WO2019041406A1 - Indecent picture recognition method, terminal and device, and computer-readable storage medium - Google Patents

Abstract

A convolutional-neural-network-based indecent picture recognition method, terminal (100) and device (200), and a computer-readable storage medium. The method comprises: performing preprocessing on data of an obtained picture to be detected (S101); respectively inputting the preprocessed data of the picture to be detected into a plurality of preset detection models to correspondingly obtain a plurality of different confidence degrees (S102); calculating all the confidence degrees according to a preset linear weighted ratio to obtain a picture confidence degree (S103); determining whether the picture confidence degree is greater than a preset confidence degree or not (S104); if the picture confidence degree is greater than the preset confidence degree, determining that the data of the picture to be detected is indecent picture data (S105). The method can automatically analyze the inputted pictures to be detected and perform classification and identification on the pictures, so that the pornography identification accuracy is greatly improved and the pornography identification cost is lowered.

Description

 Indecent picture recognition method, terminal, device and computer readable storage medium

This application is required to be submitted to the China Patent Office on August 28, 2017, the application number is CN 2017107527832, and the invention name is “a method for identifying indecent pictures based on convolutional neural networks, terminals, devices and computer readable storage media”. The priority of the Chinese Patent Application, the entire contents of which is incorporated herein by reference.

Technical field

The present invention relates to the field of computer technologies, and in particular, to an indecent picture recognition method based on a convolutional neural network, a terminal, a device, and a computer readable storage medium.

Background technique

With the explosive growth of network data and the increasing personalization needs of people, people can obtain massive amounts of data through the network every day. While enriching their spiritual life, they are exposed to massive information and are more exposed to all kinds of pornographic information. Information must be filtered and displayed to the audience, especially children, so the yellow has always been an extremely important part of information filtering. The traditional artificial yellowing has high requirements for practitioners. The Huangshi is under the dual psychological and physiological pressures. The labor cost of the enterprise is too high, and the audit efficiency is low. It cannot be applied to scenarios that require large-scale audits. At present, deep learning algorithms such as convolutional neural networks are widely used in image recognition, and have achieved remarkable results in reducing network model complexity and improving image data processing capabilities, but existing image recognition through convolutional neural networks. There is still a certain degree of error in the results obtained.

Summary of the invention

Embodiments of the present invention provide an indecent picture recognition method, a terminal, a device, and a computer readable storage medium based on a convolutional neural network, which can automatically analyze an input picture and classify and identify the picture, thereby greatly improving the accuracy of the yellowing. Reduce the cost of yellowing.

In one aspect, an embodiment of the present invention provides an indecent picture recognition method based on a convolutional neural network, the method comprising:

Pre-processing the acquired image data to be tested;

The pre-processed picture data to be tested is respectively input into a plurality of preset detection models to correspondingly obtain a plurality of different confidence levels;

Calculating all the confidences according to a preset linear weighting ratio to obtain a picture confidence, wherein the preset linear weighting ratio includes a plurality of ratios, the number of the ratios being the same as the number of the detection models, And each ratio corresponds to a different detection model;

Determining whether the picture confidence is greater than a preset reliability;

If the picture confidence is greater than the preset reliability, the picture data to be tested is determined to be indecent picture data.

On the other hand, an embodiment of the present invention further provides an indecent picture recognition terminal based on a convolutional neural network, where the terminal includes:

a pre-processing unit, configured to perform pre-processing on the acquired image data to be tested;

a detecting unit, configured to separately input the pre-processed picture data to be input into multiple preset detection models to correspondingly obtain a plurality of different confidence levels;

a confidence calculation unit, configured to calculate all confidence levels according to a preset linear weighting ratio to obtain a picture confidence, wherein the preset linear weighting ratio includes a plurality of ratios, and the ratio of the ratios The number of detection models is the same, and each ratio corresponds to a different detection model;

a determining unit, configured to determine whether the picture confidence is greater than a preset reliability;

The determining unit is configured to determine that the picture data to be tested is indecent picture data if the picture confidence is greater than a preset reliability.

In another aspect, an embodiment of the present invention further provides an indecent picture recognition device based on a convolutional neural network, the device comprising:

a memory for storing a program that implements indecent picture recognition;

a processor for running a program for realizing indecent picture recognition stored in the memory to perform the following operations:

Pre-processing the acquired image data to be tested;

The pre-processed picture data to be tested is respectively input into a plurality of preset detection models to correspondingly obtain a plurality of different confidence levels;

Calculating all the confidences according to a preset linear weighting ratio to obtain a picture confidence, wherein the preset linear weighting ratio includes a plurality of ratios, the number of the ratios being the same as the number of the detection models, And each ratio corresponds to a different detection model;

Determining whether the picture confidence is greater than a preset reliability;

If the picture confidence is greater than the preset reliability, the picture data to be tested is determined to be indecent picture data.

In still another aspect, an embodiment of the present invention further provides a computer readable storage medium storing one or more programs, the one or more programs being executable by one or more processors Execute to implement the following steps:

Pre-processing the acquired image data to be tested;

The pre-processed picture data to be tested is respectively input into a plurality of preset detection models to correspondingly obtain a plurality of different confidence levels;

Calculating all the confidences according to a preset linear weighting ratio to obtain a picture confidence, wherein the preset linear weighting ratio includes a plurality of ratios, the number of the ratios being the same as the number of the detection models, And each ratio corresponds to a different detection model;

Determining whether the picture confidence is greater than a preset reliability;

If the picture confidence is greater than the preset reliability, the picture data to be tested is determined to be indecent picture data.

In summary, the present invention has the following beneficial effects: the embodiment of the present invention performs pre-processing on the acquired image data to be tested; and inputs the pre-processed image data to be input into multiple preset detection models to obtain correspondingly a different confidence; calculating all the confidences according to the preset linear weighting ratio to obtain a picture confidence; if the picture confidence is greater than the preset reliability, determining that the picture data to be tested is an indecent picture Data, automatically analyze the input picture to be tested and classify and identify the picture, which greatly improves the accuracy of the yellowing and reduces the cost of the yellowing. In addition, the multiple training and verification of the convolutional neural network can further reduce the error. And improve the recognition accuracy of the obtained detection model.

DRAWINGS

In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly described below. It is obvious that the drawings in the following description are some embodiments of the present invention. For the ordinary technicians, other drawings can be obtained based on these drawings without any creative work.

1 is a schematic flow chart of an indecent picture recognition method based on a convolutional neural network according to an embodiment of the present invention;

2 is another schematic flowchart of an indecent picture recognition method based on a convolutional neural network according to an embodiment of the present invention;

3 is another schematic flowchart of an indecent picture recognition method based on a convolutional neural network according to an embodiment of the present invention;

4 is another schematic flowchart of an indecent picture recognition method based on a convolutional neural network according to an embodiment of the present invention;

FIG. 5 is a schematic flowchart of an indecent picture recognition method based on a convolutional neural network according to another embodiment of the present invention; FIG.

FIG. 6 is a schematic block diagram of an indecent picture recognition terminal based on a convolutional neural network according to an embodiment of the present invention; FIG.

FIG. 7 is another schematic block diagram of an indecent picture recognition terminal based on a convolutional neural network according to an embodiment of the present invention; FIG.

FIG. 8 is another schematic block diagram of an indecent picture recognition terminal based on a convolutional neural network according to an embodiment of the present invention; FIG.

FIG. 9 is another schematic block diagram of an indecent picture recognition terminal based on a convolutional neural network according to an embodiment of the present invention; FIG.

FIG. 10 is a schematic structural diagram of an indecent picture recognition device based on a convolutional neural network according to an embodiment of the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention are clearly and completely described in the following with reference to the accompanying drawings in the embodiments of the present invention. It is obvious that the described embodiments are a part of the embodiments of the present invention, but not all embodiments. Based on the embodiments of the present invention, those skilled in the art are not creative

All other embodiments obtained under the premise of the invention are within the scope of the invention.

It will be understood that the terms "comprise" and "the" when used in the specification and the appended claims "Comprising" indicates the existence of the described features, integers, steps, operations, elements and/or components, but does not exclude the presence of one or more other features, integers, steps, operations, elements, components and/or combinations thereof Add to.

It is also to be understood that the terminology of the present invention is to be construed as a As made in the specification of the invention and the appended claims

As used in the singular, the s

Please refer to FIG. 1. FIG. 1 is a schematic flow chart of an indecent picture recognition method based on a convolutional neural network according to an embodiment of the present invention. The method can be run on terminals such as smart phones (such as Android phones, IOS phones, etc.), tablets, laptops, and smart devices. The method of the invention can automatically analyze the input picture to be tested and classify and identify the picture, thereby greatly improving the accuracy of the yellowing and reducing the cost of the yellowing. As shown in FIG. 1, the method includes steps S101 to S105.

S101: Perform pre-processing on the acquired picture data to be tested.

In the embodiment of the present invention, the picture data to be tested may be a regular picture, or may be a picture obtained by extracting a video key frame from the video data. In order to improve the accuracy of the classification detection, it is necessary to perform corresponding processing on the picture data to be measured.

Specifically, as shown in FIG. 2, step S101 includes steps S201 to S202.

S201: Perform feature enhancement on the acquired image data to be tested.

In the embodiment of the present invention, in order to facilitate feature recognition, feature acquisition of the acquired image data to be tested may be performed first. Specifically, as shown in FIG. 3, step S201 includes steps S301 to S303.

S301. Adjust the acquired picture data to be tested and all preset picture data to 256*256 RGB pictures.

In the embodiment of the present invention, all preset picture data can be adjusted to 256*256 RGB pictures. Similarly, when using the detection model for detection, the used pictures must also be RGB pictures of the same format. That is to say, the acquired image data to be tested needs to be adjusted to 256*256 RGB pictures.

S302. Calculate a pixel average value of all preset picture data after being adjusted to an RGB picture.

In the embodiment of the present invention, in order to make the feature of the picture data to be measured more prominent, to improve the accuracy of the classification detection result, it is necessary to calculate the pixel average value of all the preset picture data after being adjusted to the RGB picture, that is, obtain a The so-called average map.

S303. Subtract the pixel average value from the pixels of the image data to be measured adjusted to the RGB picture to obtain the picture data to be tested after the feature enhancement.

In the embodiment of the present invention, by subtracting the average value of the pixel, the image data to be tested after the feature enhancement corresponding to the image data to be measured adjusted to the RGB image can be obtained.

S202. The feature-enhanced picture data to be tested is divided according to a preset rule to obtain one or more sub-pictures.

Wherein, all the sub-pictures are input into each detection model, so that each detection model correspondingly obtains one or more sub-confidences, and determines that the maximum sub-confidence corresponding to each detection model is the confidence level corresponding to the detection model.

In this embodiment, the image data to be tested after feature enhancement may be divided into one or more sub-pictures of 224*224. Of course, it may be divided into one or more sub-pictures of other sizes according to actual conditions, so that To better extract features and reduce the interference of incoherent features. Similarly, in the process of training the convolutional neural network, all the preset picture data are correspondingly processed correspondingly, thereby increasing the number of trained picture data and further improving the accuracy of the subsequent detection model.

Since one or more sub-pictures are input to each detection model, corresponding to one or more sub-confidences, it is necessary to determine a maximum value of one or more sub-confidences, thereby determining the maximum value as the detection model. Corresponding confidence.

S102: Input the pre-processed picture data to be tested into multiple preset detection models to correspondingly obtain a plurality of different confidence levels.

In the embodiment of the present invention, the pre-processed picture data to be tested is input to each preset detection model, so that each preset detection model can obtain a confidence level.

S103. Calculate all the confidence levels according to a preset linear weighting ratio to obtain a picture confidence, wherein the preset linear weighting ratio includes a plurality of ratios, the number of the ratios and the number of the detection models. The same, and each ratio corresponds to a different detection model.

In the embodiment of the present invention, the preset linear weighting ratio may be obtained by training the related algorithm, or obtaining the optimal linear weighting ratio according to the method of repeated experiments. For example, it is possible to comprehensively analyze the confidence obtained in each detection model for a plurality of indecent images, and then confirm an optimal linear weighting ratio. In general, when there are two detection models, the linear weighting ratio can be chosen to be 1:1, that is, the two confidence levels of the two detection models are each half, so that the final image confidence is obtained.

S104. Determine whether the picture confidence is greater than a preset reliability.

In the embodiment of the present invention, if the picture confidence is not greater than the preset reliability, it is known that the picture data to be tested is normal picture data. The preset reliability can be set correspondingly according to actual conditions. For example, when the pre-set reliability is 0.8, if the picture confidence is less than or equal to 0.8, the picture data to be tested is normal picture data.

S105. If the picture confidence is greater than the preset reliability, determine that the picture data to be tested is indecent picture data.

In the embodiment of the present invention, when the picture confidence is greater than the preset reliability, the picture data to be tested may be known as indecent picture data. For details, when the pre-set reliability is 0.8, if the picture confidence is greater than 0.8, the picture data to be tested is indecent picture data.

It can be seen that, in the embodiment of the present invention, the obtained image data to be tested is preprocessed; the preprocessed image data to be tested is respectively input into a plurality of preset detection models to correspondingly obtain a plurality of different confidence levels; Setting a linear weighting ratio to calculate all the confidence levels to obtain a picture confidence; if the picture confidence is greater than the pre-set reliability, determining that the picture data to be tested is indecent picture data, the input image can be automatically analyzed and Classification and identification of images greatly improve the accuracy of yellowing and reduce the cost of yellowing. In addition, multiple training and verification of convolutional neural networks can further reduce errors and improve the recognition accuracy of the resulting detection models. .

As shown in FIG. 4, as a preferred embodiment, step S101 in the method for identifying indecent picture based on the convolutional neural network according to the embodiment of the present invention further includes separately training a plurality of different volumes by using preset picture data. The neural network is obtained to obtain a plurality of corresponding detection models, that is, specifically including the following steps S401 to S405:

S401. The preset picture data is divided into a training set and a verification set.

In the embodiment of the present invention, the preset picture data may include indeterminate classified image data and normal picture data, which may be manually classified and filtered, and the two types of data are used as two scenes as input data to be convolved. The neural network performs learning classification to obtain a detection model that can determine whether the picture is an indecent picture. The training set includes several normal pictures as well as indecent pictures, and the verification set also includes several normal pictures as well as indecent pictures. The training set is used for routine training of the convolutional neural network, and the verification set is used for corresponding classification detection of the model obtained by the trained convolutional neural network. In order to improve the accuracy of recognition, the training set and the verification set are needed. The corresponding pre-processing, such as feature enhancement, is performed before the convolutional neural network can be input for training.

S402. Train the plurality of different convolutional neural networks with the training set to obtain a plurality of corresponding intermediate models.

In the embodiment of the present invention, Convolutional Neural Network (Convolutional Neural Network) Network, CNN) is a feedforward neural network whose artificial neurons can respond to a surrounding area of a part of the coverage and perform well for large image processing. Different convolutional neural networks include different hierarchical structures.

Through the training set, a plurality of different convolutional neural networks can be separately trained, that is, the pictures in the training set and the classification corresponding to the picture are input into a plurality of different convolutional neural networks for training, and a plurality of corresponding intermediate models can be obtained.

S403. Verify all the intermediate models by using the verification set to obtain a plurality of corresponding error sets, wherein each error set includes at least one error sample.

In the embodiment of the present invention, the image in the verification set is input into the intermediate model for classification detection to obtain the classification detection result, and when the classification detection result is inconsistent with the preset classification of the picture, the picture is identified as an error sample. Specifically, all the pictures whose classification detection result is inconsistent with the pre-classification of the picture can be classified as an error set. In summary, each error set can include at least one error sample.

S404. If the number of the error concentration error samples is greater than or equal to the preset threshold, the intermediate model corresponding to the one of the error sets is trained by using one of the error sets to obtain a corresponding new intermediate model.

In the embodiment of the present invention, if the number of the error concentration error samples is greater than or equal to the preset threshold, it indicates that the error rate of the result of the classification detection is within an unacceptable range, and the error needs to be utilized at this time. The intermediate model corresponding to one of the error sets is trained to obtain a corresponding new intermediate model, thereby further improving the accuracy of the classification detection of the intermediate model.

S405. Verify the new intermediate model again by using the verification set until the number of error concentration error samples is less than a preset threshold, and determine that the new intermediate model at this time is a corresponding preset detection model.

In the embodiment of the present invention, after obtaining a new intermediate model, it is necessary to re-verify the verification set to obtain a new classification detection result, and determine whether the number of error samples in the error set at this time is less than a preset threshold. If the number of error samples in the error set is less than the preset threshold, then it can be determined that the new intermediate model at this time is the corresponding preset detection model. And if the number of error samples in the error set is greater than or equal to the preset threshold, then step S404 may be returned.

In addition, as another preferred embodiment, the number of convolutional neural networks may be two, including a first convolutional neural network having an eight-layer structure and a second convolutional neural network having a twenty-two-layer structure. The first convolutional neural network includes five convolutional layers, two fully connected layers, and one probability statistical layer. The convolutional layer of each layer in the first convolutional neural network can filter the input picture data into a two-dimensional vector through the convolution kernel, and separately calculate its parameters in the training phase, and the full connection layer will input and weight. The vector is multiplied, so the neurons in the next layer are all connected with the neurons in the previous layer. All neurons are accelerated by the activation function, and the probability and statistics layer is used to judge the indecent picture and the normal picture. The second convolutional neural network is deeper than the first convolutional neural network layer, and the Inception structure consisting of two convolutional layers is used to widen the network, and the structure is repeated many times, and then the first convolution is performed. A similar step in the neural network outputs a test result that can be classified.

For details, refer to FIG. 5. FIG. 5 is a schematic flowchart of a method for recognizing an indecent picture based on a convolutional neural network according to another embodiment of the present invention. The method can be run on terminals such as smart phones (such as Android phones, IOS phones, etc.), tablets, laptops, and smart devices. The method mainly analyzes the input picture data and classifies and discriminates the picture, which greatly improves the accuracy of the yellowing and reduces the cost of the yellowing. As shown in FIG. 5, steps S501 to S506 of the method.

S501. The first convolutional neural network and the second convolutional neural network are respectively trained by using preset image data to obtain a corresponding first detection model and second detection model.

S502: Perform pre-processing on the acquired picture data to be tested.

S503. The preprocessed picture data to be tested is input to the first detection model and the second detection model, respectively, to correspondingly obtain the first confidence and the second confidence.

S504: Calculate the first confidence and the second confidence according to a preset linear weighting ratio to obtain a picture confidence, wherein the preset linear weighting ratio includes two ratios, and the ratio of the ratios The number of detection models is the same, and each ratio corresponds to a different detection model.

S505. Determine whether the picture confidence is greater than a preset reliability. When the picture confidence is not greater than the preset reliability, the picture data to be tested may be determined to be normal picture data.

S506. If the picture confidence is greater than the preset reliability, determine that the picture data to be tested is indecent picture data.

A person skilled in the art can understand that all or part of the process of implementing the above embodiments can be completed by a computer program to instruct related hardware, and the program can be stored in a computer readable storage medium. When executed, the flow of an embodiment of the methods as described above may be included. The storage medium may be a magnetic disk, an optical disk, or a read-only storage memory (Read-Only) Memory, ROM), etc.

Referring to FIG. 6 , corresponding to the inconsistency picture recognition method based on the convolutional neural network, the embodiment of the present invention further provides an indecent picture recognition terminal based on a convolutional neural network, where the terminal 100 includes: a pre-processing unit 101 The detecting unit 102, the confidence calculating unit 103, the determining unit 104, and the determining unit 105.

The pre-processing unit 101 is configured to perform pre-processing on the acquired picture data to be tested.

The detecting unit 102 is configured to input the pre-processed picture data to be tested into a plurality of preset detection models to correspondingly obtain a plurality of different confidence levels.

The confidence calculation unit 103 is configured to calculate all confidence levels according to a preset linear weight ratio to obtain a picture confidence, wherein the preset linear weight ratio includes a plurality of ratios, and the number of the ratios The number is the same as the number of detection models, and each ratio corresponds to a different detection model.

The determining unit 104 is configured to determine whether the picture confidence is greater than a preset reliability.

The determining unit 105 is configured to determine that the picture data to be tested is indecent picture data if the picture confidence is greater than a pre-set confidence.

As shown in FIG. 7, the terminal 100 further includes a classification unit 100a, a first training unit 100b, a verification unit 100c, and a second training unit 100d.

The classification unit 100a is configured to divide preset picture data into a training set and a verification set.

The first training unit 100b is configured to train a plurality of different convolutional neural networks by using the training set to obtain a plurality of corresponding intermediate models.

The verification unit 100c is configured to verify all the intermediate models by using the verification set to obtain a plurality of corresponding error sets, wherein each error set includes at least one error sample.

The second training unit 100d is configured to train an intermediate model corresponding to one of the error sets to obtain a correspondence if the number of error concentration errors in one of the errors is greater than or equal to a preset threshold. New intermediate model.

The verification unit 100c is further configured to verify the new intermediate model again by using the verification set until the number of error concentration error samples is less than a preset threshold, and determine a new intermediate model at this time. The model is detected for the corresponding preset.

As shown in FIG. 8, the pre-processing unit 101 includes a strengthening unit 1011 and a dividing unit 1012.

The strengthening unit 1011 is configured to perform feature enhancement on the acquired picture data to be tested.

The dividing unit 1012 is configured to divide the feature-intensified picture data to be obtained according to a preset rule to obtain one or more sub-pictures.

The detecting unit 102 is further configured to input all the sub-pictures into each detection model, so that each detection model correspondingly obtains one or more sub-confidences, and determines that the maximum sub-confidence corresponding to each detection model is The confidence level corresponding to the detection model.

As shown in FIG. 9, the strengthening unit 1011 includes an adjustment unit 1011a, an average value calculation unit 1011b, and a target calculation unit 1011c.

The adjusting unit 1011a is configured to adjust the acquired picture data to be tested and all preset picture data to 256*256 RGB pictures.

The average value calculating unit 1011b is configured to calculate a pixel average value of all preset picture data after being adjusted to an RGB picture.

The target calculation unit 1011c is configured to subtract the pixel average value from the pixels of the image data to be measured adjusted to the RGB picture to obtain the picture data to be tested after feature enhancement.

FIG. 10 is a schematic structural diagram of an indecent picture recognition device based on a convolutional neural network according to the present invention. As shown in FIG. 10, the device 200 can include an input device 201, an output device 202, a transceiver device 203, a memory 204, and a processor 205, where:

The input device 201 is configured to receive input data of an external access control device. In a specific implementation, the input device 201 in the embodiment of the present invention may include a keyboard, a mouse, a photoelectric input device, and a sound input.

Into the device, touch input device, scanner, etc.

The output device 202 is configured to output output data of the access control device to the outside. In the specific implementation,

The output device 202 described in the embodiments of the present invention may include a display, a speaker, a printer, and the like.

The transceiver device 203 is configured to send data to or receive data from other devices through a communication link. In a specific implementation, the transceiver device 203 of the embodiment of the present invention may include a transceiver device such as a radio frequency antenna.

The memory 204 is configured to store a program that implements indecent picture recognition. The memory 204 of an embodiment of the invention may be a system memory such as volatile (such as RAM), non-volatile (such as ROM, flash memory, etc.), or a combination of both. In a specific implementation, the memory 204 of the embodiment of the present invention may also be an external memory outside the system, such as a magnetic disk, an optical disk, a magnetic tape, or the like.

The processor 205 is configured to run a program for implementing indecent picture recognition stored in the memory 204 to perform the following operations:

Pre-processing the acquired image data to be tested;

The pre-processed picture data to be tested is respectively input into a plurality of preset detection models to correspondingly obtain a plurality of different confidence levels;

Calculating all the confidences according to a preset linear weighting ratio to obtain a picture confidence, wherein the preset linear weighting ratio includes a plurality of ratios, the number of the ratios being the same as the number of the detection models, And each ratio corresponds to a different detection model;

Determining whether the picture confidence is greater than a preset reliability;

If the picture confidence is greater than the preset reliability, the picture data to be tested is determined to be indecent picture data.

Further, it is also used to perform the following operations:

The preset picture data is divided into a training set and a verification set;

Training a plurality of different convolutional neural networks with the training set to obtain a plurality of corresponding intermediate models;

Verifying all intermediate models with the verification set to obtain a plurality of corresponding error sets, wherein each error set includes at least one error sample;

If the number of the error concentration error samples is greater than or equal to the preset threshold, the intermediate model corresponding to the one of the error sets is trained by using one of the error sets to obtain a corresponding new intermediate model;

The new intermediate model is verified again by using the verification set until the number of error concentration error samples is less than a preset threshold, and it is determined that the new intermediate model at this time is a corresponding preset detection model.

Further, the preprocessing the acquired image data to be tested includes:

Performing feature enhancement on the acquired image data to be tested;

Dividing the image data to be tested after feature enhancement according to a preset rule to obtain one or more sub-pictures;

Wherein, all the sub-pictures are input into each detection model, so that each detection model correspondingly obtains one or more sub-confidences, and determines that the maximum sub-confidence corresponding to each detection model is the confidence level corresponding to the detection model.

Further, performing feature enhancement on the acquired image data to be tested includes:

Adjusting the acquired image data to be tested and all preset image data to 256*256 RGB images;

Calculating a pixel average of all preset picture data after being adjusted to an RGB picture;

The pixels of the image to be tested that are adjusted to be RGB pictures are subtracted from the average of the pixels to obtain the picture data to be tested after feature enhancement.

Further, the number of convolutional neural networks is two, including a first convolutional neural network having an eight-layer structure and a second convolutional neural network having a twenty-two-layer structure, wherein the first convolutional nerve The network consists of five convolutional layers, two fully connected layers, and a probabilistic layer.

It will be understood by those skilled in the art that the embodiment of the device for identifying indecent picture based on convolutional neural network shown in FIG. 10 does not constitute a limitation on the specific configuration of the device based on the inconsistency of the convolutional neural network. In other embodiments, an apparatus based on inconsistency picture recognition of a convolutional neural network may include more or fewer components than illustrated, or some components may be combined, or different component arrangements. For example, in some embodiments, the device based on the inconsistency picture recognition of the convolutional neural network may include only the memory and the processor. In such an embodiment, the structure and function of the memory and the processor are the same as the embodiment shown in FIG. Consistent, no longer repeat them here.

The present invention provides a computer readable storage medium having stored one or more programs, the one or more programs being executable by one or more processors to implement the following steps:

Pre-processing the acquired image data to be tested;

The pre-processed picture data to be tested is respectively input into a plurality of preset detection models to correspondingly obtain a plurality of different confidence levels;

Calculating all the confidences according to a preset linear weighting ratio to obtain a picture confidence, wherein the preset linear weighting ratio includes a plurality of ratios, the number of the ratios being the same as the number of the detection models, And each ratio corresponds to a different detection model;

Determining whether the picture confidence is greater than a preset reliability;

If the picture confidence is greater than the preset reliability, the picture data to be tested is determined to be indecent picture data.

Further, the following steps are also implemented:

The preset picture data is divided into a training set and a verification set;

Training a plurality of different convolutional neural networks with the training set to obtain a plurality of corresponding intermediate models;

Verifying all intermediate models with the verification set to obtain a plurality of corresponding error sets, wherein each error set includes at least one error sample;

If the number of the error concentration error samples is greater than or equal to the preset threshold, the intermediate model corresponding to the one of the error sets is trained by using one of the error sets to obtain a corresponding new intermediate model;

The new intermediate model is again verified by using the verification set until the number of error concentration error samples is less than a preset threshold, and it is determined that the new intermediate model at this time is a corresponding detection model.

Further, the preprocessing the acquired image data to be tested includes:

Performing feature enhancement on the acquired image data to be tested;

Dividing the image data to be tested after feature enhancement according to a preset rule to obtain one or more sub-pictures;

Wherein, all the sub-pictures are input into each detection model, so that each detection model correspondingly obtains one or more sub-confidences, and determines that the maximum sub-confidence corresponding to each detection model is the confidence level corresponding to the detection model.

Further, performing feature enhancement on the acquired image data to be tested includes:

Adjusting the acquired image data to be tested and all preset image data to 256*256 RGB images;

Calculating a pixel average of all preset picture data after being adjusted to an RGB picture;

The pixels of the image to be tested that are adjusted to be RGB pictures are subtracted from the average of the pixels to obtain the picture data to be tested after feature enhancement.

Further, the number of convolutional neural networks is two, including a first convolutional neural network having an eight-layer structure and a second convolutional neural network having a twenty-two-layer structure, wherein the first convolutional nerve The network consists of five convolutional layers, two fully connected layers, and a probabilistic layer.

The foregoing storage medium of the present invention includes: a magnetic disk, an optical disk, and a read-only storage memory (Read-Only) Memory, ROM, and other media that can store program code.

The units in all embodiments of the present invention may be implemented by a general-purpose integrated circuit such as a CPU (Central Processing) Unit, central processing unit, or through ASIC (Application Specific Integrated

Circuit, ASIC) to achieve.

The steps in the method of the embodiment of the present invention may be sequentially adjusted, merged, and deleted according to actual needs.

The units in the terminal in the embodiment of the present invention may be combined, divided, and deleted according to actual needs.

A person skilled in the art can clearly understand that, for the convenience and brevity of the description, the specific working process of the device, the terminal, and the unit described above may be referred to the corresponding process in the foregoing method embodiments, and details are not described herein again.

The above is only the specific embodiment of the present invention, but the scope of the present invention is not limited thereto, and any equivalent person can be easily conceived within the technical scope of the present invention by any person skilled in the art. Modifications or substitutions are intended to be included within the scope of the invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims (20)

1. An indecent picture recognition method based on a convolutional neural network, characterized in that the method comprises:

   Pre-processing the acquired image data to be tested;

   The pre-processed picture data to be tested is respectively input into a plurality of preset detection models to correspondingly obtain a plurality of different confidence levels;

   Calculating all the confidences according to a preset linear weighting ratio to obtain a picture confidence, wherein the preset linear weighting ratio includes a plurality of ratios, the number of the ratios being the same as the number of the detection models, And each ratio corresponds to a different detection model;

   Determining whether the picture confidence is greater than a preset reliability;

   If the picture confidence is greater than the preset reliability, the picture data to be tested is determined to be indecent picture data.
2. The method according to claim 1, wherein the pre-processing of the acquired image data to be tested comprises:

   The preset picture data is divided into a training set and a verification set;

   Training a plurality of different convolutional neural networks with the training set to obtain a plurality of corresponding intermediate models;

   Verifying all intermediate models with the verification set to obtain a plurality of corresponding error sets, wherein each error set includes at least one error sample;

   If the number of the error concentration error samples is greater than or equal to the preset threshold, the intermediate model corresponding to the one of the error sets is trained by using one of the error sets to obtain a corresponding new intermediate model;

   The new intermediate model is verified again by using the verification set until the number of error concentration error samples is less than a preset threshold, and it is determined that the new intermediate model at this time is a corresponding preset detection model.
3. The method according to claim 1, wherein the preprocessing the acquired image data to be tested comprises:

   Performing feature enhancement on the acquired image data to be tested;

   Dividing the image data to be tested after feature enhancement according to a preset rule to obtain one or more sub-pictures;

   Wherein, all the sub-pictures are input into each detection model, so that each detection model correspondingly obtains one or more sub-confidences, and determines that the maximum sub-confidence corresponding to each detection model is the confidence level corresponding to the detection model.
4. The method according to claim 3, wherein the characterizing the acquired image data to be tested comprises:

   Adjusting the acquired image data to be tested and all preset image data to 256*256 RGB images;

   Calculating a pixel average of all preset picture data after being adjusted to an RGB picture;

   The pixels of the image to be tested that are adjusted to be RGB pictures are subtracted from the average of the pixels to obtain the picture data to be tested after feature enhancement.
5. The method of claim 1 wherein the number of convolutional neural networks is two, comprising a first convolutional neural network having an eight-layer structure and a second convolutional nerve having a twenty-two-layer structure A network, wherein the first convolutional neural network comprises five convolutional layers, two fully connected layers, and one probability statistical layer.
6. An indecent picture recognition terminal based on a convolutional neural network, wherein the terminal comprises:

   a pre-processing unit, configured to perform pre-processing on the acquired image data to be tested;

   a detecting unit, configured to separately input the pre-processed picture data to be input into multiple preset detection models to correspondingly obtain a plurality of different confidence levels;

   a confidence calculation unit, configured to calculate all confidence levels according to a preset linear weighting ratio to obtain a picture confidence, wherein the preset linear weighting ratio includes a plurality of ratios, and the ratio of the ratios The number of detection models is the same, and each ratio corresponds to a different detection model;

   a determining unit, configured to determine whether the picture confidence is greater than a preset reliability;

   The determining unit is configured to determine that the picture data to be tested is indecent picture data if the picture confidence is greater than a preset reliability.
7. The terminal according to claim 6, wherein the terminal further comprises:

   a classification unit, configured to divide preset picture data into a training set and a verification set;

   a first training unit, configured to use the training set to train a plurality of different convolutional neural networks to obtain a plurality of corresponding intermediate models;

   a verification unit, configured to verify all the intermediate models by using the verification set to obtain a plurality of corresponding error sets, wherein each error set includes at least one error sample;

   a second training unit, configured to: if one of the error concentration error samples is greater than or equal to a preset threshold, use the one of the error sets to train an intermediate model corresponding to the one of the error sets to obtain a corresponding new Intermediate model

   The verification unit is further configured to use the verification set to verify the new intermediate model again, until the number of the error concentration error samples is less than a preset threshold, and determine that the new intermediate model at this time is Corresponding preset detection model.
8. The terminal according to claim 6, wherein the preprocessing unit comprises:

   a strengthening unit, configured to perform feature enhancement on the acquired image data to be tested;

   a dividing unit, configured to divide the image data to be tested after feature enhancement according to a preset rule to obtain one or more sub-pictures;

   The detecting unit is further configured to input all the sub-pictures into each detection model, so that each detection model correspondingly obtains one or more sub-confidences, and determines that the maximum sub-confidence corresponding to each detection model is the Detect the confidence level corresponding to the model.
9. The terminal according to claim 8, wherein the strengthening unit comprises:

   The adjusting unit is configured to adjust the acquired picture data to be tested and all preset picture data to 256*256 RGB pictures.

   The average value calculating unit is configured to calculate a pixel average value of all preset picture data after being adjusted to an RGB picture.

   The target calculation unit is configured to subtract the pixel average value from the pixels of the image data to be measured adjusted to the RGB image to obtain the image data to be tested after the feature enhancement.
10. The terminal according to claim 6, wherein the number of convolutional neural networks is two, including a first convolutional neural network having an eight-layer structure and a second convolutional nerve having a twenty-two-layer structure A network, wherein the first convolutional neural network comprises five convolutional layers, two fully connected layers, and one probability statistical layer.
11. An indecent picture recognition device based on a convolutional neural network, comprising:

    a memory for storing a program that implements indecent picture recognition;

    a processor for running a program for realizing indecent picture recognition stored in the memory to perform the following operations:

    Pre-processing the acquired image data to be tested;

    The pre-processed picture data to be tested is respectively input into a plurality of preset detection models to correspondingly obtain a plurality of different confidence levels;

    Calculating all the confidences according to a preset linear weighting ratio to obtain a picture confidence, wherein the preset linear weighting ratio includes a plurality of ratios, the number of the ratios being the same as the number of the detection models, And each ratio corresponds to a different detection model;

    Determining whether the picture confidence is greater than a preset reliability;

    If the picture confidence is greater than the preset reliability, the picture data to be tested is determined to be indecent picture data.
12. The device according to claim 11, wherein before the pre-processing of the acquired image data to be tested, the method comprises:

    The preset picture data is divided into a training set and a verification set;

    Training a plurality of different convolutional neural networks with the training set to obtain a plurality of corresponding intermediate models;

    Verifying all intermediate models with the verification set to obtain a plurality of corresponding error sets, wherein each error set includes at least one error sample;

    If the number of the error concentration error samples is greater than or equal to the preset threshold, the intermediate model corresponding to the one of the error sets is trained by using one of the error sets to obtain a corresponding new intermediate model;

    The new intermediate model is verified again by using the verification set until the number of error concentration error samples is less than a preset threshold, and it is determined that the new intermediate model at this time is a corresponding preset detection model.
13. The device according to claim 11, wherein the pre-processing of the acquired image data to be tested comprises:

    Performing feature enhancement on the acquired image data to be tested;

    Dividing the image data to be tested after feature enhancement according to a preset rule to obtain one or more sub-pictures;

    Wherein, all the sub-pictures are input into each detection model, so that each detection model correspondingly obtains one or more sub-confidences, and determines that the maximum sub-confidence corresponding to each detection model is the confidence level corresponding to the detection model.
14. The device according to claim 13, wherein the characterizing the acquired image data to be tested comprises:

    Adjusting the acquired image data to be tested and all preset image data to 256*256 RGB images;

    Calculating a pixel average of all preset picture data after being adjusted to an RGB picture;

    The pixels of the image to be tested that are adjusted to be RGB pictures are subtracted from the average of the pixels to obtain the picture data to be tested after feature enhancement.
15. The apparatus according to claim 11, wherein the number of convolutional neural networks is two, including a first convolutional neural network having an eight-layer structure and a second convolutional nerve having a twenty-two-layer structure A network, wherein the first convolutional neural network comprises five convolutional layers, two fully connected layers, and one probability statistical layer.
16. A computer readable storage medium, characterized in that the computer readable storage medium stores one or more programs, the one or more programs being executable by one or more processors to implement the following steps:

    Pre-processing the acquired image data to be tested;

    The pre-processed picture data to be tested is respectively input into a plurality of preset detection models to correspondingly obtain a plurality of different confidence levels;

    Calculating all the confidences according to a preset linear weighting ratio to obtain a picture confidence, wherein the preset linear weighting ratio includes a plurality of ratios, the number of the ratios being the same as the number of the detection models, And each ratio corresponds to a different detection model;

    Determining whether the picture confidence is greater than a preset reliability;

    If the picture confidence is greater than the preset reliability, the picture data to be tested is determined to be indecent picture data.
17. The computer readable storage medium according to claim 16, wherein the pre-processing of the acquired image data to be tested comprises:

    The preset picture data is divided into a training set and a verification set;

    Training a plurality of different convolutional neural networks with the training set to obtain a plurality of corresponding intermediate models;

    Verifying all intermediate models with the verification set to obtain a plurality of corresponding error sets, wherein each error set includes at least one error sample;

    If the number of the error concentration error samples is greater than or equal to the preset threshold, the intermediate model corresponding to the one of the error sets is trained by using one of the error sets to obtain a corresponding new intermediate model;

    The new intermediate model is verified again by using the verification set until the number of error concentration error samples is less than a preset threshold, and it is determined that the new intermediate model at this time is a corresponding preset detection model.
18. The computer readable storage medium according to claim 16, wherein the preprocessing the acquired image data to be tested comprises:

    Performing feature enhancement on the acquired image data to be tested;

    Dividing the image data to be tested after feature enhancement according to a preset rule to obtain one or more sub-pictures;

    Wherein, all the sub-pictures are input into each detection model, so that each detection model correspondingly obtains one or more sub-confidences, and determines that the maximum sub-confidence corresponding to each detection model is the confidence level corresponding to the detection model.
19. The computer readable storage medium according to claim 18, wherein the performing feature enhancement on the acquired image data to be tested comprises:

    Adjusting the acquired image data to be tested and all preset image data to 256*256 RGB images;

    Calculating a pixel average of all preset picture data after being adjusted to an RGB picture;

    The pixels of the image to be tested that are adjusted to be RGB pictures are subtracted from the average of the pixels to obtain the picture data to be tested after feature enhancement.
20. A computer readable storage medium according to claim 16, wherein the number of convolutional neural networks is two, including a first convolutional neural network having an eight-layer structure and a second having a twenty-two-layer structure A second convolutional neural network, wherein the first convolutional neural network comprises five convolutional layers, two fully connected layers, and one probability statistical layer.