the first push influence feature, the second push influence feature and the third push influence feature.
A computer device comprising a memory storing a computer program and a processor which when executing the computer program performs the steps of:
acquiring information to be detected, wherein the information to be detected comprises an image to be detected and a text to be detected;
extracting a target object area in an image to be detected, identifying the category of a target object in the target object area, and determining a first pushing influence characteristic corresponding to information to be detected according to the category of the target object;
extracting an image text in the image to be detected, and matching the image text through a preset keyword library to obtain a second pushing influence characteristic corresponding to the information to be detected;
Inputting the text to be detected into a text classification model to be detected for text classification, obtaining a text category to be detected, and obtaining a third pushing influence characteristic corresponding to the information to be detected according to the text category to be detected;
and determining a push detection result corresponding to the information to be detected based on the first push influence feature, the second push influence feature and the third push influence feature.
A computer readable storage medium having stored thereon a computer program which when executed by a processor performs the steps of:
acquiring information to be detected, wherein the information to be detected comprises an image to be detected and a text to be detected;
extracting a target object area in an image to be detected, identifying the category of a target object in the target object area, and determining a first pushing influence characteristic corresponding to information to be detected according to the category of the target object;
extracting an image text in the image to be detected, and matching the image text through a preset keyword library to obtain a second pushing influence characteristic corresponding to the information to be detected;
inputting the text to be detected into a text classification model to be detected for text classification, obtaining a text category to be detected, and obtaining a third pushing influence characteristic corresponding to the information to be detected according to the text category to be detected;
And determining a push detection result corresponding to the information to be detected based on the first push influence feature, the second push influence feature and the third push influence feature.
According to the information push detection method, the information push detection device, the computer equipment and the storage medium, the first push influence characteristic is determined according to the category of the target object in the target object area of the image to be detected, the second push influence characteristic is obtained according to the image text in the image to be detected through the preset keyword library, meanwhile, the text to be detected is classified, the third push influence characteristic is obtained according to the category of the text to be detected, then the push detection result corresponding to the information to be detected is determined according to the first push influence characteristic, the second push influence characteristic and the third push influence characteristic, namely, the corresponding push influence characteristic is obtained according to different types of data respectively, then the push detection result corresponding to the information to be detected is determined according to different push influence characteristics, and the information push detection efficiency and the information push detection accuracy are improved.
Drawings
FIG. 1 is an application environment diagram of an information push detection method in one embodiment;
FIG. 2 is a flow chart of a method for detecting information push in an embodiment;
FIG. 3 is a flow diagram of one implementation of obtaining a first push impact feature;
FIG. 4 is a flow diagram of a method for obtaining a first push impact feature in one implementation;
FIG. 5 is a schematic flow chart of detecting an advertising image of wine in one embodiment;
FIG. 6 is a flow diagram of determining a target object class in one implementation;
FIG. 7 is a schematic diagram of the attention mechanism in one implementation;
FIG. 8 is a flow diagram of a trained target object region detection network in one implementation;
FIG. 9 is a flow diagram of a trained target object recognition network in one implementation;
FIG. 10 is a flow diagram of one implementation of a second push impact feature;
FIG. 11 is a schematic flow chart of detecting text in an advertising image of wine in one embodiment;
FIG. 12 is a flow diagram of a trained target object recognition network in one implementation;
FIG. 13 is a flow diagram of word frequency and reverse order file frequency obtained in one implementation;
FIG. 14 is a schematic diagram of a network structure of a text classification model to be detected in one implementation;
FIG. 15 is a flow chart of a push detection result in one implementation;
FIG. 16 is a flow chart of a method for detecting information push in one embodiment;
FIG. 17 is a schematic diagram of a framework of a method for detecting information push in one embodiment;
FIG. 18 is a schematic diagram of an application scenario of an information push detection method in an embodiment;
FIG. 19 is a block diagram of an information push detection device in one embodiment;
fig. 20 is an internal structural view of a computer device in one embodiment.
Detailed Description
The present application will be described in further detail with reference to the drawings and examples, in order to make the objects, technical solutions and advantages of the present application more apparent. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the application.
Computer Vision (CV) is a science of studying how to "look" a machine, and more specifically, to replace human eyes with a camera and a Computer to perform machine Vision such as recognition, tracking and measurement on a target, and further perform graphic processing to make the Computer process into an image more suitable for human eyes to observe or transmit to an instrument to detect. As a scientific discipline, computer vision research-related theory and technology has attempted to build artificial intelligence systems that can acquire information from images or multidimensional data. Computer vision techniques typically include image processing, image recognition, image semantic understanding, image retrieval, OCR, video processing, video semantic understanding, video content/behavior recognition, three-dimensional object reconstruction, 3D techniques, virtual reality, augmented reality, synchronous positioning, and map construction, among others, as well as common biometric recognition techniques such as face recognition, fingerprint recognition, and others.
Natural language processing (Nature Language processing, NLP) is an important direction in the fields of computer science and artificial intelligence. It is studying various theories and methods that enable effective communication between a person and a computer in natural language. Natural language processing is a science that integrates linguistics, computer science, and mathematics. Thus, the research in this field will involve natural language, i.e. language that people use daily, so it has a close relationship with the research in linguistics. Natural language processing techniques typically include text processing, semantic understanding, machine translation, robotic questions and answers, knowledge graph techniques, and the like.
Machine Learning (ML) is a multi-domain interdisciplinary, involving multiple disciplines such as probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory, etc. It is specially studied how a computer simulates or implements learning behavior of a human to acquire new knowledge or skills, and reorganizes existing knowledge structures to continuously improve own performance. Machine learning is the core of artificial intelligence, a fundamental approach to letting computers have intelligence, which is applied throughout various areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, confidence networks, reinforcement learning, transfer learning, induction learning, teaching learning, and the like.
The scheme provided by the embodiment of the application relates to technologies such as image processing, text processing, deep learning and the like of artificial intelligence, and is specifically described by the following embodiments:
the information push detection method provided by the application can be applied to an application environment shown in figure 1. Wherein the terminal 102 communicates with the server 104 via a network. The server 104 acquires information to be detected, which is sent by the terminal 102 and comprises an image to be detected and a text to be detected; the server 104 extracts a target object area in the image to be detected, identifies the category of the target object in the target object area, and determines a first pushing influence characteristic corresponding to the information to be detected according to the category of the target object; the server 104 extracts image texts in the images to be detected, and matches the image texts through a preset keyword library to obtain second pushing influence characteristics corresponding to the information to be detected; the server 104 inputs the text to be detected into a text classification model to be detected for text classification, a text class to be detected is obtained, and a third pushing influence characteristic corresponding to the information to be detected is obtained according to the text class to be detected; the server 104 determines a push detection result corresponding to the information to be detected based on the first push influence feature, the second push influence feature, and the third push influence feature. The terminal 102 may be, but not limited to, various personal computers, notebook computers, smartphones, tablet computers, and portable wearable devices, and the server 104 may be implemented by a stand-alone server or a server cluster composed of a plurality of servers.
In one embodiment, as shown in fig. 2, an information push detection method is provided, and the method is applied to the server in fig. 1 for illustration, it may be understood that the method may also be applied to a terminal, and in this embodiment, the method includes the following steps:
step 202, obtaining information to be detected, wherein the information to be detected comprises an image to be detected and a text to be detected.
The information to be detected refers to information to be pushed, for example, advertisement, instant messaging application friend circle message, news of a news platform, posting of a video website, social platform message and the like, and the news platform is a platform for publishing news, for example, the news platform can be a networknews platform, a hundred degree news platform, a UC news platform, various news websites and the like. The social platform refers to a platform for social contact, for example, the social platform may be a microblog application platform, a WeChat application platform, a video website platform, and the like. The information to be detected is multi-modal information, namely the information to be detected comprises information in a plurality of different forms. The image to be detected refers to an image included in the information to be detected, and the number of the image to be detected can be multiple, and the image to be detected can also be an image obtained from a video in the information to be detected. The text to be detected refers to text included in the information to be detected, and the text may be a title, a text, and the like in the information to be detected.
Specifically, the server acquires information to be detected, where the information to be detected may be information that needs to be pushed to each receiver by the sender, and the sender may be an object capable of pushing information, such as a user, an enterprise, a platform, and the like. The recipient may be an object for the user or the like to receive the information. The server may acquire the transmitted information to be detected from the sender terminal. The information to be detected can also be stored in a database. The server can acquire the information to be detected from a database, and the database can be a database of the server itself or a database in a third party server, and when the database is the database in the third party server, the server is stated to push and detect the information to be detected which needs to be pushed by the third party. In one embodiment, when the information to be detected is obtained, preprocessing is performed on an image to be detected and a text to be detected in the information to be detected, for example, converting the image to be detected into an image with a fixed size, performing word segmentation and duplication removal on the text to be detected, and the like. Obtaining a preset image to be detected and a preset text to be detected, and carrying out subsequent processing by using the preset image to be detected and the preset text to be detected.
Step 204, extracting a target object area in the image to be detected, identifying the category of the target object in the target object area, and determining a first pushing influence feature corresponding to the information to be detected according to the category of the target object.
The target object refers to an object in the image to be detected, the target object can be a plurality of objects, such as various advertisement goods, various animals, plants, human bodies and the like. The target object region refers to an image region in which a target object exists. The category of the target object refers to the category to which the target object belongs, for example, white spirit belongs to wine, and different objects have different categories. The category of the target object may also be a specific target object identification, may be a name, a number, etc. The push influencing feature refers to a feature influencing a push detection result, and comprises a push forward influencing feature and a push reverse influencing feature. The push forward influence feature is a feature that can influence the push detection result on push. The push reverse influence feature is a feature that can influence the push detection result to the prohibition push. The first push influence feature refers to a push influence feature obtained according to an image to be detected.
Specifically, the server may extract a target object region in the image to be detected using an image segmentation algorithm for segmenting regions of different image words and extracting a region of interest, and may include a threshold-based segmentation method, a region-based segmentation method, an edge-based segmentation method, a theory-specific segmentation method, a depth-learning-based segmentation, and the like. The server identifies the class of the target object in the target object region using a deep neural network algorithm, for example, using a CNN (Convolutional Neural Networks, convolutional neural network) algorithm. And comparing the category of the target object with the category of the preset pushing influence object, and obtaining a first pushing influence characteristic corresponding to the information to be detected according to the comparison result. When the category of the target object is consistent with the category of the preset pushing reverse influence object, the obtained first pushing influence characteristic is a pushing reverse influence characteristic, and when the category of the target object is inconsistent with the category of the preset pushing reverse influence object, the obtained first pushing influence characteristic is a pushing forward influence characteristic.
And 206, extracting an image text in the image to be detected, and matching the image text through a preset keyword library to obtain a second pushing influence characteristic corresponding to the information to be detected.
The image text refers to text contained in the image to be detected, and the text can be explanation of the image to be detected. The preset keyword library refers to a preset keyword database for influencing pushing, and the preset keyword library can be a keyword library for influencing pushing forward or a keyword library for influencing pushing backward. The second push influence feature refers to a push influence feature obtained according to the image text.
Specifically, the server may extract image text into the image to be detected using an image-text recognition algorithm, wherein the image-text recognition algorithm may be an OCR (Optical Character Recognition ) algorithm, a Chinese-character-location recognition algorithm in the image, a deep neural network algorithm, and the like. And then word segmentation is carried out on the image text to obtain word segmentation results, word segmentation results are searched in a preset keyword library, a matching result is determined according to the search results, and second pushing influence characteristics corresponding to the information to be detected are obtained according to the matching result.
And step 208, inputting the text to be detected into a text classification model to be detected for text classification, obtaining the text category to be detected, and obtaining a third pushing influence characteristic corresponding to the information to be detected according to the text category to be detected.
The text classification model to be detected is used for identifying push influence categories corresponding to the text to be detected, wherein the push influence categories comprise push forward influence categories and push reverse influence categories. And obtaining the forward impact characteristics according to the forward impact category, and obtaining the backward impact characteristics according to the backward impact category. The text classification model to be detected is obtained by training by using a deep neural network algorithm. The deep neural network algorithm may be a CNN algorithm, among others.
Specifically, the server inputs the text to be detected into a text classification model to be detected for text classification, namely an input layer of the text classification model to be detected is converted into a vector, the vector is input into a convolution layer of the text classification model to be detected for convolution calculation, a convolution calculation result is input into a pooling layer of the text classification model to be detected for pooling, and then the pooling result is input into a classification output layer for classification in a fully connected mode and classification results are output, so that the text category to be detected is obtained. And then obtaining a third pushing influence characteristic corresponding to the information to be detected directly according to the text category to be detected.
Step 210, determining a push detection result corresponding to the information to be detected based on the first push influence feature, the second push influence feature and the third push influence feature.
The pushing detection result refers to a detection result of whether the information to be detected is pushed or not.
Specifically, the server may input the first pushing influence feature, the second pushing influence feature and the third pushing influence feature into the feature fusion model to identify whether pushing is performed, so as to obtain a result output by the feature fusion model, and thus obtain the feature fusion model. The feature fusion model is a model obtained by training by using a machine learning algorithm according to historical data, the historical data comprises historical to-be-detected information and corresponding historical push detection results, a first push influence feature, a second push influence feature and a third push influence feature are obtained according to the historical to-be-detected information, then the machine learning algorithm is used for training, and when training is completed, the feature fusion model is obtained. The machine learning algorithm may be a linear regression algorithm, a random forest algorithm, a neural network algorithm, and the like. The server can also directly obtain a push detection result corresponding to the information to be detected according to the first push influence feature, the second push influence feature and the third push influence feature, determine the number of push reverse influence features and the number of push forward influence features from the first push influence feature, the second push influence feature and the third push influence feature, and determine the push detection result corresponding to the information to be detected according to the number of push reverse influence features and the number of push forward influence features.
According to the information push detection method, the first push influence characteristics are determined according to the categories of the target objects in the target object area of the image to be detected, the second push influence characteristics are obtained according to the image texts in the image to be detected through the preset keyword library, meanwhile, the text to be detected is subjected to text classification, the third push influence characteristics are obtained according to the categories of the text to be detected, then the push detection results corresponding to the information to be detected are determined according to the first push influence characteristics, the second push influence characteristics and the third push influence characteristics, namely, the corresponding push influence characteristics are obtained according to different types of data respectively, then the push detection results corresponding to the information to be detected are determined according to different push influence characteristics, and the information push detection efficiency and the information push detection accuracy are improved.
In one embodiment, as shown in fig. 3, step 204, extracting a target object area in the image to be detected, identifying a category of a target object in the target object area, and determining a first pushing influence feature corresponding to the information to be detected according to the category of the target object includes:
in step 302, the image to be detected is input into a target object detection model, which includes a target object area detection network and a target object identification network.
The target object detection model is used for identifying the category of a target object in an image to be detected, and is obtained by training a deep neural network according to the training image and the category of the corresponding training object. The target object region detection network is used for identifying the region of interest from the image to be detected, is a part of networks in the target object detection model, and can be obtained by training alone. The term of the target object recognition network recognizes the category of the target object in the region of interest in the image to be detected, the target object recognition network can be a partial network of the target object detection model red, and the target object recognition network can also be obtained by independent training.
Specifically, the server inputs the image to be detected into the target object detection model for detection, namely, the category of the target object is obtained after passing through the target object area detection network and the target object identification network.
And step 304, detecting the image to be detected according to the target object area detection network to obtain a target object area.
The target object area detection network is trained according to a deep neural network for target detection, wherein the deep neural network for target detection can be a SSD (Single Shot MultiBox Detector) network, a cascade-RCNN (one target detection algorithm) network, a YOLO (one target detection algorithm) network, a fast RCNN (one target detection algorithm) network and the like. The target object area is an area where a target object exists, and there may be a plurality of target object areas or only one target object area.
Specifically, the server inputs an image to be detected into a target object area detection network to detect an area of interest, and a target object area is obtained. The detection is performed by using a cascades-rcnn network, so that the accuracy of positioning the region of interest can be improved, and the performance of the whole model is further improved.
And step 306, inputting the target object area into a target object identification network for identification, and obtaining the category of the target object in the target object area.
The target object recognition network is trained according to a deep neural network for object recognition. Among these, the deep neural network for object recognition may be a res net (residual) network, an RCNN network, a Fast RCNN network, and the like.
Specifically, the server inputs the target object area into a target object identification network to identify, and obtains the category of the target object in the target object area. When a plurality of target object areas exist, the plurality of target object areas are input into a target object recognition network for recognition, and the target object category corresponding to each target object area is obtained. In one embodiment, the target object recognition network is only used for recognizing the types of the target objects which are forbidden to be pushed, so that the number of recognized types is reduced, and the recognition efficiency and accuracy are improved.
In a specific embodiment, as shown in fig. 4, a network structure diagram for detecting an image to be detected is shown, specifically: the image to be detected comprises an image text and a region of a target object, the detected image is input into an SSD network in a target object detection model, and a sensing interest region, namely a region of the target object, is obtained from the image to be detected. And inputting the region of the target object into a target object detection model ResNet network, wherein the ResNet network introduces a CBAM module to obtain the target object category in the input target object region. For example, as shown in fig. 5, a schematic diagram of detecting an image in an alcoholic beverage advertisement is shown, where when an image to be detected in the advertisement to be detected is detected by using a target object detection model, an area where a wine bottle is located is detected by using a target object area detection network, and the area where the wine bottle is located is identified, so that an identification result is an alcoholic beverage.
Step 308, matching the category of the target object with the category of the preset push reverse influence object, and obtaining a first push influence feature corresponding to the information to be detected as a first push reverse influence feature when the matching is consistent.
Step 310, when the matching is inconsistent, obtaining the first push influence feature corresponding to the information to be detected as the first push forward influence feature.
Wherein the preset pushing reverse influence object class refers to a preset object class which prohibits pushing, for example, objects such as wine, watch, adult product and the like in the advertisement image are prohibited from being pushed,
specifically, the server matches all obtained target object categories with preset push reverse influence object categories, and when preset push reverse influence object categories consistent in matching exist, first push influence features corresponding to information to be detected are obtained as first push reverse influence features. When the preset pushing reverse influence object categories which are consistent in matching do not exist, the obtained first pushing influence characteristics corresponding to the information to be detected are first pushing forward influence characteristics, namely, when all the categories of the target objects are not pushing reverse influence object categories, the obtained first pushing influence characteristics corresponding to the information to be detected are first pushing forward influence characteristics.
In the above embodiment, by using the target object detection model to detect the class of the corresponding target object in the image to be detected, accuracy and efficiency of obtaining the class of the target object are improved, and then the class of the target object is matched with the preset push reverse influence object class, so as to obtain the first push influence feature, and accuracy and efficiency of the obtained first push influence feature are improved.
In one embodiment, as shown in fig. 6, step 306, inputting the target object area into the target object recognition network to recognize, and obtaining the category of the target object in the target object area includes:
at step 602, the regional, channel and spatial attention features of the target object region are computed by the target object recognition network.
The region features refer to the extracted features of the target object region, the channel attention features refer to features obtained according to the relation among channels of the region features, the spatial attention features refer to features obtained according to the spatial relation of the region features, and the channel attention features are complemented.
Specifically, the server calculates, through the target object recognition network, region characteristics corresponding to the target object region using the region parameters, calculates channel attention characteristics using the channel attention parameters in the attention mechanism module, and calculates spatial attention characteristics using the spatial attention parameters in the attention mechanism module.
Step 604, combining the channel attention feature, the space attention feature and the region feature, performing convolution calculation according to the combined features, and performing classification recognition according to the convolution calculation result to determine the category of the target object in the target object region.
The combination means that the features are combined together, the features can be directly spliced, and the combined features can be obtained after the operation between the features, such as addition operation, dot multiplication operation, product operation and the like.
Specifically, the server splices the channel attention feature, the space attention feature and the region feature to obtain spliced features, inputs the spliced features into a convolution layer to carry out convolution operation, carries out classification and identification through an output layer according to a convolution calculation result to obtain output probability, and determines the category of the target object in the target object region according to the output probability.
In a specific embodiment, the attention mechanism module, that is, the CBAM (Convolutional Block Attention Module, convolution block attention module) module is included in the target object recognition network, as shown in fig. 7, which is a schematic structural diagram of the attention mechanism, specifically: the output feature of the previous convolution block is obtained as an input feature F, the channel attention feature F ' is calculated by using the input feature F and the channel attention Mc, and the spatial attention feature F ' is calculated by using the channel attention feature F ' and the spatial attention Ms. And then obtaining the output characteristic of the attention memorizing module according to the space attention characteristic F', and taking the output characteristic as the input characteristic of the next convolution block to continue calculation.
In the above embodiment, the expression capability of the target object region is improved by calculating the region feature, the channel attention feature and the spatial attention feature, and then the category of the target object is calculated by using the region feature, the channel attention feature and the spatial attention feature, so that the obtained category of the target object is more accurate.
In one embodiment, as shown in fig. 8, the training of the target object area detection network includes the steps of:
step 802, a training image of a labeled target object region is acquired.
Specifically, the server acquires a training image of a labeled target object area, wherein the labeled target object area is a target object in the index-injected training image, and the target object can be an object which is forbidden to push or can be pushed. The server can collect unlabeled images in advance and then label the unlabeled images to obtain training images of the labeled target object areas, wherein the server can collect the unlabeled images from the Internet to label, can obtain the unlabeled images from a server database to label, and can obtain the training images of the labeled target object areas from a third-party data platform which is used for providing the labeled training images.
Step 804, inputting the training image of the marked target object region into the initial target object region detection network for detection, and obtaining the initial target object region.
The initial target object area detection network refers to a target object area detection network initialized by network parameters. And the initial target object area is a target object area obtained by carrying out area extraction by using the initialized network parameters.
Specifically, the server inputs the training image marked with the target object area into an initial target object area detection network for detection, namely vectorizes the training image, and divides the image area of the vectorized training image by using initialized network parameters to obtain an initial target object area.
Step 806, calculating the region error information of the initial target object region and the labeled target object region, and updating the network parameters in the initial target object region detection network according to the region error information.
And step 808, obtaining the trained target object area detection network until the area error information obtained by training meets the preset first training completion condition.
The region error information is used for representing the error between the target object region obtained through training and the marked target object region. The preset first training completion condition refers to a preset condition that the target object region detection network training is completed, and may include that the region error information is smaller than a preset threshold value or the training iteration number exceeds a preset number.
Specifically, the server calculates the region error information of the initial target object region and the labeled target object region using a preset loss function, wherein the preset loss function may be a cross entropy loss function. And judging whether the region error information accords with a preset first training completion condition, wherein the preset first training completion condition can be whether the error information is smaller than a preset threshold value. The preset first training completion condition may also be whether the number of training iterations exceeds a preset number. When the region error information does not meet the preset first training completion condition, the back propagation is performed by using a back propagation algorithm according to the region error information, that is, updating network parameters in the initial target object region detection network, and the back propagation can also be performed by using a gradient descent algorithm. And obtaining the target object area detection network with updated network parameters. And taking the target object area detection network with updated network parameters as an initial target object area detection network, returning to step 802, and continuing to perform iterative execution until the training-obtained area error information accords with a preset first training completion condition, indicating that the training of the target object area detection network is completed, and obtaining the trained target object area detection network.
In one embodiment, as shown in FIG. 9, training of the target object recognition network includes the steps of:
step 902, obtaining a training image of the marked target object region, extracting the marked target object region from the training image, and obtaining a target object category label corresponding to the marked target object region.
The target object class label refers to a real target object class.
Specifically, the server acquires the training image of the marked target object region, extracts the marked target object region by using the trained target object region detection network, and can also directly cut the training image of the marked target object region to obtain the marked target object region. And simultaneously, the target object class label corresponding to the marked target object area is obtained, and the target object class label can be set when the target object area is marked. Or the target object area is obtained from a database with guaranteed category labels according to the marked target object area.
And 904, inputting the marked target object area into an initial target object recognition network for classification, obtaining the output training target object class probability, and obtaining the training target object class according to the training target object class probability.
The initial target object recognition network refers to a target object recognition network initialized by network parameters, and the training target object class probability refers to the probability of the target object class obtained through training.
Specifically, the server inputs the marked target object area into an initial target object recognition network to perform classification recognition, namely vectorizing the marked target object area, calculating according to the vectorized marked target object area and initialized network parameters to obtain each target object class probability, and obtaining a target object class corresponding to the maximum training target object class probability as a training target object class.
Step 906, calculating class error information of the target object class label and the training target object class, and updating network parameters in the initial target object identification network based on the class error information.
Step 908, obtaining the trained target object recognition network until the class error information obtained by training reaches the preset second training completion condition.
The class error information refers to the error between the true target object class and the target object class obtained through training. The preset second training completion condition refers to a preset condition for completing training of the target object recognition network, and may be whether the category error information is smaller than a preset category error threshold value or whether the training frequency reaches the maximum training iteration frequency.
Specifically, the server calculates class error information of the target object class label and the training target object class by using a preset loss function, wherein the loss function can be a cross entropy loss function. The server judges whether the category error information reaches a preset second training completion condition, when the category error information does not reach the preset second training completion condition, the counter-propagation algorithm is used for counter-propagation according to the category error information, namely, the network parameters in the initial target object recognition network are updated, the target object recognition network after the network parameters are updated is obtained, the target object recognition network after the network parameters are updated is used as the initial target object recognition network, the iteration execution is continued in step 902, until the category error information obtained by training reaches the preset second training completion condition, the fact that the category error information obtained by training is smaller than a preset category error threshold value or the training times reach the maximum training iteration times is indicated, and the trained target object recognition network is obtained.
In the above embodiment, the target object area detection network and the target object recognition network in the target object detection model are respectively trained, so that optimization can be performed in a directed manner, for example, when the target object area detection network performs poorly, the target object area detection network can be optimized independently, or when the target object recognition network performs poorly, the target object recognition network can be optimized independently, thereby improving the accuracy of training.
In one embodiment, as shown in fig. 10, step 206, extracting an image text in the image to be detected, and matching the image text through a preset keyword library to obtain a second pushing influence feature corresponding to the information to be detected, which includes:
step 1002, extracting image text in an image to be detected through an image text recognition algorithm, and segmenting the image text to obtain each word to be detected.
Wherein the image text recognition algorithm is used for recognizing text in the image, the image text recognition algorithm can comprise
Specifically, the server extracts a feature sequence in an image to be detected through an image text recognition algorithm, acquires real text distribution corresponding to the feature sequence, and converts operations such as removing integration and the like of the real text distribution corresponding to the feature sequence into each word to be detected. Among them, the image text recognition algorithm includes CRNN OCR (convolutional cyclic neural network optical character recognition) algorithm and attention OCR (attention optical character recognition) algorithm.
Step 1004, matching each word to be detected with the push reverse influence keywords in the preset keyword library.
The preset keyword library is a database for storing and transmitting reverse influence keywords. The pushing reverse influence keyword refers to a keyword with a reverse influence on whether the information to be detected is pushed, namely, a keyword which is forbidden to be pushed.
Specifically, the server searches each word to be detected in push reverse influence keywords in a preset keyword library, and when the push reverse influence keywords consistent with the word to be detected are searched, a matched consistent result is obtained. And when all the words to be detected are not found in the preset keyword library, obtaining a result of inconsistent matching.
Step 1006, when the matching is consistent, obtaining a second push influence feature corresponding to the information to be detected as a second push reverse influence feature.
And step 1008, when the matching is inconsistent, obtaining a second push influence characteristic corresponding to the information to be detected as a second push forward influence characteristic.
The second pushing reverse influence feature refers to a pushing influence feature corresponding to the fact that pushing of the word to be detected is forbidden in the image text. The second push forward influence feature refers to a push influence feature corresponding to the fact that the image text does not contain the word to be detected in the push prohibition mode.
Specifically, when the matching is consistent, the second push influence characteristic corresponding to the information to be detected is obtained as a second push reverse influence characteristic. And when the matching is inconsistent, obtaining a second push influence characteristic corresponding to the information to be detected as a second push forward influence characteristic.
In one particular embodiment, as shown in FIG. 11, a schematic diagram of a framework for image text recognition is provided. The method comprises the steps of identifying an alcohol advertisement image through an OCR text recognition algorithm to obtain each text word, wherein each text word comprises a brewer, a national good and a top-quality product, identifying the category of each text word through a preset keyword library to obtain a keyword which is reversely influenced by the brewer for pushing, and obtaining a second pushing influence characteristic which is a second pushing reverse influence characteristic.
In the embodiment, whether the pushing reverse influence keywords exist in the image text is detected through the preset keyword library, so that the second pushing influence characteristics are obtained, and the efficiency of obtaining the second pushing influence characteristics is improved.
In one embodiment, as shown in fig. 12, the establishment of the preset keyword library includes the following steps:
step 1202, acquiring a pushing reverse influence target object identifier, and acquiring a corresponding pushing reverse influence target object text according to the pushing reverse influence target object identifier;
the pushing reverse influence target object refers to a target object with a reverse influence on whether information to be detected is pushed or not, namely a target object which is forbidden to be pushed, a pushing reverse influence target object identifier is used for identifying, and the pushing reverse influence target object is used for identifying. The pushing of the target object text with reverse influence refers to whether the information to be detected is pushed or not, namely the target object text which is forbidden to be pushed.
Specifically, the server acquires the pushing reverse influence target object identifier, and according to the pushing reverse influence target object identifier, the corresponding pushing reverse influence target object text can be acquired from the internet, or the stored pushing reverse influence target object text can be directly acquired from the database. The number of the pushing reverse influence target object identifiers can be multiple, and each pushing reverse influence target object identifier obtains multiple pushing reverse influence target object texts.
And 1204, word segmentation is carried out on the text of the target object which is reversely influenced by pushing, so as to obtain each text word.
Specifically, word segmentation is a process of recombining a continuous word sequence into a word sequence according to a certain specification. The server divides words of each obtained pushing reverse influence target object text to obtain each text word corresponding to each pushing reverse influence target object text.
In step 1206, word frequency and reverse document frequency corresponding to each text word are calculated, and importance degree of each text word is calculated based on the word frequency and reverse document frequency corresponding to each text word.
The word frequency refers to the frequency of occurrence of text words in the text of the target object which is reversely influenced by pushing. Reverse document frequency refers to the frequency with which text words appear in all pushing reverse impact target object text.
Specifically, the server calculates word frequency and reverse file frequency corresponding to each text word, and then calculates the product of the word frequency and the reverse file frequency corresponding to each text word to obtain the importance degree of each text word.
Step 1208, selecting a target number of text words as push reverse influence keywords according to the importance degree of each text word, and storing the push reverse influence keywords in a keyword library to obtain a preset keyword library.
Specifically, the server sorts the text words according to the importance degree, and sequentially selects the text words with the target number from large to small according to the importance degree as push reverse influence keywords according to the sorting result. The push reverse influence keywords corresponding to the push reverse influence target object identifiers are obtained, and the push reverse influence keywords corresponding to the push reverse influence target object identifiers are stored in a keyword database to obtain a preset keyword library.
In the above embodiment, the importance degree of each text word is obtained by calculating the word frequency and the reverse file frequency corresponding to each text word, and the push reverse influence keyword corresponding to the push reverse influence target object identifier is obtained according to the importance degree of each text word, so that the obtained push reverse influence keyword is more accurate. And then establishing a keyword library, so that the subsequent use is convenient.
In one embodiment, as shown in fig. 13, step 1206, calculating word frequencies and reverse file frequencies corresponding to each text word includes:
step 1302, counting the total number of words corresponding to the text of the push reverse influence target object, and counting the occurrence times of each text word in the text of the push reverse influence target object.
In step 1304, a ratio of the number of occurrences corresponding to each text word to the total number of words is calculated, and a word frequency corresponding to each text word is obtained.
Wherein, the total number of words refers to the number of text words contained in the text of the push reverse influence target object.
Specifically, the server counts the total number of words of each text word corresponding to the pushing reverse influence target object text, counts the occurrence times of each text word in the pushing reverse influence target object text, calculates the ratio of the occurrence times corresponding to each text word to the total number of words, and obtains the word frequency corresponding to each text word. The word frequency corresponding to each text word may be calculated using the following formula (1).
Wherein TF is w Refers to the word frequency corresponding to the text word.
In step 1306, the total number of texts corresponding to the reverse pushing target object texts is counted, and the target text numbers corresponding to the text words are counted, wherein the target text numbers refer to the number of texts containing target text words in the reverse pushing target object texts, and the target text words are selected from the text words.
Specifically, the server counts the total number of the collected push reverse influence target object texts corresponding to the push reverse influence target object identifiers. Then, each text word is taken as a target text word, and the number of texts containing the target text word is calculated.
Step 1308, calculating the ratio of the target text number corresponding to each text word to the total text number, and calculating the logarithm of the ratio to obtain the reverse file frequency corresponding to each text word.
Specifically, the server calculates the ratio of the target text number corresponding to each text word to the total text number, and then calculates the logarithm of the ratio to obtain the reverse file frequency corresponding to each text word. Wherein, the reverse file frequency corresponding to the text word can be calculated using formula (2).
The IDF refers to the reverse document frequency corresponding to the text word, and the denominator is added with 1 to avoid that the denominator is 0.
In one embodiment, training of the text classification model to be detected comprises the steps of:
acquiring a training text to be detected and a corresponding training text category to be detected; and training the training text to be detected as input, training the training text category to be detected as a label by using a convolutional neural network, and obtaining a text classification model to be detected when the training is completed.
The training text class to be detected refers to a real text class to be detected during training, and the text class to be detected comprises a push class and a push prohibition class. Among these, convolutional neural networks (Convolutional Neural Networks, CNN) are a type of feed-forward neural network that includes convolutional calculations and has a deep structure. The convolutional neural network includes an input layer, a convolutional layer, a pooling layer, a full-connectivity and a softmax layer.
Specifically, the server acquires a training text to be detected and a corresponding training text category to be detected, wherein the training text to be detected is acquired from the internet, and the training text category to be detected is a category obtained after category labeling is carried out on the training text to be detected. And then the server takes the training text to be detected as input, and takes the training text category to be detected as a label to train by using a convolutional neural network, wherein an activation function used for training can be a Relu activation function, an S-type activation function or a tanh activation function. And when training is finished, obtaining a text classification model to be detected, wherein the training is finished, namely that the value of the loss function reaches a preset threshold value or the training reaches the maximum iteration number. The loss function may use a cross entropy loss function, may be a square error loss function, may be a classification loss function and a KL divergence loss function, and the like.
In a specific embodiment, as shown in fig. 14, a schematic structural diagram of a convolutional neural network is provided, which includes an input layer, a convolutional layer, a pooling layer, a full-link layer, and a softmax layer. And (3) carrying out text classification on the 'brewer, the national goods and the fine products', and obtaining the representation of the text through an input layer to obtain a vector matrix of 3x 100. Then inputting the characteristics of the text into a convolution layer for convolution to obtain the characteristics of the text, pooling the characteristics of the text through a pooling layer, namely extracting the characteristics corresponding to the maximum value from the characteristics to serve as pooling results, classifying the pooling results through a full connection and softmax layer to obtain the probability of being the push type and the probability of prohibiting the push type, obtaining the type of training output, calculating a loss value by using a loss function according to the type of training output, the push type and the prohibited push type, updating parameters in a network according to the loss value, completing one round of training, and completing training when the loss value is smaller than a preset threshold value to obtain a text classification model to be detected.
In the embodiment, the text to be detected and the corresponding text category to be detected are trained to obtain the text classification model to be detected, and then the text classification model to be detected can be deployed, so that the subsequent use is convenient.
In one embodiment, step 210, determining a push detection result corresponding to the information to be detected based on the first push influence feature, the second push influence feature, and the third push influence feature, includes the steps of:
and combining the first pushing influence feature, the second pushing influence feature and the third pushing influence feature to obtain combined features, inputting the combined features into a pushing influence detection model to carry out pushing influence detection, and obtaining a pushing detection result corresponding to the information to be detected.
The combination feature is a feature obtained by combining the first push influence feature, the second push influence feature and the third push influence feature, wherein the combination may be performed by splicing, performing vector operation, and the like, and the vector operation may be performed by adding, performing number multiplication, performing dot product, and the like. The push influence detection model is trained according to the historical push influence characteristic data and the historical push detection result by using a machine learning algorithm, wherein the machine learning algorithm can be a linear regression algorithm, a neural network algorithm, a random forest algorithm and the like. The push detection result refers to a detection result of whether to push the information to be detected, and includes push and push prohibition.
Specifically, the server can directly splice the first pushing influence feature, the second pushing influence feature and the third pushing influence feature to obtain a combined feature, input the combined investment into the pushing influence detection model to carry out pushing influence detection, and obtain an output pushing detection result, namely a pushing detection result corresponding to the information to be detected.
In the embodiment, the pushing influence detection model is used for pushing influence detection, so that the pushing detection result corresponding to the information to be detected is obtained, and the accuracy of obtaining the pushing detection result is improved.
In one embodiment, the push influencing features include a push forward influencing feature and a push reverse influencing feature. As shown in fig. 15, step 210 of determining a push detection result corresponding to the information to be detected based on the first push influence feature, the second push influence feature, and the third push influence feature includes:
in step 1502, the first push influencing feature, the second push influencing feature, and the third push influencing feature are matched with the push reverse influencing feature, and the number of matched features is counted.
Specifically, the server judges push reverse influence features in the first push influence features, the second push influence features and the third push influence features, namely the push reverse influence features are respectively matched with the first push influence features, the second push influence features and the third push influence features, namely whether the push reverse influence features are identical with the first push influence features, the second push influence features and the third push influence features or not. And counting the number of matched and consistent features, wherein the number of features refers to the number of the first push influence feature, the second push influence feature and the third push influence feature which are consistent with the push reverse influence feature, for example, when the first push influence feature is the push reverse influence feature, the second push influence feature is the push reverse influence feature, and the third push influence feature is the push forward influence feature, the push reverse influence feature is consistent with the first push influence feature and the second push influence feature, and the counted number of features is 2.
In step 1504, when the feature number exceeds the preset number, the push detection result corresponding to the information to be detected is obtained as push prohibition, and when the feature number does not exceed the preset number, the push detection result corresponding to the information to be detected is obtained as push.
The preset number may be a preset number threshold for existence of the push reverse impact feature. The push prohibition refers to the information to be checked, and the push prohibition refers to the information to be checked. The pushing means that the information to be detected is information capable of pushing.
Specifically, the feature quantity is compared with the preset quantity, when the feature quantity is the comparison result, the pushing detection result corresponding to the information to be detected is obtained to be the pushing prohibition, and when the feature quantity is not the comparison result, the pushing detection result corresponding to the information to be detected is obtained to be the pushing. In one embodiment, the server may further count the number of features that are inconsistent in matching, and when the number of features that are consistent in matching is greater than the number of features that are inconsistent in matching, obtain a push detection result corresponding to the information to be detected as push prohibition, and when the number of features that are consistent in matching is less than the number of features that are inconsistent in matching, obtain a push detection result corresponding to the information to be detected as push.
In the embodiment, the feature quantity matched with the pushing reverse influence feature is counted and compared with the preset quantity to obtain the pushing detection result corresponding to the information to be detected, so that the efficiency of obtaining the pushing detection result is improved.
In one embodiment, after step 210, after determining a push detection result corresponding to the information to be detected based on the first push influence feature, the second push influence feature, and the third push influence feature, the method further includes the steps of:
when the push detection result is push, each receiver address is obtained, and the information to be detected is sent to the receiver terminal according to each receiver address, so that the receiver terminal displays the information to be detected.
When the push detection result is that push is forbidden, generating alarm information, acquiring a sender address, and returning the alarm information to a sender terminal corresponding to the sender address.
The receiving party refers to an object for receiving the pushed information. The receiver address is used to indicate the address of the receiver terminal, and may be an IP (internet protocol) address, a MAC (physical) address, and the like. The alarm information is used for prompting that the information to be detected is the information of forbidding pushing. The sender refers to pushing the object of the information to be detected. The sender address is used to indicate the address of the sender terminal, and may be an IP address, a MAC address, and the like.
Specifically, when the server obtains the detection result of the information to be detected, performing subsequent processing according to the detection result, that is, when the push detection result is push, obtaining each receiver address, where each receiver address is a preset address where the information to be detected is to be pushed. And sending the information to be detected to the corresponding receiver terminal according to each receiver address. And when the receiving party terminal receives the information to be detected, displaying the information to be detected. When the pushing detection result is that pushing is forbidden, generating alarm information, acquiring a sender address, returning the alarm information to a sender terminal corresponding to the sender address, and pushing after the sender terminal receives the alarm information and re-modifying the information to be detected according to the alarm information. In one embodiment, when the detection result of the file pushing is that pushing is prohibited, a prompt for the presence of forbidden information is given to the management terminal, so that the management terminal further examines the information to be detected corresponding to the pushing prohibition, a further detection result is obtained, and the accuracy of the detection result is improved.
In the above embodiment, when the push detection result is push, the information to be detected is pushed, and when the push detection result is push prohibition, the alarm information is generated, so that the information to be detected can be processed later according to the push detection result.
In a specific embodiment, as shown in fig. 16, the information push detection method specifically includes the following steps:
in step 1602, information to be detected is obtained, the information to be detected including an image to be detected and a text to be detected.
In step 1604, the image to be detected is input into a target object detection model, where the target object detection model includes a target object area detection network and a target object recognition network, and the image to be detected is detected according to the target object area detection network to obtain a target object area.
In step 1606, the regional feature, the channel attention feature and the spatial attention feature of the target object region are calculated through the target object recognition network, the channel attention feature, the spatial attention feature and the regional feature are combined, the convolution calculation is performed according to the combined features, and the classification recognition is performed according to the convolution calculation result to determine the category of the target object in the target object region.
In step 1608, the category of the target object is matched with the category of the preset push reverse influence object, when the matching is consistent, the first push influence feature corresponding to the information to be detected is obtained as a first push reverse influence feature, and when the matching is inconsistent, the first push influence feature corresponding to the information to be detected is obtained as a first push forward influence feature.
In step 1610, the image text in the image to be detected is extracted by the image text recognition algorithm, and the image text is segmented to obtain each word to be detected.
Step 1612, matching each word to be detected with a push reverse influence keyword in a preset keyword library, and obtaining a second push reverse influence feature corresponding to the information to be detected as a second push reverse influence feature when the matching is consistent; and when the matching is inconsistent, obtaining a second push influence characteristic corresponding to the information to be detected as a second push forward influence characteristic.
Step 1614, inputting the text to be detected into the text classification model to be detected for text classification, so as to obtain a text category to be detected, when the text category to be detected is a push category, obtaining a third push influence feature corresponding to the information to be detected as a third push forward influence feature, and when the text category to be detected is a push prohibition category, obtaining a third push influence feature corresponding to the information to be detected as a third push reverse influence feature.
In step 1616, the first pushing influence feature, the second pushing influence feature, and the third pushing influence feature are matched with the pushing reverse influence feature, the feature quantity matched with the first pushing influence feature, the second pushing influence feature, the third pushing influence feature and the pushing reverse influence feature are counted, when the feature quantity exceeds the preset quantity, the pushing detection result corresponding to the information to be detected is obtained as the forbidden pushing, and when the feature quantity does not exceed the preset quantity, the pushing detection result corresponding to the information to be detected is obtained as the pushing.
In a specific embodiment, as shown in fig. 17, a frame schematic diagram of an information push detection method is provided, specifically:
the server pushes the wine advertisement to each user terminal, at this time, the server acquires wine advertisement information, the wine advertisement information comprises a wine advertisement title text and a wine advertisement video, a wine advertisement image is intercepted from the wine advertisement video, the wine advertisement image is identified through OCR text to obtain an image text, the image text is segmented to obtain each text word, and the judgment of whether the image text corresponding to the commodity is forbidden to be pushed or not is carried out through a preset keyword library by using each text word, and the image text contains a text related to wine, so that a detection result that the image text is forbidden to be pushed is obtained. And then inputting the wine advertisement image into a trained SSD network to obtain a region of interest, namely a wine bottle region, inputting the wine bottle region into a trained ResNet network which is introduced into the CBAM module to obtain that the type of the wine bottle in the output wine bottle region is wine, wherein the wine is a commodity which is forbidden to be pushed, and obtaining a detection result which is forbidden to be pushed. The wine advertisement title text is then input into a CNN text classification network, and push-forbidden results are obtained through an input layer, a convolution layer, a pooling layer, a full connection layer and a softmax layer. And finally, obtaining that the wine advertisement information is the information of prohibiting pushing according to the result that all three detection results are the information of prohibiting pushing, and generating alarm information to prompt that pushing is prohibited.
The application also provides an application scene, which applies the information push detection method. Specifically, the application of the information push detection method in the application scene is as follows:
the information push detection method is applied to a news website, the news website is used for sending a piece of news, a news website server is used for obtaining the news to be sent, news documents and news pictures are included in the news, a news object area in the news pictures is extracted, the category of news objects in the news object area is identified, and first push influence characteristics corresponding to the news are determined according to the category of the news objects; extracting picture texts in news pictures, and matching the picture texts through a preset keyword library to obtain second pushing influence features corresponding to news; inputting the news document into a text classification model to be detected for text classification, obtaining a news document category, and obtaining a third pushing influence characteristic corresponding to news according to the news document category; and determining a push detection result corresponding to the news based on the first push influence feature, the second push influence feature and the third push influence feature. When the pushing detection result is pushing, pushing can be performed to the user, namely, when the user enters the news website, the news is pushed to the user terminal, so that the user terminal displays the news on the news website page. When the news has the pictures and the words with the blood smell violence, and the pushing detection result is that pushing is forbidden, prompt information such as the pictures and the words with the blood smell violence in the news are generated at the moment, prompt is carried out on the management terminal, and the news is pushed after being modified through the management terminal.
The application further provides an application scene, and the application scene applies the information push detection method. Specifically, the application of the information push detection method in the application scene is as follows:
the information push detection method is applied to a WeChat application platform, as shown in fig. 18, and is a schematic diagram of a framework for push detection of the WeChat application platform, specifically:
the WeChat application platform obtains information to be detected, wherein the information to be detected can be information issued by WeChat public numbers, can be advertisements in WeChat and the like. And carrying out data preprocessing on the information to be detected through a data processing server, namely preprocessing the text in the information to be detected, and preprocessing the image in the information to be detected. And then detecting through an intelligent auditing system, and inputting an image to be detected into a target object detection model for detection to obtain a first pushing influence characteristic. And (3) using ORC to identify and obtain an image text, segmenting the image text to obtain each word to be detected, and matching each word to be detected with push reverse influence keywords in a preset keyword library to obtain a second push influence feature. Inputting the text to be detected into a text classification model to be detected for text classification, obtaining the text category to be detected, and obtaining a third pushing influence characteristic corresponding to the information to be detected according to the text category to be detected. And determining a push detection result corresponding to the information to be detected based on the first push influence feature, the second push influence feature and the third push influence feature. And when the pushing detection result is pushing, pushing is performed. When the pushing detection result is that pushing is forbidden, the information to be detected, which is forbidden to be pushed, is sent to an auditing system for further auditing, namely, manual auditing is carried out, or detection is carried out through an inspection system. And then the audit record information is stored through the audit snapshot server. And then training the offline model by the stored audit record information, wherein the training comprises an image capability model and a text capability model, and training to obtain a deep learning model. Detection is then performed by a deep learning model. Namely, a self-learning cycle is performed to detect whether the information to be detected is pushed or not. The stored audit record information can be used for data analysis through a data analysis system, for example, the reasons for prohibiting pushing are analyzed, the duty ratio for prohibiting pushing is analyzed, and the like.
It should be understood that, although the steps in the flowcharts of fig. 2, 3, 6, 8-10, 12, 13, 15 and 16 are shown in order as indicated by the arrows, these steps are not necessarily performed in order as indicated by the arrows. The steps are not strictly limited to the order of execution unless explicitly recited herein, and the steps may be executed in other orders. Moreover, at least a portion of the steps of fig. 2, 3, 6, 8-10, 12, 13, 15, and 16 may include a plurality of steps or stages, which are not necessarily performed at the same time, but may be performed at different times, and the order of the steps or stages may not necessarily be sequential, but may be performed in rotation or alternatively with other steps or at least a portion of the steps or stages in other steps.
In one embodiment, as shown in fig. 19, an information push detection apparatus 1900 is provided, which may be a software module or a hardware module, or a combination of both, forming part of a computer device, and specifically includes: an information acquisition module 1902, a first feature determination module 1904, a second feature derivation module 1906, a third feature derivation module 1908, and a result determination module 1910, wherein:
The information acquisition module 1902 is configured to acquire information to be detected, where the information to be detected includes an image to be detected and a text to be detected;
a first feature determining module 1904, configured to extract a target object area in the image to be detected, identify a class of a target object in the target object area, and determine a first pushing influence feature corresponding to the information to be detected according to the class of the target object;
a second feature obtaining module 1906, configured to extract an image text in the image to be detected, and match the image text with a preset keyword library to obtain a second pushing influence feature corresponding to the information to be detected;
a third feature obtaining module 1908, configured to input a text to be detected into the text classification model to perform text classification, obtain a text category to be detected, and obtain a third pushing influence feature corresponding to the information to be detected according to the text category to be detected;
the result determining module 1910 is configured to determine a push detection result corresponding to the information to be detected based on the first push influence feature, the second push influence feature, and the third push influence feature.
In one embodiment, the first feature determination module 1904 includes:
the image input unit is used for inputting the image to be detected into a target object detection model, and the target object detection model comprises a target object area detection network and a target object identification network;
The region obtaining unit is used for detecting the image to be detected according to the target object region detection network to obtain a target object region;
the object identification unit is used for inputting the target object area into a target object identification network for identification to obtain the category of the target object in the target object area;
the category matching unit is used for matching the category of the target object with the category of the preset push reverse influence object, and when the matching is consistent, the first push influence characteristic corresponding to the information to be detected is obtained and is the first push reverse influence characteristic; and when the matching is inconsistent, obtaining a first push influence characteristic corresponding to the information to be detected as a first push forward influence characteristic.
In an embodiment, the object recognition unit is further configured to calculate, via the object recognition network, a region feature, a channel attention feature and a spatial attention feature of the target object region; and combining the channel attention feature, the space attention feature and the region feature, performing convolution calculation according to the combined features, and performing classification recognition according to the convolution calculation result to determine the category of the target object in the target object region.
In one embodiment, the information push detection apparatus 1900 further includes:
The detection network training module is used for acquiring training images of the marked target object areas; inputting the training image of the marked target object area into an initial target object area detection network for detection to obtain an initial target object area; calculating the region error information of the initial target object region and the marked target object region, and updating the network parameters in the initial target object region detection network according to the region error information; and obtaining a trained target object area detection network until the area error information obtained by training accords with a preset first training completion condition.
In one embodiment, the information push detection apparatus 1900 further includes:
the recognition network training module is used for acquiring a training image of the marked target object area, extracting the marked target object area from the training image and acquiring a target object category label corresponding to the marked target object area; inputting the marked target object area into an initial target object recognition network for classification to obtain output training target object class probability, and obtaining training object class according to the training target object class probability; calculating class error information of the class labels of the target objects and the class of the training target objects, and updating network parameters in the initial target object identification network based on the class error information; and obtaining the trained target object recognition network until the class error information obtained by training reaches a preset second training completion condition.
In one embodiment, the second feature derivation module 1906 includes:
the text extraction unit is used for extracting image texts in the images to be detected through an image text recognition algorithm, and segmenting the image texts to obtain each word to be detected;
the word matching unit is used for matching each word to be detected with push reverse influence keywords in a preset keyword library; when the matching is consistent, obtaining a second pushing influence characteristic corresponding to the information to be detected as a second pushing reverse influence characteristic; and when the matching is inconsistent, obtaining a second push influence characteristic corresponding to the information to be detected as a second push forward influence characteristic.
In one embodiment, the information push detection apparatus 1900 further includes:
the text acquisition module is used for acquiring the pushing reverse influence target object identification and acquiring corresponding pushing reverse influence target object text according to the pushing reverse influence target object identification;
the word segmentation module is used for segmenting the text of the pushing reverse influence target object to obtain each text word;
the first calculation module is used for calculating word frequency and reverse file frequency corresponding to each text word;
the second calculation module is used for calculating the importance degree of each text word based on the word frequency and the reverse file frequency corresponding to each text word;
The selecting module is used for selecting the text words with target quantity according to the importance degree of each text word as push reverse influence keywords, and storing the push reverse influence keywords into the keyword library to obtain a preset keyword library.
In one embodiment, the first computing module is further configured to count a total number of words corresponding to the text of the push reverse influence target object, and count a number of occurrences of each text word in the text of the push reverse influence target object; calculating the ratio of the occurrence times corresponding to each text word to the total number of words to obtain the word frequency corresponding to each text word; counting the total number of texts corresponding to the reverse pushing influence target object texts, and counting the target text numbers corresponding to each text word, wherein the target text numbers refer to the text numbers of the target text words contained in the reverse pushing influence target object texts, and the target text words are selected from each text word; and calculating the ratio of the target text number corresponding to each text word to the total text number, and calculating the logarithm of the ratio to obtain the reverse file frequency corresponding to each text word.
In one embodiment, the information push detection apparatus 1900 further includes:
the classification model training module is used for acquiring a training text to be detected and a corresponding training text category to be detected; and training the training text to be detected as input, training the training text category to be detected as a label by using a convolutional neural network, and obtaining a text classification model to be detected when the training is completed.
In an embodiment, the result determining module 1910 is further configured to combine the first push influence feature, the second push influence feature, and the third push influence feature to obtain a combined feature, input the combined feature into the push influence detection model to perform push influence detection, and obtain a push detection result corresponding to the information to be detected.
In one embodiment, the push influencing features include a push forward influencing feature and a push reverse influencing feature; the result determining module 1910 is further configured to match the first push influencing feature, the second push influencing feature, and the third push influencing feature with the push reverse influencing feature, and count feature numbers consistent in matching; when the feature quantity exceeds the preset quantity, the push detection result corresponding to the information to be detected is obtained to be push forbidden, and when the feature quantity does not exceed the preset quantity, the push detection result corresponding to the information to be detected is obtained to be push.
In one embodiment, the information push detection apparatus 1900 further includes:
the pushing module is used for acquiring each receiver address when the pushing detection result is pushing, and sending information to be detected to the receiver terminal according to each receiver address so as to enable the receiver terminal to display the information to be detected;
And the push prohibition module is used for generating alarm information when the push detection result is push prohibition, acquiring the address of the sender and returning the alarm information to the sender terminal corresponding to the address of the sender.
For specific limitation of the information push detection device, reference may be made to the limitation of the information push detection method hereinabove, and no further description is given here. All or part of the modules in the information push detection device can be realized by software, hardware and a combination thereof. The above modules may be embedded in hardware or may be independent of a processor in the computer device, or may be stored in software in a memory in the computer device, so that the processor may call and execute operations corresponding to the above modules.
In one embodiment, a computer device is provided, which may be a server, and the internal structure of which may be as shown in fig. 20. The computer device includes a processor, a memory, and a network interface connected by a system bus. Wherein the processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The database of the computer device is used for storing information to be detected and detection result data. The network interface of the computer device is used for communicating with an external terminal through a network connection. The computer program, when executed by a processor, implements an information push detection method.
It will be appreciated by those skilled in the art that the structure shown in FIG. 20 is merely a block diagram of some of the structures associated with the present inventive arrangements and is not limiting of the computer device to which the present inventive arrangements may be applied, and that a particular computer device may include more or fewer components than shown, or may combine some of the components, or have a different arrangement of components.
In an embodiment, there is also provided a computer device comprising a memory and a processor, the memory having stored therein a computer program, the processor implementing the steps of the method embodiments described above when the computer program is executed.
In one embodiment, a computer-readable storage medium is provided, storing a computer program which, when executed by a processor, implements the steps of the method embodiments described above.
In one embodiment, a computer program product or computer program is provided that includes computer instructions stored in a computer readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device performs the steps in the above-described method embodiments.
Those skilled in the art will appreciate that implementing all or part of the above described methods may be accomplished by way of a computer program stored on a non-transitory computer readable storage medium, which when executed, may comprise the steps of the embodiments of the methods described above. Any reference to memory, storage, database, or other medium used in embodiments provided herein may include at least one of non-volatile and volatile memory. The nonvolatile Memory may include Read-Only Memory (ROM), magnetic tape, floppy disk, flash Memory, optical Memory, or the like. Volatile memory can include random access memory (Random Access Memory, RAM) or external cache memory. By way of illustration, and not limitation, RAM can be in the form of a variety of forms, such as static random access memory (Static Random Access Memory, SRAM) or dynamic random access memory (Dynamic Random Access Memory, DRAM), and the like.
The technical features of the above embodiments may be arbitrarily combined, and all possible combinations of the technical features in the above embodiments are not described for brevity of description, however, as long as there is no contradiction between the combinations of the technical features, they should be considered as the scope of the description.
The above examples illustrate only a few embodiments of the application, which are described in detail and are not to be construed as limiting the scope of the application. It should be noted that it will be apparent to those skilled in the art that several variations and modifications can be made without departing from the spirit of the application, which are all within the scope of the application. Accordingly, the scope of protection of the present application is to be determined by the appended claims.
Claims (22)
1. An information push detection method, characterized in that the method comprises:
acquiring information to be detected, wherein the information to be detected comprises an image to be detected and a text to be detected;
extracting a target object area in the image to be detected, identifying the category of a target object in the target object area, and determining a first pushing influence feature corresponding to the information to be detected according to the category of the target object, wherein the method comprises the following steps: calculating the regional characteristics of a target object region through a target object recognition network, calculating the channel attention characteristics and the space attention characteristics through a convolution block attention module in the target object recognition network, combining the channel attention characteristics, the space attention characteristics and the regional characteristics, carrying out convolution calculation according to the combined characteristics, and carrying out classification recognition according to a convolution calculation result to determine the category of the target object in the target object region;
Extracting an image text in the image to be detected, and matching the image text through a preset keyword library to obtain a second pushing influence characteristic corresponding to the information to be detected;
inputting the text to be detected into a text classification model to be detected for text classification, obtaining the text category to be detected, and obtaining a third pushing influence characteristic corresponding to the information to be detected according to the text category to be detected;
determining a push detection result corresponding to the information to be detected based on the first push influence feature, the second push influence feature and the third push influence feature, including: and matching the first pushing influence feature, the second pushing influence feature and the third pushing influence feature with pushing reverse influence features, counting the feature quantity matched consistently, obtaining a pushing detection result corresponding to the information to be detected as a pushing prohibition when the feature quantity exceeds a preset quantity, and obtaining the pushing detection result corresponding to the information to be detected as a pushing when the feature quantity does not exceed the preset quantity.
2. The method according to claim 1, wherein the extracting the target object region in the image to be detected, identifying a class of a target object in the target object region, and determining the first pushing influence feature corresponding to the information to be detected according to the class of the target object includes:
Inputting the image to be detected into a target object detection model, wherein the target object detection model comprises a target object area detection network and a target object identification network;
detecting the image to be detected according to the target object area detection network to obtain the target object area;
inputting the target object area into the target object identification network for identification to obtain the category of the target object in the target object area;
matching the category of the target object with a preset pushing reverse influence object category, and obtaining a first pushing influence characteristic corresponding to the information to be detected as a first pushing reverse influence characteristic when the matching is consistent;
and when the matching is inconsistent, obtaining a first push influence characteristic corresponding to the information to be detected as a first push forward influence characteristic.
3. The method according to claim 2, wherein the training of the target object area detection network comprises the steps of:
acquiring a training image of the marked target object area;
inputting the training image of the marked target object area into an initial target object area detection network for detection to obtain an initial target object area;
Calculating the region error information of the initial target object region and the marked target object region, and updating network parameters in the initial target object region detection network according to the region error information;
and obtaining a trained target object area detection network until the area error information obtained by training accords with a preset first training completion condition.
4. The method according to claim 2, wherein the training of the target object recognition network comprises the steps of:
acquiring a training image of a marked target object area, extracting the marked target object area from the training image, and acquiring a target object category label corresponding to the marked target object area;
inputting the marked target object area into an initial target object recognition network for classification to obtain output training target object class probability, and obtaining training target object class according to the training target object class probability;
calculating class error information of the target object class label and the training target object class, and updating network parameters in the initial target object identification network based on the class error information;
And obtaining the trained target object recognition network until the class error information obtained by training reaches a preset second training completion condition.
5. The method according to claim 1, wherein the extracting the image text in the image to be detected, and matching the image text through a preset keyword library, to obtain the second pushing influence feature corresponding to the information to be detected, includes:
extracting an image text in the image to be detected through an image text recognition algorithm, and segmenting the image text to obtain each word to be detected;
matching each word to be detected with push reverse influence keywords in the preset keyword library;
when the matching is consistent, obtaining a second push influence characteristic corresponding to the information to be detected as a second push reverse influence characteristic;
and when the matching is inconsistent, obtaining a second push influence characteristic corresponding to the information to be detected as a second push forward influence characteristic.
6. The method of claim 1, wherein the establishing of the pre-set keyword library comprises the steps of:
acquiring a pushing reverse influence target object identifier, and acquiring a corresponding pushing reverse influence target object text according to the pushing reverse influence target object identifier;
Word segmentation is carried out on the text of the pushing reverse influence target object, and each text word is obtained;
calculating word frequency and reverse file frequency corresponding to each text word;
calculating the importance degree of each text word based on the word frequency and the reverse file frequency corresponding to each text word;
selecting a target number of text words as push reverse influence keywords according to the importance degree of each text word, and storing the push reverse influence keywords into a keyword library to obtain the preset keyword library.
7. The method of claim 6, wherein the calculating word frequencies and reverse file frequencies for the respective text words comprises:
counting the total number of words corresponding to the push reverse influence target object text, and counting the occurrence times of each text word in the push reverse influence target object text;
calculating the ratio of the occurrence times corresponding to each text word to the total number of words to obtain the word frequency corresponding to each text word;
counting the total number of texts corresponding to the push reverse influence target object texts, and counting the target text numbers corresponding to the text words, wherein the target text numbers refer to the text numbers of the push reverse influence target object texts containing target text words, and the target text words are obtained by selecting the text words;
And calculating the ratio of the target text number corresponding to each text word to the total text number, and calculating the logarithm of the ratio to obtain the reverse file frequency corresponding to each text word.
8. The method according to claim 1, wherein the training of the text classification model to be detected comprises the steps of:
acquiring a training text to be detected and a corresponding training text category to be detected;
and taking the training text to be detected as input, taking the training text category to be detected as a label, and training by using a convolutional neural network, and obtaining the text classification model to be detected when training is completed.
9. The method of claim 1, wherein determining a push detection result corresponding to the information to be detected based on the first push influence feature, the second push influence feature, and the third push influence feature comprises:
and combining the first pushing influence feature, the second pushing influence feature and the third pushing influence feature to obtain combined features, inputting the combined features into a pushing influence detection model to carry out pushing influence detection, and obtaining a pushing detection result corresponding to the information to be detected.
10. The method of claim 1, further comprising, after the determining a push detection result corresponding to the information to be detected based on the first push influence feature, the second push influence feature, and the third push influence feature:
when the push detection result is push, acquiring each receiver address, and sending the information to be detected to a receiver terminal according to each receiver address so that the receiver terminal displays the information to be detected;
when the push detection result is that push is forbidden, generating alarm information, acquiring a sender address, and returning the alarm information to a sender terminal corresponding to the sender address.
11. An information push detection device, the device comprising:
the information acquisition module is used for acquiring information to be detected, wherein the information to be detected comprises an image to be detected and a text to be detected;
the first feature determining module is configured to extract a target object area in the image to be detected, identify a class of a target object in the target object area, and determine a first pushing influence feature corresponding to the information to be detected according to the class of the target object, where the first pushing influence feature includes: calculating the regional characteristics of a target object region through a target object recognition network, calculating the channel attention characteristics and the space attention characteristics through a convolution block attention module in the target object recognition network, combining the channel attention characteristics, the space attention characteristics and the regional characteristics, carrying out convolution calculation according to the combined characteristics, and carrying out classification recognition according to a convolution calculation result to determine the category of the target object in the target object region;
The second feature obtaining module is used for extracting image texts in the images to be detected, and matching the image texts through a preset keyword library to obtain second pushing influence features corresponding to the information to be detected;
the third feature obtaining module is used for inputting the text to be detected into a text classification model to be detected for text classification, obtaining the text category to be detected, and obtaining a third pushing influence feature corresponding to the information to be detected according to the text category to be detected;
the result determining module is configured to determine a push detection result corresponding to the information to be detected based on the first push influence feature, the second push influence feature, and the third push influence feature, and includes: and matching the first pushing influence feature, the second pushing influence feature and the third pushing influence feature with pushing reverse influence features, counting the feature quantity matched consistently, obtaining a pushing detection result corresponding to the information to be detected as a pushing prohibition when the feature quantity exceeds a preset quantity, and obtaining the pushing detection result corresponding to the information to be detected as a pushing when the feature quantity does not exceed the preset quantity.
12. The apparatus of claim 11, wherein the first feature determination module comprises:
the image input unit is used for inputting the image to be detected into a target object detection model, wherein the target object detection model comprises a target object area detection network and a target object identification network;
the region obtaining unit is used for detecting the image to be detected according to the target object region detection network to obtain the target object region;
the object identification unit is used for inputting the target object area into the target object identification network for identification to obtain the category of the target object in the target object area;
the category matching unit is used for matching the category of the target object with a preset pushing reverse influence object category, and when the matching is consistent, a first pushing influence characteristic corresponding to the information to be detected is obtained and is a first pushing reverse influence characteristic;
and when the matching is inconsistent, obtaining a first push influence characteristic corresponding to the information to be detected as a first push forward influence characteristic.
13. The apparatus of claim 12, wherein the apparatus further comprises:
The detection network training module is used for acquiring training images of the marked target object areas; inputting the training image of the marked target object area into an initial target object area detection network for detection to obtain an initial target object area; calculating the region error information of the initial target object region and the marked target object region, and updating network parameters in the initial target object region detection network according to the region error information; and obtaining a trained target object area detection network until the area error information obtained by training accords with a preset first training completion condition.
14. The apparatus of claim 12, wherein the apparatus further comprises:
the recognition network training module is used for acquiring a training image of the marked target object area, extracting the marked target object area from the training image and acquiring a target object category label corresponding to the marked target object area; inputting the marked target object area into an initial target object recognition network for classification to obtain output training target object class probability, and obtaining training target object class according to the training target object class probability; calculating class error information of the target object class label and the training target object class, and updating network parameters in the initial target object identification network based on the class error information; and obtaining the trained target object recognition network until the class error information obtained by training reaches a preset second training completion condition.
15. The apparatus of claim 11, wherein the second feature derivation module comprises:
the text extraction unit is used for extracting the image text in the image to be detected through an image text recognition algorithm, and segmenting the image text to obtain each word to be detected;
the word matching unit is used for matching each word to be detected with the push reverse influence keywords in the preset keyword library;
when the matching is consistent, obtaining a second push influence characteristic corresponding to the information to be detected as a second push reverse influence characteristic;
and when the matching is inconsistent, obtaining a second push influence characteristic corresponding to the information to be detected as a second push forward influence characteristic.
16. The apparatus of claim 11, wherein the apparatus further comprises:
the text acquisition module is used for acquiring the pushing reverse influence target object identification and acquiring corresponding pushing reverse influence target object text according to the pushing reverse influence target object identification;
the word segmentation module is used for segmenting the text of the push reverse influence target object to obtain each text word;
the first calculation module is used for calculating word frequency and reverse file frequency corresponding to each text word;
The second calculation module is used for calculating the importance degree of each text word based on the word frequency and the reverse file frequency corresponding to each text word;
and the selecting module is used for selecting the text words with target quantity according to the importance degree of each text word as push reverse influence keywords, and storing the push reverse influence keywords into a keyword library to obtain the preset keyword library.
17. The apparatus of claim 16, wherein the first computing module is further configured to count a total number of words corresponding to the push reverse impact target object text and count a number of occurrences of the respective text word in the push reverse impact target object text; calculating the ratio of the occurrence times corresponding to each text word to the total number of words to obtain the word frequency corresponding to each text word; counting the total number of texts corresponding to the push reverse influence target object texts, and counting the target text numbers corresponding to the text words, wherein the target text numbers refer to the text numbers of the push reverse influence target object texts containing target text words, and the target text words are obtained by selecting the text words; and calculating the ratio of the target text number corresponding to each text word to the total text number, and calculating the logarithm of the ratio to obtain the reverse file frequency corresponding to each text word.
18. The apparatus of claim 11, wherein the apparatus further comprises:
the classification model training module is used for acquiring a training text to be detected and a corresponding training text category to be detected; and taking the training text to be detected as input, taking the training text category to be detected as a label, and training by using a convolutional neural network, and obtaining the text classification model to be detected when training is completed.
19. The apparatus of claim 11, wherein the result determining module is further configured to combine the first push influence feature, the second push influence feature, and the third push influence feature to obtain a combined feature, input the combined feature into a push influence detection model to perform push influence detection, and obtain a push detection result corresponding to the information to be detected.
20. The apparatus of claim 11, wherein the apparatus further comprises:
the pushing module is used for acquiring each receiver address when the pushing detection result is pushing, and sending the information to be detected to a receiver terminal according to each receiver address so that the receiver terminal displays the information to be detected;
And the push prohibition module is used for generating alarm information when the push detection result is push prohibition, acquiring a sender address and returning the alarm information to a sender terminal corresponding to the sender address.
21. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that the processor implements the steps of the method of any one of claims 1 to 10 when the computer program is executed.
22. A computer readable storage medium storing a computer program, characterized in that the computer program when executed by a processor implements the steps of the method of any one of claims 1 to 10.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010756055.0A CN113537206B (en) | 2020-07-31 | 2020-07-31 | Push data detection method, push data detection device, computer equipment and storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010756055.0A CN113537206B (en) | 2020-07-31 | 2020-07-31 | Push data detection method, push data detection device, computer equipment and storage medium |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113537206A CN113537206A (en) | 2021-10-22 |
| CN113537206B true CN113537206B (en) | 2023-11-10 |
Family
ID=78094255
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010756055.0A Active CN113537206B (en) | 2020-07-31 | 2020-07-31 | Push data detection method, push data detection device, computer equipment and storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113537206B (en) |
Families Citing this family (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113779308B (en) * | 2021-11-12 | 2022-02-25 | 冠传网络科技(南京)有限公司 | Short video detection and multi-classification method, device and storage medium |
| CN114399699A (en) * | 2021-12-06 | 2022-04-26 | 北京达佳互联信息技术有限公司 | Target recommendation object determination method and device, electronic equipment and storage medium |
| CN114496256A (en) * | 2022-01-28 | 2022-05-13 | 北京百度网讯科技有限公司 | Event detection method and device, electronic equipment and storage medium |
| CN116522011B (en) * | 2023-05-16 | 2024-02-13 | 深圳九星互动科技有限公司 | Big data-based pushing method and pushing system |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108038209A (en) * | 2017-12-18 | 2018-05-15 | 深圳前海微众银行股份有限公司 | Answer system of selection, device and computer-readable recording medium |
| CN110555482A (en) * | 2019-09-06 | 2019-12-10 | 腾讯科技（深圳）有限公司 | Vulgar picture identification method and device based on artificial intelligence and electronic equipment |
| WO2020000847A1 (en) * | 2018-06-25 | 2020-01-02 | 中译语通科技股份有限公司 | News big data-based method and system for monitoring and analyzing risk perception index |
| CN111080628A (en) * | 2019-12-20 | 2020-04-28 | 湖南大学 | Image tampering detection method and device, computer equipment and storage medium |
| CN111104898A (en) * | 2019-12-18 | 2020-05-05 | 武汉大学 | Image scene classification method and device based on target semantics and attention mechanism |
| US10671878B1 (en) * | 2019-01-11 | 2020-06-02 | Capital One Services, Llc | Systems and methods for text localization and recognition in an image of a document |
| CN111310026A (en) * | 2020-01-17 | 2020-06-19 | 南京邮电大学 | Artificial intelligence-based yellow-related terrorism monitoring method |
-
2020
- 2020-07-31 CN CN202010756055.0A patent/CN113537206B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108038209A (en) * | 2017-12-18 | 2018-05-15 | 深圳前海微众银行股份有限公司 | Answer system of selection, device and computer-readable recording medium |
| WO2020000847A1 (en) * | 2018-06-25 | 2020-01-02 | 中译语通科技股份有限公司 | News big data-based method and system for monitoring and analyzing risk perception index |
| US10671878B1 (en) * | 2019-01-11 | 2020-06-02 | Capital One Services, Llc | Systems and methods for text localization and recognition in an image of a document |
| CN110555482A (en) * | 2019-09-06 | 2019-12-10 | 腾讯科技（深圳）有限公司 | Vulgar picture identification method and device based on artificial intelligence and electronic equipment |
| CN111104898A (en) * | 2019-12-18 | 2020-05-05 | 武汉大学 | Image scene classification method and device based on target semantics and attention mechanism |
| CN111080628A (en) * | 2019-12-20 | 2020-04-28 | 湖南大学 | Image tampering detection method and device, computer equipment and storage medium |
| CN111310026A (en) * | 2020-01-17 | 2020-06-19 | 南京邮电大学 | Artificial intelligence-based yellow-related terrorism monitoring method |
Non-Patent Citations (4)
| Title |
|---|
| "The Study of ROI Detection Based on Visual Attention Mechanism";Qing-hua Chen et.al.;《2010 6th International Conference on Wireless Communications Networking and Mobile Computing (WiCOM)》;全文 * |
| "基于深度学习的不良信息检测技术的研究";谢鑫;《中国优秀硕士学位论文全文数据库信息科技辑》;第2020年卷(第07期);参见摘要、第2-6章，具体参见第9-10，16-18，22-28，34-39页，56-58页 * |
| 基于Anchor-free的交通标志检测;范红超;李万志;章超权;;地球信息科学学报(第01期);第92-103页 * |
| 胡涛.《面向城市疾病制图分析的地址匹配研究》.吉林大学出版社,2019,第74页. * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113537206A (en) | 2021-10-22 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110866140B (en) | Image feature extraction model training method, image searching method and computer equipment | |
| CN113537206B (en) | Push data detection method, push data detection device, computer equipment and storage medium | |
| CN108694225B (en) | Image searching method, feature vector generating method and device and electronic equipment | |
| CN110472090B (en) | Image retrieval method based on semantic tags, related device and storage medium | |
| CN109918560B (en) | Question and answer method and device based on search engine | |
| CN105426354B (en) | The fusion method and device of a kind of vector | |
| CN111708888B (en) | Classification method, device, terminal and storage medium based on artificial intelligence | |
| CN111582409A (en) | Training method of image label classification network, image label classification method and device | |
| CN107679070B (en) | Intelligent reading recommendation method and device and electronic equipment | |
| CN112148831B (en) | Image-text mixed retrieval method and device, storage medium and computer equipment | |
| CN113076905B (en) | Emotion recognition method based on context interaction relation | |
| CN112258254B (en) | Internet advertisement risk monitoring method and system based on big data architecture | |
| CN111444387A (en) | Video classification method and device, computer equipment and storage medium | |
| CN115982473B (en) | Public opinion analysis arrangement system based on AIGC | |
| CN116226785A (en) | Target object recognition method, multi-mode recognition model training method and device | |
| CN111782793A (en) | Intelligent customer service processing method, system and equipment | |
| CN117076688A (en) | Knowledge question-answering method and device based on domain knowledge graph and electronic equipment | |
| Maynard et al. | Multimodal sentiment analysis of social media | |
| CN112100212A (en) | Case scenario extraction method based on machine learning and rule matching | |
| CN116955707A (en) | Content tag determination method, device, equipment, medium and program product | |
| CN115730058A (en) | Reasoning question-answering method based on knowledge fusion | |
| CN114329004A (en) | Digital fingerprint generation method, digital fingerprint generation device, data push method, data push device and storage medium | |
| CN115630357B (en) | Method for judging behavior of collecting personal information by application program crossing boundary | |
| CN114896436B (en) | Network structure searching method based on characterization mutual information | |
| CN112035670B (en) | Multi-modal rumor detection method based on image emotional tendency |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| REG | Reference to a national code |
Ref country code: HK Ref legal event code: DE Ref document number: 40055190 Country of ref document: HK |
|
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |