CN104182735A - Training optimization pornographic picture or video detection method based on convolutional neural network - Google Patents
Training optimization pornographic picture or video detection method based on convolutional neural network Download PDFInfo
- Publication number
- CN104182735A CN104182735A CN201410406286.3A CN201410406286A CN104182735A CN 104182735 A CN104182735 A CN 104182735A CN 201410406286 A CN201410406286 A CN 201410406286A CN 104182735 A CN104182735 A CN 104182735A
- Authority
- CN
- China
- Prior art keywords
- layer
- convolutional neural
- neural networks
- classification
- pornographic
- 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.)
- Pending
Links
Landscapes
- Image Analysis (AREA)
Abstract
The invention relates to a training optimization pornographic picture or video detection method based on a convolutional neural network. A final detection model is obtained through circuit training, classification detection accuracy can be improved to a largest extent, and deficiencies that a perfect complexion model and high detection accuracy are difficult to establish due to ambient light diversity and race diversity and the detection accuracy is always low due to external factors, such as human body gesture diversity, sheltering and the like can be overcome. When the method disclosed by the invention is used for picture data or video data to detect pictures or a video image sequence, only a classification result shows that the pictures or video image sequence is pornographic, a picture set or videos are defined as pornographic data, and a deficiency that a universal rule is difficult to establish for different pornographic videos is overcome.
Description
Technical field
The present invention relates to a kind of pornographic image detection method, more particularly, relate to a kind of pornographic image based on convolutional neural networks or video detecting method of training optimization.
Background technology
Along with the widespread use of internet, user also can run into flame when obtaining a large amount of useful informations, especially the most serious with pornographic image and pornographic video.The method that existing pornographic image or pornographic video detect is mainly to utilize the colour of skin to identify, but in actual applications, existing detection method exists the factor of following restriction Detection accuracy:
1) variation of ambient light, the diversity of ethnic group etc. make to be difficult to the complexion model of Erecting and improving and the accuracy rate of detection;
2) human body attitude diversity, the extraneous factor such as block and make the accuracy rate that detects always not high;
3) for different pornographic videos, be difficult to set up more common rule.
Summary of the invention
The object of the invention is to overcome the deficiencies in the prior art, a kind of impact that not only can adapt to various environment and attitude and the extraneous factor such as block is provided, and can set up more common detection rule, improve the pornographic image based on convolutional neural networks or the video detecting method of the training optimization of the accuracy rate detecting.
Technical scheme of the present invention is as follows:
Train the pornographic image based on convolutional neural networks or the video detecting method of optimization, default view data or video data; And utilization has been carried out convolutional neural networks and has been trained the Preliminary detection model obtaining to carry out classification and Detection for the first time; If there is wrong sample in classification and Detection result for the first time, to wrong sample, utilize convolutional neural networks to train for the second time, obtain transition detection model for the first time; And utilize transition detection model for the first time to carry out classification and Detection for the second time, if there is wrong sample in classification and Detection result for the second time, to wrong sample, utilize convolutional neural networks to train for the third time, obtain transition detection model for the second time, the above step that circulates, until wrong sample is less than default threshold values or does not exist in testing result, obtains final detection model; Utilize final detection model to carry out classification and Detection to view data to be measured or video data.
As preferably, default view data or video data comprise pornographic data and the normal data of presorting, using these two classes data as two scenes, as input data, offer convolutional neural networks and carry out learning classification, obtain judging whether view data or video data are pornographic and normal detection model.
As preferably, the process that training is optimized is: the transition detection model obtaining is carried out to class test to default view data or video data, the sample that the classification and Detection result view data different from the result of presorting or video data are decided to be to classification error.
As preferably, the sample of classification error is specially: the environment classification label that classification and Detection result is obtained is collected with the inconsistent sample image data of the environment classification label of presorting or video data, re-starts classification.
As preferably, the order of the network structure of convolutional neural networks is: input layer, some groups layer, full articulamentum, probability statistics layer.
As preferably, further, group's layer comprises convolutional layer, active coating successively.
As preferably, in convolutional layer, the size of core must be odd number, and is not more than the wide or high of this layer of input; Intermediate representation does not change during by convolutional layer wide and high, and whether port number does not limit and change; In active coating, active coating does not change wide, high, the port number that convolutional layer represents; After active coating is arranged on convolutional layer or full articulamentum; By the intermediate representation after full articulamentum, be 1 dimension, and after full articulamentum, cannot through group's layer, process again; Probability between real-valued the becoming [0,1] that probability statistics layer produces full articulamentum.
As preferably, probability statistics layer is exported the probability of each label, and selects the label of maximum probability as the type of present image.
As preferably, further, one deck of down-sampling layer and normalization layer or two-layer is set after active coating.
As preferably, down-sampling layer does not change the port number of intermediate representation, and the down-sampling layer size that is core to the drawdown ratio of image; Normalization layer does not change any size of intermediate representation.
Beneficial effect of the present invention is as follows:
Method of the present invention, through circuit training, obtain final detection model, can improve classification and Detection accuracy rate in maximum program ground, overcome the variation because of ambient light, the diversity of ethnic group etc. and make to be difficult to the complexion model of Erecting and improving and the accuracy rate of detection, and human body attitude diversity, the extraneous factor such as block and make the accuracy rate that detects not high deficiency always.
When method of the present invention is used for view data or video data, carry out the detection to picture, or the detection to sequence of video images, as long as classification results is pornographic, be that this image collection of definable or video are pornographic data, overcome for different pornographic videos and be difficult to set up more common regular deficiency.
Embodiment
Below in conjunction with embodiment, the present invention is described in further detail.
The invention provides a kind of pornographic image based on convolutional neural networks or video detecting method of training optimization, default view data or video data; And utilization has been carried out convolutional neural networks and has been trained the Preliminary detection model obtaining to carry out classification and Detection for the first time; If there is wrong sample in classification and Detection result for the first time, to wrong sample, utilize convolutional neural networks to train for the second time, obtain transition detection model for the first time; And utilize transition detection model for the first time to carry out classification and Detection for the second time, if there is wrong sample in classification and Detection result for the second time, to wrong sample, utilize convolutional neural networks to train for the third time, obtain transition detection model for the second time, the above step that circulates, until wrong sample is less than default threshold values or does not exist in testing result, be that error rate reaches tolerance interval or no longer changes, obtain final detection model; Utilize final detection model to carry out classification and Detection to view data to be measured or video data.
Detection model is mainly divided into two classes: pornographic data and normal data.Default view data or video data comprise pornographic data and the normal data of presorting, can be by manually carrying out category filter.Using these two classes data as two scenes, as input data, offer convolutional neural networks and carry out learning classification, obtain judging whether view data or video data are pornographic and normal detection model.
The process that training is optimized is: the transition detection model obtaining is carried out to class test to default view data or video data, the sample that the classification and Detection result view data different from the result of presorting or video data are decided to be to classification error, be specially: the environment classification label that classification and Detection result is obtained is collected with the inconsistent sample image data of the environment classification label of presorting or video data, re-start classification, adjust network structure.And the process of optimizing is the sample image after reclassifying is carried out to training study again, so repeat " training---adjusting network structure---test---is training " process until the error rate of classification and Detection result in tolerance interval.
The network of convolutional neural networks is multilayer, and in the present invention, the order of network structure is: input layer, some groups layer (at least one group of group's layer), full articulamentum, probability statistics layer.Group's layer comprises convolutional layer, active coating successively.According to implementing needs, further, one deck of down-sampling layer and normalization layer or two-layer is set after active coating.In convolutional layer, active coating, down-sampling layer, normalization layer, the core of each layer size and output size can carry out regulating arbitrarily, and each layer has an input and produce an output, and the output of every one deck is as the input of lower one deck.
In the present embodiment, the input of input layer size is Height x Weight x Channel, and wherein Weight, Height are the wide and high of input layer image, and Channel is the port number of input layer image; Because the present invention uses the hard-wired reason of GPU, Weight=Height; If input data are image, Channel value can only be 1 or 3.If video data, the port number that Channel number is image and total product of inputting frame number.
Convolutional layer:
1) size of core must be odd number, and is not more than the wide or high of this layer of input;
2) when intermediate representation is by convolutional layer, do not change widely and high, port number is variable can be constant; Can be any positive integer in theory, because the present invention uses the hard-wired reason of GPU, be 16 multiple here.
Active coating:
1) that active coating does not change that convolutional layer represents is wide, height or port number;
2) activation function that active coating is used includes but not limited to following type function:
f(x)=1/(1+e
-x);
F (x)=a*tanh (b*x), a, b is any non-zero real;
f(x)=max(0,x);
f(x)=min(a,max(0,x));
f(x)=log(1+e
x);
f(x)=|x|;
f(x)=x
2;
f(x)=ax+b;
3) active coating is followed after convolutional layer or full articulamentum.
Down-sampling layer:
1) down-sampling layer does not change the port number of intermediate representation;
2) down-sampling layer is the size of core to the drawdown ratio of image: the down-sampling layer that core is m * n can cause intermediate representation to be reduced into last layer (1/m) * (1/n), m and n can be random natural number in theory, because the present invention uses the hard-wired reason of GPU, m=n.For example, 15 x 15 x 32, by after the down-sampling of 3 x 3, become 5 x 5 x 32; 15 x 15 x 32, by after the down-sampling of 5 x 5, become 3 x 3 x 32; But 15 x 15 x 32 can not carry out the down-sampling of 2 x 2, because 15 can not be divided exactly by 2; Be not, input size must be 2 inferior power, 16,32,64 etc., as long as input size guarantees to be sampled by all down-sampling layers.
Normalization layer:
1) normalization layer does not change any size of intermediate representation;
2) normalization layer not necessarily, adds normalization layer and conventionally can improve precision and increase calculated amount; Whether add normalization layer, see the precision of actual lifting after adding and the speed of loss.
General combination is: convolution---activates---down-sampling---normalization.
Following situation is special:
1) when interpolation normalization layer has but increased a lot of operand to precision improvement is less, cancel normalization layer, adopt following combination: convolution---activation---down-sampling;
2) in advance, effect is basic identical for normalization layer, adopts following combination: convolution---activates---normalization---down-sampling.
3) cancel down-sampling layer: convolution---activate; Or convolution---activation---normalization; Down-sampling essence is in order to increase robustness, has in passing the effect of the operand that reduces succeeding layer simultaneously; In a network, conventionally have which floor down-sampling, but not all " convolution---activate " all to follow down-sampling below.
Full articulamentum:
1) by the intermediate representation after full articulamentum, can become 1 dimension, be no longer 3 dimensions;
2) output of full articulamentum can be any;
3) once connect through complete, just cannot carry out convolution, down-sampling or normalization;
4) after full articulamentum, can connect active coating, or continue to connect full connection.
Probability statistics layer (SoftMax layer):
After being connected on full articulamentum, effect is the probability between real-valued the becoming [0,1] that full articulamentum is produced.Probability statistics layer is exported the probability of each label, and selects the label of maximum probability as the type of present image.That is,, by the input layer of view data to be detected or video data input neural network, the probability that obtains each label at last probability statistics layer is (interval [0,1] real-valued in), always have 2 types, i.e. 2 data, these 2 data and equal 1.Select the label of maximum probability as the type of this image.
In the present embodiment, it is as shown in table 1 that pornographic image detects the network structure of using:
Table 1: network structure table
The number of plies | Type | Core size | Output size | Explain |
1 | Input layer | ? | 32x32x3 | ? |
2 | Convolutional layer | 5x5 | 32x32x32 | ? |
3 | Active coating | ? | 32x32x32 | f(x)=|x| |
4 | Down-sampling layer | 2x2 | 16x16x32 | ? |
5 | Normalization layer | ? | 16x16x32 | Use local normalization |
6 | Convolutional layer | 5x5 | 16x16x16 | ? |
7 | Active coating | ? | 16x16x16 | f(x)=x 2 |
8 | Down-sampling layer | 2x2 | 8x8x16 | ? |
9 | Normalization layer | ? | 8x8x16 | Use local normalization |
10 | Full articulamentum | ? | 7 data | ? |
11 | Probability statistics layer | ? | 7 data | ? |
In the present embodiment, it is as shown in table 2 that pornographic video detects the network structure of using:
Table 2: network structure table
The number of plies | Type | Core size | Output size | Explain |
1 | Input layer | ? | 128x128x16 | ? |
2 | Convolutional layer | 7x7 | 122x122x64 | ? |
3 | Active coating | ? | 122x122x64 | f(x)=max(0,x) |
4 | Normalization | 5x5 | 122x122x64 | Use local normalization |
5 | Down-sampling layer | 2x2 | 61x61x64 | ? |
6 | Full articulamentum | ? | 20 | ? |
7 | Full articulamentum | ? | 2 | ? |
8 | Probability statistics layer | ? | 2 data | ? |
While utilizing pornographic detection model of the present invention to carry out pornographic detection to view data to be detected, as long as detect picture result, be pornographic, judge that picture is as porny.
While utilizing pornographic detection model of the present invention to carry out pornographic detection to video data to be detected, first extract the image of video, in the present embodiment, mainly to take every 8 frames to get the mode of a frame to video, video data is converted into image sequence to be detected, as long as testing result is pornographic, judge that this video is as pornographic video.
Above-described embodiment is only for the present invention is described, and not as limitation of the invention.So long as according to technical spirit of the present invention, to above-described embodiment change, modification etc. all will drop in the scope of claim of the present invention.
Claims (10)
1. the pornographic image based on convolutional neural networks or a video detecting method of training optimization, is characterized in that, default view data or video data; And utilization has been carried out convolutional neural networks and has been trained the Preliminary detection model obtaining to carry out classification and Detection for the first time; If there is wrong sample in classification and Detection result for the first time, to wrong sample, utilize convolutional neural networks to train for the second time, obtain transition detection model for the first time; And utilize transition detection model for the first time to carry out classification and Detection for the second time, if there is wrong sample in classification and Detection result for the second time, to wrong sample, utilize convolutional neural networks to train for the third time, obtain transition detection model for the second time, the above step that circulates, until wrong sample is less than default threshold values or does not exist in testing result, obtains final detection model; Utilize final detection model to carry out classification and Detection to view data to be measured or video data.
2. training according to claim 1 is optimized the pornographic image based on convolutional neural networks or video detecting method, it is characterized in that, default view data or video data comprise pornographic data and the normal data of presorting, using these two classes data as two scenes, as input data, offer convolutional neural networks and carry out learning classification, obtain judging whether view data or video data are pornographic and normal detection model.
3. training according to claim 2 is optimized the pornographic image based on convolutional neural networks or video detecting method, it is characterized in that, the process that training is optimized is: the transition detection model obtaining is carried out to class test to default view data or video data, the sample that the classification and Detection result view data different from the result of presorting or video data are decided to be to classification error.
4. training according to claim 3 is optimized the pornographic image based on convolutional neural networks or video detecting method, it is characterized in that, the sample of classification error is specially: the environment classification label that classification and Detection result is obtained is collected with the inconsistent sample image data of the environment classification label of presorting or video data, re-starts classification.
5. training according to claim 1 is optimized the pornographic image based on convolutional neural networks or video detecting method, is characterized in that, the order of the network structure of convolutional neural networks is: input layer, some groups layer, full articulamentum, probability statistics layer.
6. training according to claim 5 is optimized the pornographic image based on convolutional neural networks or video detecting method, is characterized in that, further, group's layer comprises convolutional layer, active coating successively.
7. training according to claim 6 is optimized the pornographic image based on convolutional neural networks or video detecting method, is characterized in that, in convolutional layer, the size of core must be odd number, and is not more than the wide or high of this layer of input; Intermediate representation does not change during by convolutional layer wide and high, and whether port number does not limit and change; In active coating, active coating does not change wide, high, the port number that convolutional layer represents; After active coating is arranged on convolutional layer or full articulamentum; By the intermediate representation after full articulamentum, be 1 dimension, and after full articulamentum, cannot through group's layer, process again; Probability between real-valued the becoming [0,1] that probability statistics layer produces full articulamentum.
8. training according to claim 7 is optimized the pornographic image based on convolutional neural networks or video detecting method, is characterized in that, probability statistics layer is exported the probability of each label, and selects the label of maximum probability as the type of present image.
9. training according to claim 6 is optimized the pornographic image based on convolutional neural networks or video detecting method, is characterized in that, further, one deck of down-sampling layer and normalization layer or two-layer is set after active coating.
10. training according to claim 8 is optimized the pornographic image based on convolutional neural networks or video detecting method, is characterized in that, down-sampling layer does not change the port number of intermediate representation, and the down-sampling layer size that is core to the drawdown ratio of image; Normalization layer does not change any size of intermediate representation.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201410406286.3A CN104182735A (en) | 2014-08-18 | 2014-08-18 | Training optimization pornographic picture or video detection method based on convolutional neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201410406286.3A CN104182735A (en) | 2014-08-18 | 2014-08-18 | Training optimization pornographic picture or video detection method based on convolutional neural network |
Publications (1)
Publication Number | Publication Date |
---|---|
CN104182735A true CN104182735A (en) | 2014-12-03 |
Family
ID=51963762
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201410406286.3A Pending CN104182735A (en) | 2014-08-18 | 2014-08-18 | Training optimization pornographic picture or video detection method based on convolutional neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN104182735A (en) |
Cited By (19)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105095911A (en) * | 2015-07-31 | 2015-11-25 | 小米科技有限责任公司 | Sensitive picture identification method and apparatus, and server |
CN105808610A (en) * | 2014-12-31 | 2016-07-27 | 中国科学院深圳先进技术研究院 | Internet picture filtering method and device |
WO2016124103A1 (en) * | 2015-02-03 | 2016-08-11 | 阿里巴巴集团控股有限公司 | Picture detection method and device |
CN106060653A (en) * | 2016-06-08 | 2016-10-26 | 武汉鸿瑞达信息技术有限公司 | Method for filtering improper action in live internet video |
CN106446932A (en) * | 2016-08-30 | 2017-02-22 | 上海交通大学 | Machine learning and picture identification-based evolvable prohibited picture batch processing method |
CN106446782A (en) * | 2016-08-29 | 2017-02-22 | 北京小米移动软件有限公司 | Image identification method and device |
CN106503610A (en) * | 2015-09-08 | 2017-03-15 | 阿里巴巴集团控股有限公司 | Video frequency identifying method and device |
CN106874857A (en) * | 2017-01-19 | 2017-06-20 | 腾讯科技(上海)有限公司 | A kind of living body determination method and system based on video analysis |
CN106951848A (en) * | 2017-03-13 | 2017-07-14 | 平安科技(深圳)有限公司 | The method and system of picture recognition |
CN107016356A (en) * | 2017-03-21 | 2017-08-04 | 乐蜜科技有限公司 | Certain content recognition methods, device and electronic equipment |
CN107205162A (en) * | 2016-03-16 | 2017-09-26 | 阿里巴巴集团控股有限公司 | Video frequency identifying method and device |
CN107665333A (en) * | 2017-08-28 | 2018-02-06 | 平安科技(深圳)有限公司 | A kind of indecency image identification method, terminal, equipment and computer-readable recording medium based on convolutional neural networks |
CN107665353A (en) * | 2017-09-15 | 2018-02-06 | 平安科技(深圳)有限公司 | Model recognizing method, device, equipment and computer-readable recording medium based on convolutional neural networks |
CN107871314A (en) * | 2016-09-23 | 2018-04-03 | 商汤集团有限公司 | A kind of sensitive image discrimination method and device |
CN109214426A (en) * | 2018-08-08 | 2019-01-15 | 王新维 | A kind of method and deep neural network model of the detection of object appearance |
CN110222649A (en) * | 2019-06-10 | 2019-09-10 | 北京达佳互联信息技术有限公司 | Video classification methods, device, electronic equipment and storage medium |
CN110418163A (en) * | 2019-08-27 | 2019-11-05 | 北京百度网讯科技有限公司 | Video frame sampling method, device, electronic equipment and storage medium |
WO2020048183A1 (en) * | 2018-09-04 | 2020-03-12 | 上海海事大学 | Vessel type identification method based on coarse-to-fine cascaded convolutional neural network |
CN111199233A (en) * | 2019-12-30 | 2020-05-26 | 四川大学 | Improved deep learning pornographic image identification method |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN1445715A (en) * | 2002-03-15 | 2003-10-01 | 微软公司 | System and method for mode recognising |
CN103544506A (en) * | 2013-10-12 | 2014-01-29 | Tcl集团股份有限公司 | Method and device for classifying images on basis of convolutional neural network |
CN103679185A (en) * | 2012-08-31 | 2014-03-26 | 富士通株式会社 | Convolutional neural network classifier system as well as training method, classifying method and application thereof |
CN103971342A (en) * | 2014-05-21 | 2014-08-06 | 厦门美图之家科技有限公司 | Image noisy point detection method based on convolution neural network |
CN103984959A (en) * | 2014-05-26 | 2014-08-13 | 中国科学院自动化研究所 | Data-driven and task-driven image classification method |
-
2014
- 2014-08-18 CN CN201410406286.3A patent/CN104182735A/en active Pending
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN1445715A (en) * | 2002-03-15 | 2003-10-01 | 微软公司 | System and method for mode recognising |
CN103679185A (en) * | 2012-08-31 | 2014-03-26 | 富士通株式会社 | Convolutional neural network classifier system as well as training method, classifying method and application thereof |
CN103544506A (en) * | 2013-10-12 | 2014-01-29 | Tcl集团股份有限公司 | Method and device for classifying images on basis of convolutional neural network |
CN103971342A (en) * | 2014-05-21 | 2014-08-06 | 厦门美图之家科技有限公司 | Image noisy point detection method based on convolution neural network |
CN103984959A (en) * | 2014-05-26 | 2014-08-13 | 中国科学院自动化研究所 | Data-driven and task-driven image classification method |
Cited By (24)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105808610A (en) * | 2014-12-31 | 2016-07-27 | 中国科学院深圳先进技术研究院 | Internet picture filtering method and device |
WO2016124103A1 (en) * | 2015-02-03 | 2016-08-11 | 阿里巴巴集团控股有限公司 | Picture detection method and device |
CN105989330A (en) * | 2015-02-03 | 2016-10-05 | 阿里巴巴集团控股有限公司 | Picture detection method and apparatus |
US10235603B2 (en) | 2015-07-31 | 2019-03-19 | Xiaomi Inc. | Method, device and computer-readable medium for sensitive picture recognition |
CN105095911B (en) * | 2015-07-31 | 2019-02-12 | 小米科技有限责任公司 | Sensitization picture recognition methods, device and server |
CN105095911A (en) * | 2015-07-31 | 2015-11-25 | 小米科技有限责任公司 | Sensitive picture identification method and apparatus, and server |
CN106503610B (en) * | 2015-09-08 | 2020-05-26 | 阿里巴巴集团控股有限公司 | Video identification method and device |
CN106503610A (en) * | 2015-09-08 | 2017-03-15 | 阿里巴巴集团控股有限公司 | Video frequency identifying method and device |
CN107205162A (en) * | 2016-03-16 | 2017-09-26 | 阿里巴巴集团控股有限公司 | Video frequency identifying method and device |
CN106060653A (en) * | 2016-06-08 | 2016-10-26 | 武汉鸿瑞达信息技术有限公司 | Method for filtering improper action in live internet video |
CN106446782A (en) * | 2016-08-29 | 2017-02-22 | 北京小米移动软件有限公司 | Image identification method and device |
CN106446932A (en) * | 2016-08-30 | 2017-02-22 | 上海交通大学 | Machine learning and picture identification-based evolvable prohibited picture batch processing method |
CN107871314A (en) * | 2016-09-23 | 2018-04-03 | 商汤集团有限公司 | A kind of sensitive image discrimination method and device |
CN106874857A (en) * | 2017-01-19 | 2017-06-20 | 腾讯科技(上海)有限公司 | A kind of living body determination method and system based on video analysis |
CN106874857B (en) * | 2017-01-19 | 2020-12-01 | 腾讯科技(上海)有限公司 | Living body distinguishing method and system based on video analysis |
CN106951848A (en) * | 2017-03-13 | 2017-07-14 | 平安科技(深圳)有限公司 | The method and system of picture recognition |
CN107016356A (en) * | 2017-03-21 | 2017-08-04 | 乐蜜科技有限公司 | Certain content recognition methods, device and electronic equipment |
CN107665333A (en) * | 2017-08-28 | 2018-02-06 | 平安科技(深圳)有限公司 | A kind of indecency image identification method, terminal, equipment and computer-readable recording medium based on convolutional neural networks |
CN107665353A (en) * | 2017-09-15 | 2018-02-06 | 平安科技(深圳)有限公司 | Model recognizing method, device, equipment and computer-readable recording medium based on convolutional neural networks |
CN109214426A (en) * | 2018-08-08 | 2019-01-15 | 王新维 | A kind of method and deep neural network model of the detection of object appearance |
WO2020048183A1 (en) * | 2018-09-04 | 2020-03-12 | 上海海事大学 | Vessel type identification method based on coarse-to-fine cascaded convolutional neural network |
CN110222649A (en) * | 2019-06-10 | 2019-09-10 | 北京达佳互联信息技术有限公司 | Video classification methods, device, electronic equipment and storage medium |
CN110418163A (en) * | 2019-08-27 | 2019-11-05 | 北京百度网讯科技有限公司 | Video frame sampling method, device, electronic equipment and storage medium |
CN111199233A (en) * | 2019-12-30 | 2020-05-26 | 四川大学 | Improved deep learning pornographic image identification method |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN104182735A (en) | Training optimization pornographic picture or video detection method based on convolutional neural network | |
US11176418B2 (en) | Model test methods and apparatuses | |
CN103971342B (en) | A kind of image noise detection method based on convolutional neural networks | |
CN107316066B (en) | Image classification method and system based on multi-channel convolutional neural network | |
CN105657402B (en) | A kind of depth map restoration methods | |
CN109359681B (en) | Field crop pest and disease identification method based on improved full convolution neural network | |
CN107392314A (en) | A kind of deep layer convolutional neural networks method that connection is abandoned based on certainty | |
CN107729819A (en) | A kind of face mask method based on sparse full convolutional neural networks | |
CN104036474A (en) | Automatic adjustment method for image brightness and contrast | |
CN110135386B (en) | Human body action recognition method and system based on deep learning | |
CN110287777B (en) | Golden monkey body segmentation algorithm in natural scene | |
CN108564549A (en) | A kind of image defogging method based on multiple dimensioned dense connection network | |
CN111832484A (en) | Loop detection method based on convolution perception hash algorithm | |
CN108960404A (en) | A kind of people counting method and equipment based on image | |
CN105389596A (en) | Method for enabling convolutional neural network to be suitable for recognition of pictures of various sizes | |
CN104200238B (en) | TV station symbol recognition method and TV station symbol recognition device | |
CN105956570B (en) | Smiling face's recognition methods based on lip feature and deep learning | |
CN103150733A (en) | Self-adapting multi-stage weighted median filtering algorithm applied to digital images | |
CN109872326B (en) | Contour detection method based on deep reinforced network jump connection | |
CN113239914B (en) | Classroom student expression recognition and classroom state evaluation method and device | |
CN110348448A (en) | A kind of license plate character recognition method based on convolutional neural networks | |
CN106780446A (en) | It is a kind of to mix distorted image quality evaluating method without reference | |
CN114663392A (en) | Knowledge distillation-based industrial image defect detection method | |
CN105590301A (en) | Impulse noise elimination method of self-adaption normal-inclined double cross window mean filtering | |
CN102750700A (en) | Fast robust fuzzy C-means image segmentation method combining neighborhood information |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
WD01 | Invention patent application deemed withdrawn after publication | ||
WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20141203 |