Looking for breakthrough ideas for innovation challenges? Try Patsnap Eureka!

Internet live pornography detecting method based on deep convolutional neural network

An Internet live broadcast and deep convolution technology, applied in the field of Internet live broadcast detection, can solve the problems of low precision, high model misjudgment rate, visual fatigue of patrol personnel, etc., and achieve the effect of improving work efficiency and detection rate

Active Publication Date: 2018-06-12
TIANGE TECH HANGZHOU
View PDF6 Cites 48 Cited by
  • Summary
  • Abstract
  • Description
  • Claims
  • Application Information

AI Technical Summary

Problems solved by technology

[0006] 2. Advantages and disadvantages of the algorithm: Advantages: The algorithm is intuitive, easy to understand and implement; Disadvantages: For scenes with more exposed skin areas but not pornographic images (such as swimming, beach sunbathing, etc.) The model misjudgment rate is high
[0009] 2. Advantages and disadvantages of the algorithm: Advantages: The detection accuracy has been greatly improved; Disadvantages: The selection of codewords and the size of the codebook in the algorithm are difficult to determine, and seriously affect the detection accuracy of the algorithm. In addition, the large amount of redundancy in the image background information also affects The judgment of the classifier
[0012] (1) The accuracy of the detection algorithm of traditional pornographic image content is not very high in practical applications, which may easily cause omissions in the detection of pornographic live content;
[0013] (2) With the rapid growth of live broadcast content, the existing supervision and inspection tasks have increased sharply, further increasing the operating costs of the live broadcast platform and the work intensity of patrol personnel;
[0014] (3) The supervision method mainly relies on inspectors to conduct live broadcast review, which requires the inspectors to conduct 24-hour uninterrupted online review. The long-term video wall review is likely to cause visual fatigue of the inspectors, which in turn increases the risk of omissions in inspection of pornographic content

Method used

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
View more

Image

Smart Image Click on the blue labels to locate them in the text.
Viewing Examples
Smart Image
  • Internet live pornography detecting method based on deep convolutional neural network
  • Internet live pornography detecting method based on deep convolutional neural network
  • Internet live pornography detecting method based on deep convolutional neural network

Examples

Experimental program
Comparison scheme
Effect test

Embodiment 1

[0060] Example 1 Live Video Sample Image Preprocessing

[0061] At present, data enhancement is a common method for deep learning model training. The main methods are: 1) image mirroring: mirroring the positive sample image left and right, up and down; 2) image rotation: rotating the positive sample image at different angles; 3) Image pixel change: Properly change the pixel values ​​of different channels of the positive sample image; 4) Add noise to the image: Add different types and degrees of noise processing to the positive sample image; 5) Image Gaussian blur: Perform different processing on the positive sample image Degree of Gaussian blurring.

[0062] After classification of artificial pornographic positive sample images, 50,000 live broadcast pornographic images were selected as positive samples for model training, and a data enhancement ratio of 1:5 was used to generate a total of 300,000 positive sample images (50,000 original images, 25 enhanced samples 10,000), in...

Embodiment 2

[0071] Example 2 Enhancement and preprocessing of sample image data based on training model iteration

[0072] In data enhancement and sample preprocessing, a large number of positive samples are manually screened, data enhancement processing, and image size processing are involved. In order to reduce the cost of sample processing and improve sample processing efficiency, this patent proposes a method based on training model iteration Sample image data enhancement and preprocessing methods, specifically as figure 2 Shown:

[0073] A small number of positive samples and negative samples are screened without data enhancement processing. The processing method of image size is a direct normalization method, and the convolutional neural network is used for model training to obtain an initial training model with a low detection rate;

[0074] Carry out a slider cropping process with a certain step size on the positive sample images of different sizes (the step size is set accordin...

Embodiment 3

[0078] Embodiment 3 Convolutional neural network model training based on deep learning

[0079] Such as Figure 4 As shown, the deep learning-based convolutional neural network model training method of the present invention adopts the idea of ​​image classification and recognition based on multiple convolutional neural network models for model training.

[0080]At present, the mainstream deep learning training frameworks include: caffe, TensorFlow, Torch, keras, CNTK, paddlepaddle, etc. Among them, caffe and TensorFlow are the most used training frameworks. Different frameworks have their own advantages and disadvantages, so I won’t repeat them here. Choose Model configuration and training can be done with the appropriate framework. The existing classic convolutional neural networks are more representative: LeNet, Alexnet, VGGNet, GoogLeNet, ResNet, etc. Different models have different differences in model complexity, model parameters, training time, correct recognition rate,...

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to View More

PUM

PropertyMeasurementUnit
Image resolutionaaaaaaaaaa
Login to View More

Abstract

The invention discloses an internet live pornography detecting method based on a deep convolutional neural network. The detecting accuracy of the content of live pornography is greatly improved by means of a deep learning technology, a sample image data strengthening and preprocessing method based on convolutional neural network training model iteration, a model configuration and training method based on a deep learning convolutional neural network, an image combined testing method based on multi-tailoring and multi-models, an online image dynamic cutting method of internet live video supervision and other optimizing methods, thus the conversion of the supervision mode that the algorithm model pornographic content automatic detecting plays the principal role, and artificial real-time checking plays the subsidiary role is achieved, the supervision efficiency of a live platform is improved, the labor cost for operation is lowered, the work intensity of inspection personnel is reduced, and the high-speed increasing requirement of the live content is met.

Description

technical field [0001] The invention belongs to the field of Internet live broadcast detection, and in particular relates to a method for detecting Internet live pornographic images based on a deep convolutional neural network. Background technique [0002] With the rapid development of the mobile Internet, live video broadcasting has quickly entered people's lives. In the process of development, incidents involving pornographic live broadcasting have occurred from time to time, which have seriously hindered the development environment of live video broadcasting and brought adverse effects on society. Therefore, it is increasingly urgent to increase the regulatory scrutiny of live video broadcasting. However, due to the huge content of live video broadcasts, relying solely on human review is far from meeting actual needs, and it is very important to further improve the detection technology level of pornographic live broadcast content. [0003] At present, there are two main...

Claims

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to View More

Application Information

Patent Timeline
no application Login to View More
IPC IPC(8): G06K9/00G06K9/46G06K9/62H04N21/24
CPCH04N21/2407G06V20/40G06V10/464G06F18/2411
Inventor 周建政明建华邓豪
Owner TIANGE TECH HANGZHOU
Who we serve
  • R&D Engineer
  • R&D Manager
  • IP Professional
Why Patsnap Eureka
  • Industry Leading Data Capabilities
  • Powerful AI technology
  • Patent DNA Extraction
Social media
Patsnap Eureka Blog
Learn More
PatSnap group products

Browse by: Latest US Patents, China's latest patents, Technical Efficacy Thesaurus, Application Domain, Technology Topic, Popular Technical Reports.

© 2024 PatSnap. All rights reserved.Legal|Privacy policy|Modern Slavery Act Transparency Statement|Sitemap|About US| Contact US: help@patsnap.com