KR102629764B1 - Apparatus and Method for face Recognition based on Crawling and Deep Learning - Google Patents

Abstract

The present invention relates to a face recognition device and method based on crawling and deep learning, which stores data crawled from SNS and websites, recognizes facial features through a deep learning learning process of face photos, and creates keyframes from crawled images. The purpose is to separate and analyze the face detection image and extract the user's face image through matching and comparison with facial landmarks and provide it to the user.

Description

Face recognition device and method based on crawling and deep learning {Apparatus and Method for face Recognition based on Crawling and Deep Learning}

The present invention relates to a face recognition method based on crawling and deep learning. More specifically, it stores data crawled from SNS and websites, recognizes facial features through a deep learning learning process of face photos, and crawls video. This relates to a face recognition method based on crawling and deep learning, which is characterized by separating and analyzing face detection images from key frames and extracting user face images through matching and comparison with facial landmarks.

With the recent development of image processing technology, it is widely used to search for similar images by comparing feature points extracted from a query image with feature points of images stored in a database.

Conventionally, images are crawled on the Web and the characteristic points of each image are created into a database. In the case of web-based image search, images similar to the query image may be provided based on data previously stored on the server.

For example, a similar image may be an image that has the same semantic tag as the query image. For example, when a selfie image is used as a query image, a selfie image with a similar composition may be provided as a similar image. However, the database used for web-based image search may not contain information about images containing the same person as the person in the query image. Accordingly, the search results may not provide the user with an image of the same person as the person in the query image.

In this way, similar image search can be performed based on the tags of the query image. In this case, images in the database that have the same tag as the tag of the query image may be provided to the user as similar images.

In the case of web-based image search, a large number of images are stored in the database. Therefore, the number of images corresponding to each tag is relatively large. On the other hand, the number of images that can be stored on a user device is relatively small compared to a web server.

Therefore, when a similar image search is performed for images of a user device based on a tag, the number of similar images that can correspond to the tag of the query image is relatively limited.

Additionally, if the tag identified from the query image does not correspond to the context of the actual query image, images that are not similar to the actual query image may be provided as tag-based image search results.

However, for user face recognition, more accurate and specific face recognition technology based on various facial landmarks is required, and deep learning technology is used to continuously perform face recognition learning and improve user face matching performance. This is being demanded.

Domestic Patent Publication No. 10-2019-0021130 (publication date 2019.03.05.) Domestic Patent Publication No. 10-2019-0069920 (publication date 2019.06.20.)

Therefore, the present invention stores data crawled from SNS and websites, recognizes facial landmarks through a deep learning learning process of face photos, and separates and analyzes face detection images from crawled image keyframes to determine facial landmarks. The purpose is to provide a face recognition method based on crawling and deep learning that extracts the user's face image through matching and comparison and provides it to the user.

To achieve the purpose of the present invention, a face recognition device based on crawling and deep learning uses a crawler to crawl data from SNS (Social Network Service) or a website, and an image that matches the user's face image in the crawled data. In the crawling and deep learning-based face recognition device that extracts, the crawling data of the crawler is stored in a temporary storage DB, and the image information and facial feature point data for each key frame analyzed through the deep learning process are stored in the main server DB. The main server that does; a media downloader that downloads video or image data of the crawling data stored in the temporary storage DB and stores it in the storage server DB; A deep learning server that receives key frame image data containing a face image from the main server, extracts facial feature points for each key frame from the face data using an inference model through learning and re-learning, and provides the facial feature points to the main server; And a web server that compares the crawled facial feature point data stored in the feature point DB from the main server with the specific facial feature point data, extracts data greater than a set match value, and displays the results.

Here, the deep learning server extracts facial feature points from the face data input from the main server using an inference module and a feature point module and outputs them to the main server, and extracts feature point data from the feature point DB through the main server. Requesting re-learning server; It includes a learning server that extracts an inference model by learning the face data and feature point data collected from the main server, and replaces the new inference model extracted through the learning server with the existing inference model and sends it to the inference server. It is characterized by providing.

The deep learning server includes an input layer that receives R, G, and B images generated by extracting facial feature points; A learning layer that finds the location of facial feature points by learning with the content described in the layer file based on the prepared learning materials; and an output layer that outputs coordinate values of facial feature points in the learning layer.

The crawling and deep learning-based face recognition process to achieve the purpose of the present invention uses a crawler to crawl data from SNS (Social Network Service) or a website, stores it through the main server, and specifies the user's identity in the crawled data. In a face recognition method based on crawling and deep learning that extracts an image matching a face image, the crawler crawls data from a designated website or SNS, stores the crawled data, and stores a key containing a face image from the stored crawling data. First process of extracting frame (Keyframe) data; A second process of transmitting the facial image keyframe data extracted in the first process to a deep learning server to request facial feature points; A third process in which the deep learning server that receives the facial image keyframe data extracts facial feature points for each keyframe for the inference model through a learning and re-learning process; And a fourth process of storing the facial feature point data extracted in the third step, comparing and analyzing the specific facial feature point, and extracting a matching facial image.

The first process includes step 11 of crawling images or videos by the main server according to the queue of SNS or website registered by the crawler and storing them in a temporary storage DB; and step 12 of registering the image or video information stored in the temporary storage DB in a frame analysis queue in the media downloader and processing the queue by the frame analysis program. 13 steps of analyzing the key frame containing the original image by the frame analysis program, confirming the facial information in the key frame, extracting the key frame and face image, and storing them in the storage server DB. .

In addition, deep learning in the second process includes step 21 of extracting and cropping only the face image from the crawled image; Step 22 of extracting facial feature points from the extracted and cropped face image; 23 steps to correct and project the position of the face image; 24 steps to extract and encode baseline measurements from the calibrated face images; And 25 steps of learning a facial feature classifier and training to generate feature points through feature classification of facial images and neural network training.

Additionally, in addition to the third process, step 31 of the deep learning server requesting face image data and feature point data required for learning from the re-learning server; Step 32, where the main server requests feature point data from the re-learning server to the feature point DB; Step 33, wherein the main server randomly extracts facial feature points from the feature point DB and transmits them to the learning server of the deep learning server; Step 34 of replacing the newly learned inference model using facial feature points extracted from the learning server with the existing inference model; and step 35 in which the inference server extracts the facial feature point data using the newly replaced inference model in step 34 and provides the facial feature point data to the main server.

The crawling and deep learning-based face recognition device and method according to the present invention stores data crawled from SNS and websites, recognizes facial feature points through a deep learning learning process of face photos, and recognizes faces from the crawled image keyframes. By separating and analyzing the images, you can extract only the same face images of the user by matching and comparing them with facial features. In particular, you can learn the data of the face images of users who do not want their faces exposed, and upload the matching face images to the Internet. There is an effect of being able to extract it by crawling on it.

In addition, if the crawled video is large data between 10 and 100MB, it is difficult to process it. To solve this, apply a NoSQL database using the characteristics of 'Horizontal Scalability' and 'High Availability' and distribute it across multiple databases. It can be stored for quick access and handle large amounts of data. It is an environment that uses cloud servers and real servers simultaneously to handle large amounts of data, which is effective in processing high volumes of data. Temporary storage DB is used to implement a platform environment. , It can be processed effectively through split processing into the storage server DB and main server DB.

1 is a configuration diagram of a crawling and deep learning-based face recognition device according to an embodiment of the present invention;
Figure 2 is an overall flowchart of the crawling and deep learning-based face recognition process according to an embodiment of the present invention.

The features and effects of the present invention described above will become more apparent through the following detailed description in conjunction with the accompanying drawings, and accordingly, those skilled in the art will be able to easily implement the technical idea of the present invention. You will be able to.

Since the present invention can be subject to various changes and can have various forms, specific embodiments will be illustrated and described in detail in the text.

However, this is not intended to limit the present invention to a specific disclosed form, and should be understood to include all changes, equivalents, and substitutes included in the spirit and technical scope of the present invention.

The terms used in this application are only used to describe specific embodiments and are not intended to limit the invention.

The crawling and deep learning-based face recognition device and method according to a preferred embodiment of the present invention will be described in detail with reference to the attached drawings as follows.

Figure 1 is a configuration diagram of a crawling and deep learning-based face recognition device according to an embodiment of the present invention, which includes a crawler (②) that collects data (①) by crawling data from SNS (Social Network Service) or a website. 100), the crawling data (③) of the crawler 100 is stored in the temporary storage DB (120) (④), and the image information and facial feature point data for each key frame analyzed through the deep learning process are stored in the main server DB ( 151), the image/video data of the crawling data stored in the temporary storage DB 120 is downloaded (⑤) through the media downloader 121, and in the keyframe of the image/video data A storage server DB (122) that separates and stores only facial images (⑥), and facial image data stored in the storage server DB (122) is input through the main server (150) (⑦) through learning and re-learning. A deep learning server 130 that extracts facial feature points for each key frame from facial data using an inference model and provides them to the main server (⑧), and the deep learning server 130 input to the main server 150. A main server DB 151 that stores the inference result feature point data (⑨), and a web server that compares the crawled facial feature points stored in the main server DB 151 with the user's specific facial feature point data and extracts data exceeding a set match level. (160); It consists of a web client 170 that displays the facial data extracted from the web server 160.

Here, the deep learning server 130 extracts facial feature points from the face data input from the main server 150 using an inference module and a feature point module and outputs them to the main server 150, where the main server ( 150), a re-learning server 131 that requests feature point data from the feature point DB 152 (⑪, ⑫, ⑬), and a re-learning server 131 that requests feature point data from the feature point DB 152 through the main server 150. It consists of a learning server 132 that extracts an inference model by learning the feature point data (⑭⑮).

The specific operation of the present invention configured as described above will be described in detail with reference to the attached FIGS. 1 and 2 as follows.

First, the present invention collects images/videos from multiple websites or SNS using a crawler when a user wants to extract facial images from multiple websites or SNS containing the same face image as his or her own face image, and uses deep learning to collect images/videos from multiple websites or SNS. By comparing and analyzing the facial feature points learned and calculated through the process and the specific facial feature points extracted through the deep learning process of the user's face image, a facial image with a matching degree higher than the reference value is extracted.

Referring to Figure 2, the crawler is used by installing Selenium in the Chrome browser environment, and a selector (Selector: a selector such as Xpath/css selector that can define the title, image location, video location, etc.) template After defining it in advance, it is copied and used. In addition, the crawler connects to the site, checks the list to be crawled, registers in the crawler queue, performs crawling based on the contents registered in the crawler queue, uploads to the storage, and registers video or image data in the temporary storage DB (120). .

Specify a website or SNS for crawling. The crawler 100 crawls videos or images from a designated website (①), collects large amounts of data of 10 to 100 MB per video (②), and transmits the crawling result data to the main server 150 (③). ) is done.

In this way, large-capacity data processing allows data to be processed more quickly in a distributed environment such as a NoSQL (Not Only SQL) database that utilizes the characteristics of horizontal scalability and high availability.

In other words, in order to process the large amount of data, a platform environment is implemented to split the processing into temporary storage DB 120, storage server DB 122, and main server DB 151, and cloud server and real server. ) is designed to be used simultaneously.

For example, the temporary storage DB 120 temporarily stores the crawling data generated by the crawler 100 by the main server 150 (④), before uploading it to the storage server DB 122. It is used as an inspection and synchronization process, and images/videos are downloaded (⑤) using the media downloader (121) and stored in the storage server DB (122) (⑥).

A storage for storing the images analyzed by the crawler 100 and location information of the images are stored.

To describe the operation process of the crawler 100 in more detail, the crawler 100 first stores a video or image in the temporary storage DB 120 according to a queue of a registered Jiji website or SNS, and stores the video or image in the temporary storage DB 120. Enter video/image information stored in DB 120 and register it in the frame analysis queue of the frame analysis program.

The frame analysis program includes the original image as the basic frame of the frames between each key frame, and several key frames are included in the compressed video, and are less than the number of frames. The keyframe is extracted and registered in the queue of the frame analysis program.

In addition, the frame analysis program processes the cue and checks the facial information in the key frame, and at this time, uses a machine learning algorithm such as PCA (Principal Component Analysis) or LDA (Linear Discriminant Analysis) to check the facial information.

If there is facial information by the frame analysis program, key frames and image information are uploaded to the temporary storage DB 120, and when facial feature point information verified using an algorithm for processing the uploaded image information is confirmed, it is used as main information. Register in server DB (151).

In addition, in the case of crawler updates, availability is secured using Docker images, and crawler scalability is required, allowing real-time increase/decrease of crawlers using Docker. In the cloud server, Docker is used closest to the real server, and by using Docker, all crawlers are run in the same environment, and when the image is changed, all crawler systems are updated to use the same image.

The data crawled through this process is stored in the temporary storage DB 120. Since the temporary storage DB 120 has a large capacity of images, a separate storage cluster is created and stored to manage the stored information. Managed video information includes name, original URL, hash value of stored video, key frame location, key frame information, face detection information, analysis target site cue information, image/video data, time information, etc. This saved location information is registered.

The temporary storage DB (100), storage server DB (120), and main server DB (150) are closely connected to each other, enable tracking of the absolute path of the linked storage, and perform distributed data storage and hash data storage. It has a function.

The video data crawled through the above process sends a keyframe image containing a face to the deep learning server 130 at the request of the main server 150 to request facial feature points (⑦), and the deep learning server ( 130) performs deep learning learning, and the layer configuration of the deep learning server 130 is as follows.

First, the face detection process detects the face position using the "Tinyfaces" model, which is a face detection model. At this time, rotation, scaling, and rotation are performed to bring the eyes, nose, and mouth to the center as much as possible. 3D warps, including shear, are performed, and when performing this process, the eyes and mouth are aligned to approximately the same position in the image.

The face image detected in this way is a layer that receives the 3-channel image (R, G, B) generated by extracting the facial feature points from the input layer, and can be changed to a grayscale image.

In the learning layer, learning is conducted based on pre-prepared learning materials as a supervised learning process. Learning is conducted based on the SGD (Stochastic Gradient Descent) algorithm, and actual learning is performed using the content described in the layer file. 68 facial feature points It is a layer with 68 outputs that finds the location of points.

The output layer is a layer that outputs the x, y coordinates of the 68 facial feature points in the learning layer, and all face images are adjusted to the same size.

Meanwhile, the deep learning server 130 is a server that serves models developed with “Tensorflow” or “Pytorch” and is distributed according to required performance to be processed on several servers, and the learning server 120 provides 68 facial feature points. Design a model that extracts in the following order.

(1) Prepare training data and verification data labeling 68 facial feature points in a ratio of 8:2 (prevent overfitting by selecting data that is as unrelated as possible for each data).

(2) Proceed with learning using the designed model, repeat the process with verification data (proceed with verification -> modify model), and continuously work to prevent overfitting of the model and improve performance and recognition rate. Export the developed model in a serviceable form.

(3) Preparing a model for distribution and serving with a deep learning server

(4) Creation of Docker file with new model for deep learning server

(5) After removing the server from the deep learning server list, sequentially shut down the prepared deep learning servers and upload and restart a new Docker file.

(6) Check whether it has been applied as a new version of the model

(7) Registration and operation in the deep learning server list

(8) Inspection and re-learning process of data received through deep learning server

(9) Save and inspect data with a high compliance rate in a separate folder or server through a deep learning server to check whether the data can be used for re-learning.

(10) Improved detection rate by proportionally adding and relearning the training data or verification data of the learning server.

The operation of the deep learning server 130 designed in this way will be described in detail as follows.

The deep learning server 130 separates only the face from the image. For this purpose, face location extraction and cropping are performed using the "Tinyfaces" model, which is excellent for detecting even small faces.

Next, facial landmarks are extracted. 68 facial feature points are extracted, including the top of the chin, the outer edge of the eye, and the inner edge of the eyebrow, which exist in all faces. Specifically, it calculates the horizontal, vertical, and curvature of the eye gradient (point), calculates the interval between the right end point of the left eye and the left end point of the right eye, calculates the positions of the lower end points on both sides of the eyebrows, and calculates the positions of the upper end points on both sides of the eyebrows. Calculate the curvature of the eyebrows, calculate the position of both lower ends of the lips and the size of the philtrum and lower part, and calculate 17 feature points for other facial shapes.

In addition, through position correction and projection of the face image, rotation and scale -> shear steps are performed to bring the eyes and mouth to the center.

Next, the corrected face is encoded. That is, basic measurements are extracted from each photo, and the size of the ears, the distance between the eyes, the length of the nose, the length of the lips, etc. are encoded.

Next, classify facial features and train a neural network. In other words, it is classified using a facial feature classifier (SVM: Support-vector machine), and instead of training a neural network to recognize objects in each photo, it is trained (embedded) to infer 68 measurements for each face.

The learning server 140 continuously learns the model designed by the deep learning server 130, and continuously learns the face data inferred by the crawling and analysis processor, so the more it learns, the more it is detected. Performance improves.

The facial feature point data extracted from the deep learning server 130 is transmitted to the main server 150 (⑧), and the main server 150 stores the input facial feature point data in the main server DB 151, The facial feature point data stored in the main server DB (151) is encrypted to encrypt personal information and stored in the feature point DB (152), and the unique value of the feature point DB (152) and the unique value of the storage server DB (122) are combined. Save. (⑨)

The main server 150 matches the detection rate with the deep learning server 130, and the main server DB 151 provides video temporary storage DB 120 information, video date and time, and keyframe location information within the video. , Facial feature point information detected in the image is stored and used for face comparison.

In addition, the main server 150 essentially requires a process for quick accessibility and fault repair, and utilizes the automatic recovery process or backup recovery of the NoSQL server.

In addition, high performance of the DB is essential by processing large amounts of data. To achieve this, data is stored in memory or in a distributed database to prevent data storage and reading optimization problems.

In addition, the process of comparing facial data requires a lot of data operations. Real-time processing requires processing on SQL commands, and a controller is added to the database or the KNN (K-Nearest Neighbor) algorithm is implemented with SQL commands.

Through comparative analysis with the user's face data requested from the web server 160, a degree of match is calculated according to the degree of matching, and face images above a certain degree of match are extracted.

In other words, the web server 160 is a server that interoperates with the web client 170 to learn the user's face photo, etc. through the learning server 132, which learns the user's face photo, etc., and serves the learned deep learning. The model is sent to a deployed served model provider that transmits it to the inference server.

68 user facial feature point information is obtained from the served model provider, and the user facial feature points are compared and analyzed with the facial feature points stored in the main server DB 151 to calculate the degree of matching, and as a result, Facial images exceeding a certain level of match are extracted and transmitted to the web server 160.

The serving model supplier (Deployed Served Model) is replaced through continuous distribution in the learning server 132 and re-learning server 131, and the model is updated through the replaced model.

The re-learning server 131 requests collection of face data and specific point data required for learning (⑪,⑫), and the main server 150 randomly extracts facial feature points for learning from the feature point DB 152. Request them to come. (⑬)

The specific point DB 152 and the main server 151 transmit randomly extracted facial singular points to the deep learning server 130 through the main server 150 (⑭, ⑮), and the learning server 132 receives the collected facial singularities ( ), and use this to extract a new inference model. ( )

Replace the new inference model with the existing inference model from the learning server, ( ) The inference server extracts the facial feature point data using the newly replaced inference model and provides it to the main server 150. (⑧)

The inspection and redistribution process pre-processes the data and optimizes the model to ensure that there is no correlation as much as possible to solve the problem of overfitting. Due to the nature of deep learning, model performance problems depending on the amount of learning data are the output of the learning evaluation process of the deep learning server. The performance problem of a single server is solved through a process of re-learning, and when the instantaneous throughput is high through load balancing using Docker, a new server is registered to ensure balanced and fast throughput.

When uploading an image, the web server 150 and the web client 170 retrieve facial feature point information using the distributed model of the deep learning server 130, and use the KNN algorithm to retrieve facial feature point information from the main DB server. By comparing with the facial feature point data of 150, facial data with a certain degree of matching or higher is obtained and provided to the user through the screen of the web client 170.

As described above, the crawling and deep learning-based face recognition device and method according to the embodiment of the present invention, although described with limited embodiments and drawings, are not limited to this embodiment and are commonly known in the technical field to which the present invention pertains. Of course, various modifications and variations can be made by those skilled in the art within the technical spirit of the present invention and the equivalent scope of the claims described below.

100: Crawler 110: Website, SNS
120: Temporary storage DB 121: Media downloader
122: Storage server DB 130: Deep learning server
131: Re-learning server 132: Learning server
150: main server 151: main server DB
152: Feature point DB 160: WEB server
170: Web client

Claims (17)

In a crawling and deep learning-based face recognition device that crawls data from SNS (Social Network Service) or a website using a crawler and extracts an image matching the user's face image from the crawled data,
A main server that stores crawling data of the crawler in a temporary storage DB and stores image information and facial feature point data for each key frame analyzed through a deep learning process in a main server DB;
A media downloader that downloads video or image data of the crawling data stored in the temporary storage DB and stores it in the storage server DB;
A deep learning server that receives key frame image data containing a face image from the main server, extracts facial feature points for each key frame from the face data using an inference model through learning and re-learning, and provides the facial feature points to the main server; and
a feature point DB that encrypts and stores facial feature point data extracted from the deep learning server;
A web server that compares crawled facial feature point data stored in the feature point DB from the main server with specific facial feature point data, extracts data equal to or higher than a set match value, and displays the results,
The crawled data is distributed and stored in temporary storage DB, storage server DB, and main server DB. Each temporary storage DB, main server DB, and storage server DB are closely connected to each other, and the absolute path of each linked DB is tracked. Contains hash data to make it possible,
The deep learning server extracts facial feature points from the facial data input from the main server using an inference module and a feature point module and outputs them to the main server,
a re-learning server that requests feature point data from the feature point DB through the main server;
a learning server that extracts an inference model by learning face data and feature point data collected from the main server; and
An inference server that extracts feature point data of the face using an inference model newly extracted through the learning server and then provides the feature point data to the main server,
The deep learning server includes an input layer that receives R, G, and B images generated by extracting facial feature points;
A learning layer that finds the location of facial feature points by learning with the content described in the layer file based on the prepared learning materials; and
A face recognition device based on crawling and deep learning, comprising: an output layer that outputs coordinate values of facial feature points in the learning layer. According to paragraph 1,
The main server transmits the face data to the deep learning server, and stores facial feature point data extracted from the deep learning server and unique values of the storage server DB in the main server DB. Learning-based face recognition device. According to paragraph 1,
The crawling data stored in the temporary storage DB includes video hash values, key frame positions, key frame information, face detection information, analysis target site queue information, URL, image and video data. A face recognition device based on crawling and deep learning. delete According to paragraph 1,
The main server transmits feature point data randomly extracted from the feature point DB to the deep learning server in response to a feature point data request from the deep learning server, and the deep learning server sends the feature point data received from the main server to the learning server. A face recognition device based on crawling and deep learning that extracts an inference model and replaces the existing inference model with a new inference model. delete delete delete According to paragraph 1,
The learning server performs learning on the model generated by the deep learning server and continuously retrains the data generated by the crawling and analysis processor to increase the detection rate. face recognition device. According to paragraph 1,
The specific facial feature point data is transmitted to the deep learning server when a user photo is uploaded from the web server to the main server, and facial recognition based on crawling and deep learning is characterized in that the user's specific facial feature point is extracted through a learning process. Device. A face recognition method based on crawling and deep learning that uses a crawler to crawl data from SNS (Social Network Service) or websites, stores it through the main server, and extracts images that match the user's specific face image from the crawled data. In
A first process in which the crawler crawls data from a designated website or SNS, stores the crawled data in a temporary storage DB, and extracts keyframe data including a face image from the stored crawling data;
A second process of transmitting the facial image keyframe data extracted in the first process to a deep learning server to request facial feature points;
A third process in which the deep learning server that receives the facial image keyframe data extracts facial feature points for each keyframe for the inference model through a learning and re-learning process; and
A fourth process of storing the facial feature point data extracted in the third step in the main server DB, comparing and analyzing the specific facial feature point and extracting a matching facial image,
The first process includes step 11 of crawling images or videos by the main server according to the queue of SNS or website registered by the crawler and storing them in a temporary storage DB; and
12 steps of registering the image or video information stored in the temporary storage DB in a frame analysis queue in the media downloader and processing the queue by a frame analysis program;
The frame analysis program analyzes key frames containing the original image, checks facial information in the key frames, extracts the key frames and face images, and stores them in the storage server DB.
In the third process, step 31, the deep learning server requests facial image data and feature point data required for learning from the re-learning server;
Step 32, wherein the main server requests feature point data from the re-learning server to a feature point DB where facial feature point data is encrypted and stored;
Step 33, wherein the main server randomly extracts facial feature points from the feature point DB and transmits them to the learning server of the deep learning server;
Step 34 of replacing the newly learned inference model using the facial feature points extracted from the learning server with the existing inference model and transmitting it to the inference server; and
A face recognition method based on crawling and deep learning, characterized in that the inference server extracts the facial feature point data using the newly replaced inference model in step 34 and provides the facial feature point data to the main server. delete According to clause 11,
In the second process, deep learning includes 21 steps of extracting and cropping only face images from crawled images;
Step 22 of extracting facial feature points from the extracted and cropped face image;
23 steps to correct and project the position of the face image;
24 steps to extract and encode baseline measurements from the calibrated face images; and
A face recognition method based on crawling and deep learning, comprising 25 steps of learning a facial feature classifier and training to generate feature points through feature classification of facial images and neural network training. According to clause 13,
In step 22, the extraction of facial features includes horizontal, vertical, and curvature values of eye tilt;
Distance between the right endpoint of the left eye and the left endpoint of the right eye;
Curvature value of eyebrows;
The position of both lower ends of the lips and the size of the philtrum and lower end; and
A face recognition method based on crawling and deep learning, characterized by including 17 points about the shape of the face. delete According to clause 11,
The specific facial feature point is a face recognition method based on crawling and deep learning, characterized in that the main server transmits the user's facial image requested by the user through a web server to the deep learning server to calculate specific facial feature point data. . According to clause 16,
A face recognition method based on crawling and deep learning, wherein the web server compares and analyzes specific facial feature point data extracted from the deep learning server and facial feature point data extracted in the third process to calculate the degree of matching.