TW201606706A - Method for filtering unsightly image from a batch of images - Google Patents

Abstract

A method for filtering unsightly image from a batch of images includes the following steps. Receive a batch of images uploaded according to a template by a terminal device. Execute an eye detection of each image, to determine whether the image has eye feature or not. When the image doesn't have the eye feature, execute a face detection of the image without the eye feature, to determine whether the image has front-face feature or not. When the image does have the eye feature or does have the front-face feature, execute a color-level comparison of the image with the eye feature or the front-face feature according to a complexion range, to determine whether the image has skin pattern or not. When the image has the skin pattern, execute a mouth detection of the image with the skin pattern, to determine whether the image has mouth feature or not. Decide each image with the mouth feature as unsightly image and set all unsightly images to be disabled items, such that all unsightly images are not slot in a network platform. Decide each image without the front-face feature, the skin pattern or the mouth feature as normal image and slot all normal images in a network platform according to the template.

Description

Indecent picture filtering method

The invention relates to a picture publishing technology of a network platform, in particular to a filtering method for an indecent picture.

With the rapid development of Internet communication and information technology, the rise of the Internet has brought about a fast and convenient communication environment and the flow of information, so that more and more people are relying on the Internet for entertainment, research and business activities.

In the case of e-commerce, the seller can display the sales information on the network platform through the web server, and the buyer browses and selects the goods to be purchased on the same network platform through the terminal device and directly Complete the transaction online. In e-commerce, the seller can provide various information such as sellers and goods to the buyer without having to store the storefront, thus saving a lot of store rent, and the buyer can purchase the goods and go home without going out, so e-commerce has become a business necessity. Business model.

In order to sell the goods to be sold, the seller usually introduces the characteristics and uses of the products in a large number of pictures, texts or videos to attract consumers to purchase. In the e-commerce website platform, if the product image appears indecent, it will affect the goodwill of the company and the seller who provide the website platform.

Although the traditional image recognition technology can directly filter suspicious image data, its identification time and/or its recognition effect cannot be satisfactory. Another method of image filtering is to use a database comparison method. However, this method can only be used to compare the known pictures stored in the database. As for the indecent pictures that are not stored in the database, they can only be judged as normal pictures. Therefore, a more efficient image filtering method is needed.

In an embodiment, the filtering method of the indecent picture includes: receiving a plurality of pictures that are batch-uploaded by a terminal device based on a template, performing eye detection of each picture to determine whether each picture has an eye feature, and when the picture does not have an eye feature Performing face detection without a picture of the eye feature to determine whether the picture has a positive face feature, and when the picture has an eye feature or a positive face feature, having an eye feature or a positive face feature according to a skin color range The color scale of the picture is compared to determine whether the picture has a skin tile. When the picture has a skin tile, the mouth detection of the picture with the skin tile is performed to determine whether the picture has a mouth feature and will not have a positive face feature. A picture that does not have a skin color block or has no mouth feature is determined to be a normal picture and all normal pictures are based on a template to a network platform, and a picture having a mouth feature is determined as an indecent picture and all indecent pictures are set For the ban project, it is not on the Internet platform.

In summary, the filtering method of the indecent picture according to the present invention is applied to a network server, which can provide a network platform with relatively fast and accurate filtering and detecting indecent pictures.

Referring to Figures 1 to 3, the filtering method of the indecent picture is applied to a network server 10. Here, the network server 10 provides a network platform 20. The user can connect to the web server 10 via the network 40 using a terminal device 30 to upload and publish a plurality of pictures on the network platform 20. The network server 10 is provided with at least a processing unit 110, a storage unit 130, a network module 150, a first counter 170 and a second counter 172, and a storage unit 130, a network module 150, and a first counter 170. The second counter 172 is electrically connected to the processing unit 110. Therefore, the filtering method of the indecent picture according to the present invention can be implemented on the network server 10 by the processing unit 110 executing the firmware or software algorithm stored in the storage unit 130. Here, the network server 10 can connect to the network 40 via the network module 150 and communicate with the terminal device 30 via the network 40. Here, the web platform 20 can be a shopping website, a blog, an electronic photo album, or a social networking website.

The processing unit 110 receives the plurality of pictures that the terminal device 30 batch uploads based on the template of the network platform 20 (step S10), and stores the received pictures in the storage unit 130. Then, the processing unit 110 sequentially performs filtering detection of each received picture to determine that the uploaded picture is a normal picture or an indecent picture.

In the filtering detection process of each picture, the processing unit 110 first performs eye detection of the picture (step S20) to determine whether the picture has an eye feature (step S22). When the picture does not have an eye feature, the processing unit 110 performs face detection without a picture of the eye feature (step S30) to determine whether the picture has a positive face feature (step S32). When the picture has an eye feature or has a positive face feature, the processing unit 110 performs a tone scale comparison with an eye feature or a picture having a positive face feature according to a skin color range (step S40) to determine whether the picture has a skin tile ( Step S42). When the picture has a skin tile, the processing unit 110 performs mouth detection with a picture of the skin tile (step S50) to determine whether the picture has a mouth feature (step S52). Here, the processing unit 110 determines a picture that does not have a positive face feature, does not have a skin color block or does not have a mouth feature as a normal picture (step S60), and determines a picture having a mouth feature as an indecent picture (step S70) .

Then, the processing unit 110 successively performs the filtering detection of the next picture (ie, returns to step S20) until the filtering detection of all the pictures is completed. After the filtering detection of all the pictures is completed, the processing unit 110 uploads all the normal pictures based on the template selected at the time of uploading to the network platform, and sets all the indecent pictures as the disabled items so as not to be put on the network platform (step S80). . Here, an indecent picture refers to an erotic picture or a picture with a bare body.

The steps S20, S30, and S50 can be achieved by using a knowledge identification technology or a statistical identification technology, respectively. Among them, the knowledge identification method mainly uses the prior knowledge to regard the face as a combination of organ characteristics, and then detects according to the characteristics of the eyes, eyebrows, mouth, nose and other organs and the geometric positional relationship between them. The statistical identification method is a model in which a face is regarded as a whole (for example, a two-dimensional primitive matrix), and then a face image space is constructed from a statistical viewpoint through a large number of face image samples, and then detected according to the similarity. In the field of machine learning, many algorithms for identification techniques are processes that classify or cluster things. The way in which the total number of known object categories is generally identified is referred to as classification, and the training material is tagged (eg, explicitly specifying whether the material has facial features or no facial features). Clustering is a method in which there is uncertainty in the total number of categories that can be processed or training is unlabeled, and it does not require information about the object class in the learning phase.

In some embodiments, in steps S20, S30, S50, processing unit 110 uses a Haar-1ike feature to detect features (eye features, face features, or mouth features) to be determined. Here, the processing unit 110 accelerates the Haar feature evaluation using an integral image (Integral Image), and uses the AdaBoost (Adaptive Enhancement) algorithm to train a strong classifier that distinguishes between faces and non-human faces and uses a filter cascade. Concatenate strong classifiers together.

Referring to FIG. 4, the detecting step using the Haar-1ike feature includes converting the color gradation of the picture to a gray level (ie, converting to a grayscale picture) (step S91), finding this in a predetermined window. a complex rectangular feature in the picture (step S93), calculating an eigenvalue of the rectangular features (ie, Haar features) using an integral map (step S95), and determining whether to have a temptation based on the calculated eigenvalues The measured feature (eye feature, face feature or mouth feature) (step S97).

Here, the rectangular feature includes four boundary features, eight line features, and two center features, and each feature is composed of 2 to 3 rectangles. Such rectangular features can be expressed as Equation 1.

Formula 1

Where ω _i is the weight of the rectangle, RectSum( r _i ) is the gray level integral of the image in the rectangular range, and N is the number of rectangles constituting the rectangular feature.

The calculation formula of the rectangular feature is as shown in Equation 2.

Equation 2

Where W is the width of the picture, H is the length of the picture, w is the width of the rectangular feature, and h is the length of the rectangular feature. Therefore, W×H is the size of the picture, w×h is the size of the rectangular feature (the left side of FIG. 5), and the maximum scale factor of the rectangular feature in the horizontal and vertical directions is as shown in Equation 3.

Equation 3

For the rectangular feature of 45° rotation, w and h are represented as shown on the right side of FIG. 5, and the calculation formula is as shown in Equation 4.

Equation 4

At this time, the maximum scale factor of the square characteristic of the 45° rotation in the horizontal and vertical directions is as shown in Equation 5.

Equation 5

In an embodiment of the gradation comparison, referring to FIG. 6, the processing unit 110 sequentially compares the plurality of pixels of the picture with the skin color range (step S410) to determine whether the currently aligned pixels fall into the picture. The skin color range (step S412). When the processing unit 110 determines that the currently aligned pixel falls within the skin color range, the first count value in the first counter 170 and the second count value in the second counter 172 are both accumulated by one (step S420). Here, the first count value indicates the cumulative number of pixels that have completed the alignment, and the second count value indicates the cumulative number of pixels that fall within the skin color range. When the processing unit 110 determines that the currently aligned pixels do not fall within the skin color range, the first count value in the first counter 170 is cumulatively incremented by 1, but the second count value in the second counter 172 is not accumulated (step S430). . Then, the processing unit 110 determines whether the ratio of the first count value to the total pixel count of the picture is less than a first ratio (step S440), and determines whether the ratio of the second count value to the first count value is less than a second Ratio (step S450).

When the ratio of the first count value to the total pixel count of the picture is less than the first ratio or the ratio of the second count value to the first count value is not less than the second ratio, the processing unit 110 continues to perform the ratio of the next pixel. Yes (ie, return to step S412) until the alignment of all pixels in this picture is completed. When the ratio of the first count value to the total pixel count of the picture is not less than the first ratio and the ratio of the second count value to the first count value is less than the second ratio, the processing unit 110 can directly determine that the picture does not have the skin. The tile (step S470) without continuing the comparison of the next pixel.

After completing the comparison of all the pixels, the processing unit 110 further determines whether the proportion of the second count value to the total number of pixels of the picture falls within the ratio range (step S460). When the ratio of the second count value to the total number of pixels of the picture does not fall within the ratio range, the processing unit 110 determines that the picture does not have a skin tile (step S470). On the other hand, when the ratio of the second count value to the total number of pixels of the picture falls within the ratio range, the processing unit 110 determines that the picture has a skin tile (step S480).

Here, the skin color range can be determined according to the region where the network platform provides services. For example, there are 36 skin colors #1~#36 in the color scale, as shown in Figure 7. The world skin color distribution can be divided into 8 skin color ranges G1~G8. Among them, the skin color range G1 is the skin color #1~#12, the skin color range G21 is the skin color #12~#14, the skin color range G3 is the skin color #15~#17, the skin color range G4 is the skin color #18~#20, and the skin color range G5 is Skin color #21~#23, skin color range G6 is skin color #24~#26, skin color range G7 is skin color #27~#29, and skin color range G8 is skin color #30~#36. The provider of the network platform 20 can select one of the skin color ranges G1 to G8 according to the service area as a program parameter for use in the color scale comparison. In other words, the skin color range used for the color scale comparison is one of the skin color ranges G1 to G8, and is stored in the storage unit 130 in advance.

Wherein, the first ratio may be 60%, that is, the cumulative number of pixels that have completed the comparison is 60% of the total number of pixels of the picture. The second ratio may be 50%, that is, the cumulative amount of pixels falling within the skin color range is 50% of the cumulative number of pixels that have completed the comparison.

Wherein the ratio range is composed of the first threshold value and the second threshold value, and the first threshold value is smaller than the second threshold value. Here, the first threshold may be greater than or equal to 20%, that is, the cumulative amount of pixels falling within the skin color range is greater than or equal to 20% of the total number of pixels in the picture. Moreover, the second threshold is less than 100%, that is, the cumulative amount of pixels falling within the skin color range is less than 100% of the total number of pixels in the picture. Preferably, the first threshold is 20% to 30% and the second threshold is 90%.

In the process of gradation alignment, the processing unit 110 determines that the picture is first converted from the first color space to the second color space, and then determines whether the picture has a skin tile by using a picture of the second color space (ie, contiguous Step S410) is performed. Here, the first color space may be RGB, and the second color space may be YUV (Luminance/Chrominance/Chroma; brightness/chroma/density), HSV (Hue/Saturation/Value; hue/saturation/lightness) or YIQ (Luminance/In phase/Quadrature phase; brightness/color).

Here, the storage unit 130 can be implemented by one or more storage elements. The storage element may be, for example, a memory or a scratchpad, but is not limited thereto. The memory can be, for example, a read only memory, a random access memory, a non-permanent memory, a permanent memory, a static memory, a dynamic memory, a flash memory, and/or any device that stores digital information. Processing unit 110 may be implemented by one or more processing elements. Processing elements can be microprocessors, microcontrollers, digital signal processors, microprocessors, central processing units, field programmable gate arrays, programmable logic devices, state devices, logic circuits, analog circuits, digital circuits, and/or any A device that operates on signals (analog and/or digits) based on operational instructions.

In summary, the filtering method of the indecent picture according to the present invention is applied to a network server, which can provide a network platform with relatively fast and accurate filtering and detecting indecent pictures. For example, taking the first ratio as 60%, the second ratio as 50%, and the ratio ranging from 25% to 90% as an example, the filtering method of the indecent picture of one embodiment of the present invention is used to complete 25G and the indecent picture is occupied. Filter detection of 25% of the images takes 11.45 hours and the accuracy is 87%. Manual testing takes 48 hours.

10‧‧‧Web server
110‧‧‧Processing unit
130‧‧‧storage unit
150‧‧‧Network Module
170‧‧‧First counter
172‧‧‧second counter
20‧‧‧Internet platform
30‧‧‧ Terminal equipment
40‧‧‧Network
S10‧‧‧Receive terminal device batch uploading multiple pictures based on template
S20‧‧‧A picture of eye detection
Does S22‧‧‧ have eye characteristics?
S30‧‧‧ Face detection for this picture
Does S32‧‧‧ have a positive face feature?
S40‧‧‧The color scale comparison of this picture
Does S42‧‧‧ have skin patches?
S50‧‧‧The mouth detection of this picture
Does S52‧‧‧ have mouth characteristics?
S60‧‧‧Determine this picture as a normal picture
S70‧‧‧Determining this picture as indecent
S80‧‧‧Put all normal pictures based on the selected template to the network platform, and set all indecent pictures as disabled items so that they are not on the network platform
S91‧‧‧ Convert pictures to grayscale pictures
S93‧‧‧ Find the complex rectangular features in this picture in a given window
S95‧‧‧Using integral graphs to calculate the eigenvalues of these rectangular features
S97‧‧‧Determine whether there are features to be detected based on the calculated eigenvalues
S410‧‧‧Compare the multiple pixels of this picture in order of skin color
Does S412‧‧‧ fall into the skin color range?
S420‧‧‧ Add 1 to the first count value and the second count value
S430‧‧‧To add 1 to the first count value, but not to accumulate the second count value
Is S440‧‧‧ less than the first ratio?
Is S450‧‧‧ less than the second ratio?
Does S460‧‧‧ fall within the ratio?
S470‧‧‧Determined that there is no skin patch
S480‧‧‧Determined with a skin patch
W‧‧‧ Pictures wide
H‧‧‧The length of the picture
w ‧‧‧The width of the rectangular feature
h ‧‧‧The length of the rectangular feature
#1~#36‧‧‧ Skin color
G1‧‧‧ skin color range
G2‧‧‧ skin color range
G3‧‧‧ skin color range
G4‧‧‧ skin color range
G5‧‧‧ skin color range
G6‧‧‧ skin color range
G7‧‧‧ skin color range
G8‧‧‧ skin color range

[FIG. 1] is a flowchart of a filtering method of an indecent picture according to an embodiment of the present invention. [Fig. 2] is a system diagram showing a filtering method of an indecent picture to which an embodiment of the present invention is applied. [Fig. 3] is a schematic block diagram of the network server in Fig. 2. [Fig. 4] is a detailed flowchart of a detecting step using a Haar-1ike feature according to an embodiment of the present invention. [Fig. 5] is a schematic view showing a rectangular feature according to an embodiment of the present invention. [Fig. 6] is a detailed flowchart of steps S40 and S42 of an embodiment of Fig. 1. [Fig. 7] A schematic view of a skin color table according to an embodiment of the present invention. [Fig. 8] Fig. 8 is a schematic view showing the distribution of world skin color according to an embodiment of the present invention.

S10‧‧‧Receive terminal device batch uploading multiple pictures based on template

S20‧‧‧A picture of eye detection

Does S22‧‧‧ have eye characteristics?

S30‧‧‧ Face detection for this picture

Does S32‧‧‧ have a positive face feature?

S40‧‧‧The color scale comparison of this picture

Does S42‧‧‧ have skin patches?

S50‧‧‧The mouth detection of this picture

Does S52‧‧‧ have mouth characteristics?

S60‧‧‧Determine this picture as a normal picture

S70‧‧‧Determining this picture as indecent

S80‧‧‧Put all normal pictures based on the selected template to the network platform, and set all indecent pictures as disabled items so that they are not on the network platform

Claims (9)

A method for filtering an indecent picture, comprising: receiving a plurality of pictures that are batch-uploaded by a terminal device based on a template; performing eye detection of each picture to determine whether each picture has an eye feature; when the picture does not have the eye a feature, performing face detection of the picture without the eye feature to determine whether the picture has a positive face feature; when the picture has the eye feature or having the positive face feature, having a skin color range Having the eye feature or the gradation of the picture having the positive face feature to determine whether the picture has a skin tile; when the picture has the skin tile, performing a mouth of the picture having the skin tile Detecting to determine whether the picture has a mouth feature; determining the picture that does not have the face feature, does not have the skin color block or does not have the mouth feature as a normal picture, and bases all the normal picture on the template Laid into a network platform; and determine the picture with the characteristics of the mouth as an indecent picture, and set all the indecent pictures as banned The project can not be put on the network platform. The filtering method of the indecent picture according to claim 1, wherein the color gradation matching step comprises: sequentially comparing the plurality of pixels of the picture with the skin color range to determine whether the pixel of the current comparison falls. Entering the skin color range; when it is determined that the pixel of the current comparison falls within the skin color range, a first count value and a second count value are cumulatively increased by 1, wherein the first count value indicates that the comparison has been completed. a cumulative number of the pixels, and the second count value represents a cumulative number of the pixels falling within the skin color range; when it is determined that the current alignment does not fall within the skin color range, the first count value Adding 1 to the total, but not accumulating the second count value; determining whether the ratio of the first count value to the total number of pixels of the picture is less than a first ratio; determining the second count value at the first count value Whether the ratio is less than a second ratio; when not less than the first ratio and less than the second ratio, determining that the picture does not have the skin tile; when less than the first ratio or not less than the second ratio, determining The second count value accounts for the total amount of the pixels Whether the ratio falls within a range of ratios; when falling within the ratio range, it is determined that the picture has the skin tile; and when it does not fall within the ratio range, it is determined that the picture does not have the skin tile. A filtering method of an indecent picture as claimed in claim 2, wherein the first ratio is 60%, and the second ratio is 50%. The filtering method of the indecent picture according to claim 2 or 3, wherein the ratio range is composed of a first threshold value and a second threshold value, the first threshold value is smaller than the second threshold value, the first The threshold is greater than or equal to 20% and the second threshold is less than 100%. The filtering method of the indecent picture according to claim 4, wherein the first threshold is 20% to 30%. The filtering method of the indecent picture according to claim 4, wherein the second threshold is 90%. The filtering method of the indecent picture according to claim 1, wherein the face detecting step comprises: converting the picture into a grayscale picture; and detecting the gray level picture by using a Haar-1ike feature The positive face feature. The filtering method of the indecent picture according to claim 1, wherein the eye detecting step comprises: converting the picture into a grayscale picture; and detecting the eye feature in the grayscale picture by a Haar feature. The filtering method of the indecent picture according to claim 1, wherein the mouth detecting step comprises: converting the picture into a grayscale picture; and detecting the mouth feature in the grayscale picture by a Haar feature.