TW201329874A - Image recognition system and its operating method - Google Patents

Abstract

Provided is an image recognition system and its operating method primarily used for filtering pornographic images. The system includes at least one terminal recognition module and a cloud learning module. In addition to an image analysis unit, the terminal recognition module further includes a confirmation unit, an error report unit and a first update unit. The cloud learning module includes a second update unit, a learning unit and a cloud database. With the use of error report, the system can be continuously improved so as to more precisely filter pornographic images. Besides, a feature vector capture is employed to calculate and find feature points for being classified, so as to achieve the function of local analysis while analyzing the global images.

Description

Image recognition system and its operation method

The invention relates to an image recognition system and an operation method thereof, in particular to automatically updating and learning, so as to filter erotic images more accurately.

For families with minor children in their homes, in order to avoid the problem of pornography, they usually choose to install blocking software to judge the use of blocking multimedia materials to block pornographic pictures and videos. Some use firewalls to block web pages or IP. For example, Chunghwa Telecom’s erotic goalkeeper.

In the previous patent case, the patent of the Republic of China Patent No. I262416 has already described the use of the erotic image library for the full picture comparison, and the document of the Republic of China Patent Publication No. 201115481 also proposes the use of local analysis. Judge.

However, the use of a firewall way, the operator can easily convert web pages or IP, the effect is not good, in order to filter pornographic pictures may cause other normal web pages or data streams to be blocked, resulting in inconvenience, and the whole picture is compared, At present, it is often compared in a stacking manner, and the chance of error is high. In addition, in a multi-partial manner, since the judgment of the face and the specific organ is complicated, and the problem of the webpage delay or the like is easily caused, the user is criticized. Therefore, there is a need for a system and method capable of improving the conventional technique.

The main object of the present invention is to provide an image recognition system, which is mainly used for filtering pornographic images. The system comprises: at least one terminal identification module installed on a terminal of a user; and a cloud learning module installed on the cloud a cloud server is connected to each terminal identification module through a network, wherein each terminal identification module includes a monitoring unit, an image analysis unit, a classification unit, a confirmation unit, a reward unit, and a first An update unit, the monitoring unit receives a multimedia material from the outside, and extracts at least one image from the multimedia data, the image analysis unit is connected to the monitoring unit, the classification unit and the confirmation unit, and is received by the monitoring unit Determining the at least one image, and determining whether the at least one image is an erotic image according to a classification criterion stored in the classification unit, and if the image is determined to be an erotic image, the image analysis unit transmits the at least one image To the confirmation unit, and if it is determined that it is not an erotic image, output the multimedia material; The unit again determines whether the at least one image is an erotic image, and if it is an erotic image, blocks the multimedia material, if not the erotic image, transmits the at least one to the reward unit, the reward unit and the confirmation unit and the first update unit The connecting is configured to receive the at least one image, and transmit the image to the first update unit, and then transmit the at least one to the cloud learning module that is connected to the first update unit by using the first update unit.

The cloud learning module includes a basic database, a cloud database, a second update unit, an identification unit, and a learning unit. The basic database stores an initial pattern data, and the cloud database is used to create a supplementary pattern. The second update unit is connected to the first update unit, the cloud database, and the identification unit, and configured to receive the at least one image of the first update unit, the identification unit and the cloud database, the second The update unit and the learning unit are connected to determine whether the image analysis unit determines an error. If the image analysis unit determines that the error is determined, the determination unit sends a determination error signal to the learning unit, and transmits the at least one image to the cloud database. And if it is determined that the image analysis unit does not determine the error, sending a determination error-free signal to the second update unit, the learning unit receiving the determination error signal, according to the at least one image in the cloud database Comparing the initial pattern data of the basic database, and generating an update file, and then transmitting the second update unit The first updating unit updates the file transferred to the sorting unit, to reset the classification criterion.

Another object of the present invention is to provide an operation method of an image recognition system, which includes at least one terminal identification module installed in at least one user terminal, and a cloud learning module installed in a cloud server. The operation method includes: an image capturing step, wherein a monitoring unit of the terminal identification module captures at least one image from the external multimedia data; a global image determining step, and an image analyzing unit in each terminal identification module Determining, according to a classification criterion in a classifier, the at least one image to determine whether there is an erotic image; and an image outputting step, when the global image determining step determines that the at least one image does not have an erotic image, a multimedia data output; a manual determining step, when determining that the at least one image has an erotic image in the global image determining step, by using a confirmation unit of each terminal identification module, by a user installing the image recognition system Double confirmation to determine whether the at least one image has an erotic image; a blocking image a step of blocking the multimedia material output when the manual determining step determines that the at least one has an erotic image; and an error reporting step, when the manual determining step determines that the at least one does not have an erotic image, transmitting the at least one image to a first update unit of each terminal identification module is further transmitted to the cloud learning module by the first update unit; a recognition determining step, after the error reporting step, the cloud learning module The second update unit transmits the captured image to an identification unit of the cloud learning module to determine whether it is an error return; and an update cloud database step, if the identification determination step determines that the error is not an error, the identification unit issues a Determining an error signal to a learning unit of the cloud learning module, and transmitting the at least one image to the cloud database to create the supplementary pattern data; and calculating a difference value step, the learning unit according to the supplementary pattern of the cloud database The data is compared with the initial pattern data of a basic database, and a difference value is calculated, the learning list And generating an update file according to the difference value, and transmitting the update file to the classification unit through the second update unit and the first update unit; and an update classification unit step, resetting and storing the classification according to the update file The classification criteria in the unit.

The invention utilizes multiple confirmation mechanisms and error return functions, and can enable the computer mechanism to automatically learn and update the files in the cloud database to make the judgment more accurate. In addition, the present invention uses the feature vector extraction method to calculate The feature points are compared, and in the whole image analysis, the functions of local analysis of color block area ratio, color block distribution, organ characteristics, etc. are simultaneously achieved, and the deficiency of the global image superposition method in the conventional technology is improved. Further, the user database file of the cloud database can achieve the effect of customization when the judgment standards of different countries and families are different.

The implementation of the present invention will be described in more detail below with reference to the drawings and component symbols, so that those skilled in the art can implement the present specification after studying the present specification.

Referring to the first figure, a schematic diagram of a unit of the image recognition system of the present invention. As shown in the first figure, the image recognition system 1 of the present invention is mainly used for filtering color images. The system includes at least one terminal identification module 10 (only one of which is shown in the figure) and a cloud learning module 30. The identification module 10 is installed on the user's terminal. The terminal identification module 10 includes a monitoring unit 11, an image analysis unit 13, a classification unit 15, a confirmation unit 17, a reward unit 19, and a first update unit. 21 and a feedback unit 23.

The monitoring unit 11 receives a multimedia material from an external (for example, the Internet) and extracts at least one image from the multimedia material. The image analyzing unit 13 is connected to the monitoring unit 11, the classifying unit 15 and the confirming unit 17, receives the image captured by the monitoring unit 11, and determines whether the captured image is based on the classification criterion stored in the classifying unit 15. An erotic image, when the erotic image is determined, the captured image is transmitted to the confirmation unit 17, and when it is determined that the image is not erotic, the multimedia material is output; the confirmation unit 17 determines again whether the captured image is erotic If so, the multimedia material is blocked, and if not, the captured image is transmitted to the reward unit 19. The reward unit 19 is connected to the confirmation unit 17 and the first update unit 21, receives the captured image, and transmits the captured image to the first update unit 21, and then transmits the image to the cloud learning module 30 by the first update unit 21. The feedback unit 23 is connected to the reward unit 19 for detecting the multimedia material output by the image analyzing unit 13. When the multimedia material has an erotic image, the image is captured and transmitted to the reward unit 19.

The cloud learning module 30 is installed on a cloud server and is connected to the terminal identification module 10 through a network. The cloud learning module 30 includes a basic database 31, a cloud database 33, and a second updating unit 35. An identification unit 37 and a learning unit 39, the basic database 31 stores initial pattern data, the cloud database 33 is used to create supplementary pattern data, the second updating unit 35 and the first updating unit 21, the cloud database 33 and the identification unit 37 Connecting, receiving the captured image of the first update unit 21, the identification unit 35 is connected to the cloud database 33, the second update unit 35, and the learning unit 39 for determining whether the image analysis unit 13 determines an error. When the image analyzing unit 13 determines an error, a determination error signal is sent to the learning unit 39, and the captured image is transmitted to the cloud database 33. However, if the identification unit 37 determines that the image analyzing unit 13 does not determine the error, A determination error-free signal is sent to the second update unit 35. Upon receiving the determination error signal, the learning unit 39 compares the file of the captured image with the data of the basic database 31 according to the cloud database 33, and generates an update file, and then passes through the second update unit 35 and the first update. The unit 21 transmits to the classification unit 15 to reset the classification judgment reference. Further, the cloud database further establishes a user database to provide different classification criteria for different users. The image recognition system of the present invention can be continuously updated to make the filtered pornographic images more accurate. Further, since each family and each country have different degrees of conservatism, different classification criteria can be generated for the classifiers of the respective terminals. .

Referring to the second and third figures, a flow chart of the method of operation of the image recognition system of the present invention. As shown in the second and third figures, the operation method S1 of the image recognition system of the present invention includes an image capture step S10, a global image determination step S20, a video output step S30, a manual determination step S40, a blocked image output step S50, and an error. The report returns to step S60, the identification determination step S80, the update of the cloud database step S90, the calculation of the difference value step S100, the update of the classification unit step S110, and the determination of the malicious return step S120.

In the image capturing step S10, the monitoring unit 11 captures at least one image from the multimedia data from the outside; in the global image determining step S20, the image analyzing unit 13 analyzes the captured image according to the classification judgment criterion of the classifier 15, It is determined whether or not there is an erotic image. If it is determined that the captured image does not have an erotic image, the image output step S30 is performed to output the multimedia material. If it is determined that the captured image has an erotic image, a manual determination step S40 is performed. The manual determination step S40 is performed by the user who installs the image recognition system by the confirmation unit 17 to determine whether it is an erotic image. If the determination is indeed an erotic image, the blocked image output step S50 blocks the output of the multimedia data. If there is an erotic image, the error returning step S60 is performed, and the captured image is transmitted to the first updating unit 21 and then transmitted to the cloud learning module 30.

After the error reporting step S60, the captured image is transmitted to the second updating unit 35 of the cloud learning module 30 through the first updating unit 21, and in the identification determining step S80, the second updating unit 35 transmits the captured image. The identification unit 37 determines whether the error is a return. If the determination is correct, the identification unit 37 sends a determination error signal to the learning unit 39, and then updates the cloud database to the S90, and transmits the captured image to the The cloud database 33 is used to create the supplementary pattern data to further establish the user database. However, if the identification unit 37 determines that the image analyzing unit 13 has not determined the error, the determination of the malicious return step S120 is performed, and the identification unit 37 issues a determination error-free signal to the first The unit 35 is updated without performing any modified actions.

After updating the cloud database step S90, the calculating difference value step S100 is performed, and the learning unit 39 calculates a difference value according to the supplementary pattern data in the cloud database 33 and the initial pattern data comparison of the basic database 31, and learns a difference value. The unit 39 generates an update file according to the difference value, and transmits the update file to the classification unit 15 through the second update unit 35 and the first update unit 21. Finally, the update classification unit is performed in step S110, and the classification judgment reference stored in the classification unit 15 is reset based on the update file.

Further, as shown in the second figure, the operation method of the image recognition system of the present invention further includes an output monitoring step S70, wherein the output monitoring step S70 is to monitor the output image by the feedback unit 23, and if the output multimedia material is found to have a pornographic pattern, Then, the pattern is captured, and an error returning step S60 is performed.

Referring to the fourth figure, a detailed flowchart of the global image determining step of the present invention. As shown in the fourth figure, the global image determining step S20 includes a scale space limit value detecting step S21, a feature point positioning step S23, a direction assigning step S25, a feature point describing step S27, and a feature vector determining step S29. The scale space limit value detecting step S21 first performs a Gaussian operation on each point of the captured image, and the operation is performed as shown in equation (1):

L(x,y,σ)=G(x,y,σ) I(x, y) [Equation (1)], where L is the processed image, I is the initial image, σ is the standard deviation, and G(x, y, σ) is the Gaussian function, with equation (2) Indicates:

Then, the L value of the two adjacent points of the processed image is subtracted at least twice, and the difference of Gaussian value (DoG) can be calculated, and the Gauss value difference between the two points can be expressed as D(x, y, σ ), and the following equation (3) represents: D(x, y, σ) = (G(x, y, σ ₁ ) - G(x, y, σ ₂ )) I(x,y)=L(x,y,kσ ₁ )-L(x,y,σ ₂ )[Equation (3)], where σ1 is the standard deviation of the first subtraction, and σ2 is the second The standard deviation of the sub-phase subtraction is such that at least one maximum value and at least one minimum value of the Gauss value difference are found.

Next, a feature point locating step S23 is performed, and the maximum value and the minimum value obtained by the scale space limit value detecting step S21 are approximated by the expansion equation of Equation (4) to formulate a plurality of feature points. [Equation (4)], where D(x) is a Gaussian value difference, X = (x, y, σ) T, and T is a transpose. Next, the direction assigning step S25 is performed, and the direction and gradient are calculated for the feature points, the direction θ(x, y) is calculated as the equation (5), and the gradient m(x, y) is calculated as the equation ( 6), and calculate the direction and gradient of the surrounding points around the feature points,

Then, the feature point description step S27 is performed, and the data in the quasi-reference is determined according to the classification in the classifier 15, and the direction and gradient weighting are performed on the relevant position of each feature point to establish a feature vector, and finally the feature of the feature point is established. The vector is subjected to the feature vector determining step S29. If the feature vector value is within a feature vector range, the image is determined to be an erotic image, and the artificial image determining step S40 is performed, and if the feature vector value is outside the feature vector range, the determination is not pornographic. The picture is subjected to image output step S30.

The invention is characterized in that the feature vector is used to calculate and find the feature points for comparison, and in the whole image analysis, the functions of locally analyzing the color patch area ratio, the color block distribution, the organ characteristics, and the like are simultaneously achieved, and The invention has improved the shortcomings of the global image superposition method in the conventional technology. In addition, the present invention utilizes the multiple confirmation mechanism and the error return function, and can automatically learn and update the computer mechanism by using the files in the cloud database to make the judgment more accurate and further. With the user database file of the cloud database, the customization effect can be achieved when the judgment standards of different countries and families are different.

The above is only a preferred embodiment for explaining the present invention, and is not intended to limit the present invention in any way, and any modifications or alterations to the present invention made in the spirit of the same invention. All should still be included in the scope of the intention of the present invention.

1. . . Image recognition system

10. . . Terminal identification module

11. . . Monitoring unit

13. . . Image analysis unit

15. . . Classification unit

17. . . Confirmation unit

19. . . Return unit

twenty one. . . First update unit

twenty three. . . Feedback unit

30. . . Cloud learning module

31. . . Basic database

33. . . Cloud database

35. . . Second update unit

37. . . Identification unit

39. . . Learning unit

S1. . . Image recognition system operation method

S10, S20, S21, S23, S25, S27, S29, S30, S40, S50, S60, S70, S80, S90, S100, S110, S120. . . step

The first picture shows the unit diagram of the image recognition system.

The second and third figures are flow charts of the method of operation of the image recognition system of the present invention.

The fourth figure is a detailed flowchart of the global image determining step of the present invention.

1. . . Image recognition system

10. . . Terminal identification module

11. . . Monitoring unit

13. . . Image analysis unit

15. . . Classification unit

17. . . Confirmation unit

19. . . Return unit

twenty one. . . First update unit

twenty three. . . Feedback unit

30. . . Cloud learning module

31. . . Basic database

33. . . Cloud database

35. . . Second update unit

37. . . Identification unit

39. . . Learning unit

Claims (9)

An image recognition system for filtering pornographic images, the system comprising: at least one terminal identification module installed on a user's terminal; and a cloud learning module installed on a cloud server The network is connected to each of the terminal identification module signals, wherein each of the terminal identification modules includes a monitoring unit, an image analysis unit, a classification unit, a confirmation unit, a reward unit, and a first update unit, and the monitoring unit Receiving a multimedia material from the outside, and extracting at least one image from the multimedia data, the image analyzing unit is connected to the monitoring unit, the classifying unit and the confirming unit, and receives the at least one captured by the monitoring unit And determining, according to the classification criterion of the classification unit, the at least one image is an erotic image, and if the image is determined to be an erotic image, the image analysis unit transmits the at least one image to the confirmation unit, and if When the determination is not an erotic image, the multimedia material is output; the confirmation unit determines the at least one image again Whether it is an erotic image, if it is an erotic image, the multimedia material is blocked, if it is not an erotic image, the at least one is transmitted to the reward unit, and the reward unit is connected to the confirmation unit and the first update unit for receiving the at least An image is transmitted to the first update unit, and the at least one is transmitted to the cloud learning module connected to the first update unit by using the first update unit, where the cloud learning module includes a basic database a cloud database, a second update unit, an identification unit, and a learning unit. The basic database stores an initial pattern data, the cloud database is used to create a supplementary pattern data, and the second updating unit and the second An update unit, the cloud database, and the identification unit are connected to receive the at least one image of the first update unit, and the identification unit is connected to the cloud database, the second update unit, and the learning unit for determining Whether the image analysis unit determines an error, and if it is determined that the image analysis unit determines an error, a determination error message is sent. Up to the learning unit, and transmitting the at least one image to the cloud database, and if it is determined that the image analysis unit does not determine an error, sending a determination error-free signal to the second update unit, the learning unit receiving the determination The error signal is compared with the initial pattern data of the at least one image in the cloud database, and an update file is generated, and the update is performed by the second update unit and the first update unit. The file is transmitted to the sorting unit to reset the classification judgment criterion. The system of claim 1, wherein each of the terminal identification modules further comprises a feedback unit, wherein the feedback unit is coupled to the reward unit for detecting the multimedia data output by the image analysis unit. When the multimedia material has an erotic image, the pornographic image is captured and transmitted to the reward unit. The system of claim 1, wherein the cloud database further establishes a user database to provide different classification criteria for different users. An image recognition system includes at least one terminal identification module installed on at least one user terminal, and a cloud learning module installed in a cloud server, the operation method comprising: an image capture a monitoring unit of the terminal identification module captures at least one image from the external multimedia data; a global image determining step, and an image analyzing unit in each terminal identification module is classified according to a classifier Determining a reference, analyzing the at least one image to determine whether there is an erotic image; and an image outputting step of outputting the multimedia material when the at least one image does not have an erotic image in the global image determining step; When the global image determining step determines that the at least one image has an erotic image, the user who installs the image recognition system performs double confirmation by using a confirmation unit of each terminal identification module to determine the at least one image. Whether there is pornographic image; a block image output step, when the manual judgment step Determining that the at least one has an erotic image, blocking the multimedia data output; and an error reporting step, when the manual determining step determines that the at least one does not have an erotic image, transmitting the at least one image to each of the terminal identification modules The first update unit is further transmitted to the cloud learning module by the first update unit; a recognition determining step, after the error reporting step, a second update unit of the cloud learning module will capture the The image is transmitted to an identification unit of the cloud learning module to determine whether the error is a return; and the step of updating the cloud database, if the identification determining step determines that the error is not an error, the identifying unit sends a determination error signal to the cloud learning module. a learning unit of the group, simultaneously transmitting the at least one image to the cloud database to establish the supplementary pattern data; and a step of calculating the difference value, the learning unit is based on the supplementary pattern data of the cloud database and a basic database The initial pattern data is compared, and a difference value is calculated, and the learning unit generates a difference according to the difference value. A new file, and then transmits the updated profile to the classification unit through the second and the first updating unit updating means; and a step of updating the classification unit, re-set stored in the classification unit is classified according to the updated reference profile is determined. The operation method of claim 4, wherein in the identification determining step, if the identification unit determines that the image analysis unit does not determine an error, performing a determination malicious return step, the identification unit sends a determination error-free signal The second update unit is given without performing any modified actions. The operating method of claim 4, wherein the cloud database further updates a user database to provide different classification criteria for different users. The method of operation of claim 4, further comprising an output monitoring step, wherein a feedback unit of each terminal identification module monitors the output of the multimedia material, and if the outputted multimedia material is found to have a pornographic pattern, Capture the pattern and proceed with the error reporting step. The method of claim 4, wherein the global image determining step comprises: a scale space limit detecting step, performing a Gaussian operation on each point of the image obtained by the system, and then performing a Gaussian operation, and then Performing at least two subtraction operations on the operation values of the two adjacent points of the processed image, and calculating a Gaussian value difference, and finding at least one maximum value and at least one minimum value of the Gauss value difference; a feature point positioning step And the maximum value and the minimum value obtained by the scale space limit detecting step are approximated by an expansion method to define a plurality of feature points; a direction assigning step, and the direction and the gradient are performed for the feature points Calculating, and calculating the direction and gradient of the surrounding points of the feature points; a feature point description step, determining the quasi-reference according to the classification in the classifier, and performing direction and gradient weighting on the relevant positions of each feature point to establish a plurality of feature vectors; and a feature vector determining step of confirming the feature points according to the feature vector values, if the feature vector values are in a feature In the vector range, the artificial image determination step is performed, and if the feature vector values are outside the feature vector range, it is determined that the image is not an erotic image, and the image output step is performed. The method of operation of claim 8, wherein the Gaussian operation is based on equation (1): L(x, y, σ) = G(x, y, σ) I(x,y) [Equation (1)], where L is the processed image, I is the initial image, σ is the standard deviation, and G(x, y, σ) is the Gaussian function, expressed by equation (2) : The Gauss value difference is expressed by equation (3): D(x, y, σ) = (G(x, y, σ ₁ ) - G(x, y, σ ₂ )) I(x, y) = L(x, y, kσ ₁ ) - L(x, y, σ ₂ ) [Equation (3)], where σ ₁ is the standard deviation of the first subtraction, and σ ₂ is The standard deviation of the second subtraction; the expansion is expressed by equation (4): Where D is the Gaussian value difference, and X = (x, y, σ) ^T , T is the shift term; the direction is calculated as in equation (5), and the gradient m(x, y) is calculated as the equation (6) ), [Equation (6)], where θ(x, y) represents a direction, and m(x, y) represents a gradient.