WO2016106966A1 - 人物标注方法和终端、存储介质 - Google Patents

人物标注方法和终端、存储介质 Download PDF

Info

Publication number
WO2016106966A1
WO2016106966A1 PCT/CN2015/073337 CN2015073337W WO2016106966A1 WO 2016106966 A1 WO2016106966 A1 WO 2016106966A1 CN 2015073337 W CN2015073337 W CN 2015073337W WO 2016106966 A1 WO2016106966 A1 WO 2016106966A1
Authority
WO
WIPO (PCT)
Prior art keywords
character
person
marked
picture
characters
Prior art date
Application number
PCT/CN2015/073337
Other languages
English (en)
French (fr)
Inventor
周江
吴钊
张本好
邓伟洪
Original Assignee
中兴通讯股份有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 中兴通讯股份有限公司 filed Critical 中兴通讯股份有限公司
Publication of WO2016106966A1 publication Critical patent/WO2016106966A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16—Human faces, e.g. facial parts, sketches or expressions
    • G06V40/168—Feature extraction; Face representation

Definitions

  • the present invention relates to the field of communications, and in particular, to a character labeling method, a terminal, and a storage medium.
  • the embodiment of the present invention provides a character labeling method, a terminal, and a storage medium, and aims to solve the problem that the character labeling accuracy is low and the labeling efficiency is low.
  • an embodiment of the present invention provides a character labeling method, where the character labeling method includes:
  • the corresponding character is identified by using the character feature similarity to cluster the corresponding characters in the unlabeled person picture.
  • the step of combining the skin color filtering and the face detection to obtain the character features of the marked person in the labeled person image comprises:
  • the step of combining the skin color filtering and the face detection to obtain the character features of the marked person in the labeled person picture comprises:
  • the character features of the manually labeled characters are obtained.
  • the following steps are:
  • the following steps are:
  • the character images of the clustering are displayed by means of source image display, embedded display and/or thumbnail display.
  • the embodiment of the present invention further provides a terminal, where the terminal includes:
  • Obtaining a module configured to combine skin color filtering and face detection to obtain a character feature of the marked person in the marked person picture
  • the labeling module is configured to identify the corresponding person by using the character feature similarity according to the acquired character characteristics of the marked person, so as to cluster the corresponding characters in the unlabeled person picture.
  • the terminal further includes:
  • Forming a module configured to retrieve a person's picture by retrieving a local file; collecting all the characters' pictures to form a set of character pictures.
  • the acquiring module includes:
  • the identification unit is configured to identify the manual annotation in the character picture by combining the skin color filtering and the face detection Character
  • the obtaining unit is configured to obtain a character feature of the manually labeled person according to the identified manually labeled person.
  • the terminal further includes:
  • the sorting module is configured to perform iterative labeling and/or recommendation sorting on the corresponding characters in the character image that is still not labeled after the cluster labeling.
  • the terminal further includes:
  • the display module is configured to display the picture of the person marked by the cluster by means of a source image display, an embedded display, and/or a thumbnail display.
  • an embodiment of the present invention provides a computer storage medium, where the computer storage medium stores computer executable instructions, and the computer executable instructions are used to execute the character labeling method provided by the first aspect of the present invention.
  • the character labeling method, the terminal, and the storage medium provided by the embodiments of the present invention combine the skin color filtering and the face detection to obtain the character features of the marked characters in the labeled person image; and utilize the character characteristics according to the obtained character characteristics of the marked characters
  • the similarity identifies the corresponding person to cluster the corresponding characters in the unlabeled person picture.
  • the invention has high accuracy and efficiency.
  • FIG. 1 is a schematic flow chart of a first embodiment of a method for marking a person according to the present invention
  • FIG. 2 is a schematic flow chart of a second embodiment of a method for marking a person according to the present invention
  • 3 is a detailed flow chart of the steps of combining the skin color filtering and the face detection to obtain the character features of the marked person in the labeled person picture;
  • FIG. 4 is a schematic flow chart of a third embodiment of a method for marking a person according to the present invention.
  • FIG. 5 is a schematic flow chart of a fourth embodiment of a method for marking a person according to the present invention.
  • FIG. 6 is a schematic diagram of functional modules of a first embodiment of a terminal according to the present invention.
  • FIG. 7 is a schematic diagram of functional modules of a second embodiment of a terminal according to the present invention.
  • FIG. 8 is a schematic diagram of functional modules of the acquisition module of FIG. 6;
  • FIG. 9 is a schematic diagram of functional modules of a third embodiment of a terminal according to the present invention.
  • FIG. 10 is a schematic diagram of functional modules of a fourth embodiment of a terminal according to the present invention.
  • An embodiment of the present invention provides a method for labeling a person.
  • the method for labeling a person includes:
  • Step S100 combining skin color filtering and face detection to combine skin color filtering and face detection to obtain character features of the marked person in the labeled person picture.
  • the characters marked in the picture of the person may be manually labeled, or may be automatically marked by the terminal, or may be automatically labeled by the result of several manual markings.
  • the manual annotation can be multi-graphed at the same time, or it can be manually modified or removed manually;
  • the automatic annotation can also be a single or multiple icon annotation for a single character or multiple characters at the same time.
  • the face information is used to detect and identify the character features of the person's picture, and combined with skin color filtering and face detection techniques. Face detection is based on the framework of the V-J face detector. The traditional VJ face detector extracts the Haar features by scanning a large number of detection frames, and then performs the Adaboost algorithm to quickly filter out the non-human face frames.
  • each face detection box is firstly tested for skin color, and the skin color detection result can quickly and effectively assist in judging whether it is a human face, thereby filtering out most of the area as a previous level of filtering.
  • the skin color likelihood of each pixel is obtained by a large number of training pictures, and the average skin color likelihood of the detection frame is calculated.
  • the retained picture is sent to the lower classifier. Otherwise filter Drop it. It then traverses all the pictures in the current local path, saves the face data, and creates a data unit for each face.
  • This data unit contains person information, serial number information, feature vector information, source information, coordinates, and posture information of the face.
  • the source information is the source map path of the face is the read value
  • the serial number information is the design value
  • the coordinate information is the face detection.
  • the result value, feature vector and pose information are the calculated result values.
  • the dimension of the final feature vector is 200 dimensions
  • 100-dimensional is the LBP feature after dimension reduction
  • LBP (Local Binary Pattern) 100-dimensional is the HG (Histogram of Oriented Gradient). feature.
  • the HOG feature first uses the SDM algorithm to locate the feature points, and locates the left and right eyes, the left and right mouth corners and the nose tip.
  • the coordinates of the left and right corners can be used to calculate the position of the center of the mouth, using the position of the center of the mouth and the left and right eyes.
  • Point as the standard of alignment, map the three points to a fixed position on the 100*100 size image by affine change, and extract the LBP feature and the HOG feature from the face image of the 100*100 size, and then use the advance
  • the trained PCA and LDA dimension reduction matrices respectively reduce the dimension values of LBP and HOG features by 100 dimensions, respectively, and perform model normalization and tandem into 200-dimensional feature vectors.
  • the corresponding rotation mapping matrix can be inversely obtained, and a standard frontal face frame is used by the rotation mapping matrix. After the transformation, without rotation, the same three-dimensional face frame can be obtained, and then the face frame with obvious three-dimensional visual effect can be obtained through perspective projection.
  • Step S200 Identify the corresponding person by using the character feature similarity according to the obtained character feature of the marked person, and perform clustering labeling on the corresponding person in the unlabeled person picture.
  • the corresponding character is identified by using the character feature similarity, and the corresponding character in the unlabeled person image is clustered according to the identified corresponding character, wherein the clustering label uses the K-means clustering method, and the preset A k value, clustering all characters, the clustering results are divided into k categories, and the cluster center of each class is displayed and prompted to read, and the preset character information can be read.
  • the terminal maintains a file that saves the preset character. At this time, loading the file, you can select a character in the characters provided by the file to mark, and also allow the input of the label.
  • the automatic labeling is triggered, and the similarity calculation is performed for each cluster set, all the pictures in the set and the cluster center picture, and the image with similarity greater than the threshold and the cluster center adopt the same Label the characters. You can repeat the above labeling operation by pressing the K-means clustering on the remaining unlabeled pictures until you abandon the loop or complete the labeling of all the characters.
  • the character labeling method provided by the embodiment obtains the character features of the marked person in the labeled person image by combining the skin color filtering and the face detection; and identifying the corresponding person by using the character feature similarity according to the acquired character characteristics of the marked character To cluster and mark the corresponding characters in the unlabeled person's picture, the accuracy and labeling efficiency are high.
  • step S100 includes:
  • Step S100A Acquiring a local file to obtain a character picture; collecting all the character pictures to form a character picture set.
  • the terminal retrieves the local file, specifies a local path containing the image, acquires the image file, performs face detection, acquires the character image, collects all the character images, forms a character image collection, and saves all the character images in the character image collection.
  • the character labeling method provided in this embodiment forms a collection of character images through face detection and image collection, so as to facilitate rapid labeling, and greatly improve labeling efficiency.
  • FIG. 3 is a schematic diagram of a refinement process of step S100 in FIG. 1. As shown in FIG. 3, the step S100 includes:
  • Step S110 combining the skin color filtering and the face detection to identify the person manually marked in the character picture.
  • the terminal converts the touch screen signal into a marking signal according to the touch screen action of the picker, thereby combining the skin color filtering and the face detection to identify the manually labeled person in the character picture.
  • Step S120 Obtain a character of a manually labeled person according to the identified manually labeled person Sign.
  • the terminal acquires features of the manually labeled person, such as the eye mask and facial morphological features, by means of face recognition and the like according to the identified manually labeled characters.
  • the character labeling method provided by the third embodiment after the step S200, includes:
  • Step S300 performing iterative labeling and/or recommendation sorting on the corresponding characters in the character picture that is still not marked after the cluster labeling.
  • the terminal can start the iterative labeling, select one or more pictures of the same person, and after completing the labeling of the character, trigger the recommended sorting of the terminal, and perform the unlabeled face image according to the current labeled character.
  • the sorting is displayed according to the similarity from high to low, so that it is convenient for manual further marking, and the front face is selected for labeling, thereby saving the searching time.
  • the recommendation item is a similarity ranking result of the currently selected face to be labeled and different characters.
  • the recommended implementation manner of the option is: comparing the currently selected face image with all the labeled images, sorting the comparison results by similarity, and comparing the characters corresponding to the comparison result as the option recommendation characters. Then trigger the first iteration, calculate the similarity between the picture of the marked person and all the unlabeled pictures, and automatically mark the unlabeled picture with the similarity higher than the threshold. After the last round of labeling, you can continue to select manual labeling or iteration. Since in the previous round, it is possible to automatically mark some pictures, so the iterative input picture changes. If you repeat it again, there may be more other pictures being labeled. .
  • the annotations can be manually added, and then the iterative annotation and the recommended sorting are repeated, until the character annotation of all face images is completed.
  • the above similarity uses the cosine similarity, and the cosine similarity is the cosine of the angle ⁇ of the two feature vectors. In the case of eigenvector normalization, the cosine similarity of the two vectors is proportional to its inner product.
  • the specific implementation manner of the recommended sorting can be described as follows. Let all the sets of face images be O, the picture set A marked as the current character, the unmarked picture set be B, and the picture set marked as other characters is C.
  • the character labeling method provided in this embodiment performs iterative labeling and/or recommendation sorting on the corresponding characters in the character images that are still not marked after the cluster labeling, thereby improving the efficiency of labeling.
  • the step S200 includes:
  • step S400 the character picture marked by the cluster is displayed by means of source image display, embedded display and/or thumbnail display.
  • the terminal stores the character pictures of all the picture collections according to the recognized character characteristics according to the identified corresponding characters, for example, storing the character pictures of the same person in the same folder for storage, and counting the number of the character pictures after classification.
  • the classification display may be performed by a source map, a face information embedded display, or a thumbnail display.
  • the detected face is highlighted by the gesture information, and the character name is also displayed on the picture by using the freetype (font) display Chinese method.
  • the images are grouped by the labeled characters to display thumbnails of all the pictures containing the corresponding characters, wherein the grouping can include single and multi-person pictures, and the face information is embedded in the display, thereby highlighting the selected characters, and Unmarking is available in the thumbnail display.
  • the character labeling method provided in this embodiment improves the labeling efficiency by means of source image display, embedded display and/or thumbnail display.
  • FIG. 6 is a schematic flowchart of a first embodiment of a terminal according to the present invention. As shown in FIG. 6, in the first embodiment, the terminal includes:
  • the obtaining module 10 is configured to combine the skin color filtering and the face detection to obtain the character features of the marked person in the labeled person picture;
  • the labeling module 20 is configured to identify the corresponding person by using the character feature similarity according to the obtained character feature of the marked person, to perform clustering labeling on the corresponding person in the unlabeled person picture.
  • the characters marked in the picture of the person may be manually labeled, or may be automatically marked by the terminal, or may be automatically labeled by the result of several manual markings.
  • the manual annotation can be multi-graphed at the same time, or it can be manually modified or removed manually;
  • the automatic annotation can also be a single or multiple icon annotation for a single character or multiple characters at the same time.
  • the acquiring module 10 of the terminal detects and recognizes the character of the person's picture by using the face information, and combines the skin color filtering and the face detection. Face detection is based on the framework of the V-J face detector.
  • the traditional VJ face detector extracts the Haar feature by scanning a large number of detection frames, and then performs the Adaboost algorithm to quickly filter out the non-human face frame.
  • the acquisition module 10 must contain most of the skin color according to the human face.
  • the prior knowledge of the area is used to detect the skin color of each face detection box.
  • the skin color detection result can quickly and effectively assist in judging whether it is a human face, and thus filtering out most areas as the previous level filtering.
  • the skin color likelihood of each pixel is obtained by a large number of training pictures, and the average skin color likelihood of the detection frame is calculated.
  • the average skin color likelihood is greater than the entire image, the retained picture is sent to the lower classifier. Otherwise it is filtered out.
  • This data unit contains person information, serial number information, feature vector information, source information, coordinates, and posture information of the face.
  • the source information is the source map path of the face as the read value
  • the serial number information is the design value
  • the coordinate information is the result value of the face detection, the feature vector and the pose information.
  • the dimension of the final feature vector is 200 dimensions, 100-dimensional is the LBP feature after dimension reduction, and LBP (Local Binary Pattern) 100-dimensional is the feature of HOG (Histogram of Oriented Gradient).
  • the HOG feature first uses the SDM algorithm to locate the feature points, and locates the left and right eyes, the left and right mouth corners and the nose tip.
  • the coordinates of the left and right corners can be used to calculate the position of the center of the mouth, using the position of the center of the mouth and the left and right eyes.
  • Point as the standard of alignment, map the three points to a fixed position on the 100*100 size image by affine change, and extract the LBP feature and the HOG feature from the face image of the 100*100 size, and then use the advance
  • the trained PCA and LDA dimension reduction matrices respectively reduce the dimension values of LBP and HOG features by 100 dimensions, respectively, and perform model normalization and tandem into 200-dimensional feature vectors.
  • the corresponding rotation mapping matrix can be inversely obtained, and a standard frontal face frame is used by the rotation mapping matrix. After the transformation, without rotation, the same three-dimensional face frame can be obtained, and then the face frame with obvious three-dimensional visual effect can be obtained through perspective projection.
  • the labeling module 20 of the terminal uses the character feature similarity to identify the corresponding person, and according to the identified corresponding person, the corresponding character in the unlabeled person image is clustered, wherein the clustering label uses the K-means poly.
  • Class method preset a k value, cluster all characters, the clustering result is divided into k categories, take the cluster center of each class for display and prompt labeling, and can read the preset character information for selection and labeling.
  • the terminal maintains a file that saves the preset character. When loading the file, you can select a character to be marked in the characters provided by the file, and also allow the input of the annotation.
  • the automatic labeling is triggered, and the similarity calculation is performed for each cluster set, all the pictures in the set and the cluster center picture, and the image with similarity greater than the threshold and the cluster center adopt the same Label the characters. You can repeat the above labeling operation by pressing the K-means clustering on the remaining unlabeled pictures until you abandon the loop or complete the labeling of all the characters.
  • the terminal obtained by the embodiment obtains the character features of the marked person in the labeled person image by combining the skin color filtering and the face detection; and identifying the corresponding person by using the character feature similarity according to the obtained character feature of the marked character; Clustering and labeling the corresponding characters in the unlabeled person pictures, the accuracy and labeling efficiency are high.
  • FIG. 7 is a schematic flowchart of a second embodiment of a terminal according to the present invention. As shown in FIG. 7, the terminal further includes:
  • the forming module 30 is configured to obtain a character picture by retrieving a local file, and collect all the character pictures to form a character picture set.
  • the terminal forming module 30 retrieves the local file, specifies a local path containing the image, acquires the image file, performs face detection, acquires the character image, collects all the character images, forms a character image collection, and saves all the character images in the In the collection of character pictures.
  • the terminal provided in this embodiment forms a set of character pictures through face detection and picture collection, so as to facilitate rapid labeling, which greatly improves the labeling efficiency.
  • FIG. 8 is a schematic diagram of the function module of the acquiring module in FIG. 6. As shown in FIG. 8, the acquiring module 10 includes:
  • the identification unit 11 is configured to identify a person manually marked in the person picture in combination with skin color filtering and face detection;
  • the obtaining unit 12 is configured to acquire a character feature of the manually labeled person according to the identified manually labeled person.
  • the recognition unit 11 of the terminal converts the touch screen signal into a sign signal according to the touch screen action of the pickup person, thereby identifying the person manually marked in the person picture in combination with the skin color filtering and the face detection.
  • the acquiring unit 12 of the terminal acquires features of the manually labeled person, such as eye mask and facial morphological features, by means of face recognition or the like according to the identified manually labeled person.
  • FIG. 9 is a schematic flowchart of a third embodiment of a terminal according to the present invention. As shown in FIG. 9, the terminal further includes:
  • the sorting module 30 is configured to perform iterative labeling and/or recommendation sorting on the corresponding characters in the character pictures that are still not labeled after the cluster labeling.
  • the terminal may start iterative labeling, select one or more pictures of the same person, and after completing the labeling of the character, trigger the recommended sorting of the terminal, according to the current labeling character, the unlabeled
  • the face image carries out a comprehensive sorting of the current characters, and the display is sorted according to the similarity from high to low, thereby facilitating manual further labeling, and selecting the front face to mark, saving the search time.
  • the recommendation operation module there is an option recommendation when inputting, and the recommendation item is a similarity ranking result of the currently selected face to be labeled and different characters.
  • the recommended implementation manner of the option is: comparing the currently selected face image with all the labeled images, sorting the comparison results by similarity, and comparing the characters corresponding to the comparison result as the option recommendation characters. Then trigger the first iteration, calculate the similarity between the picture of the marked person and all the unlabeled pictures, and automatically mark the unlabeled picture with the similarity higher than the threshold. After the last round of labeling, you can continue to select manual labeling or iteration. Since in the previous round, it is possible to automatically mark some pictures, so the iterative input picture changes. If you repeat it again, there may be more other pictures being labeled. .
  • the annotations can be manually added, and then the iterative annotation and the recommended sorting are repeated, until the character annotation of all face images is completed.
  • the above similarity uses the cosine similarity, and the cosine similarity is the cosine of the angle ⁇ of the two feature vectors. In the case of eigenvector normalization, the cosine similarity of the two vectors is proportional to its inner product.
  • the specific implementation manner of the recommended sorting can be described as follows. Let all the sets of face images be O, the picture set A marked as the current character, the unmarked picture set be B, and the picture set marked as other characters is C.
  • the terminal provided in this embodiment performs iterative labeling and/or recommendation sorting on the corresponding characters in the character picture that has not been marked after the clustering is marked, thereby improving the efficiency of labeling.
  • FIG. 10 is a schematic flowchart of a fourth embodiment of a terminal according to the present invention. As shown in FIG. 10, on the basis of the first embodiment, the terminal further includes:
  • the display module 40 is configured to display the character images of the clustering by means of source image display, embedded display, and/or thumbnail display.
  • the display module 40 of the terminal stores the character pictures of all the picture sets according to the recognized character features according to the identified corresponding characters, for example, storing the character pictures of the same person in the same folder for storage, and counting the classified characters.
  • the number of pictures is displayed in a classified manner, and the display manner may be displayed according to the source image, or may be embedded in the face information display, or may be a thumbnail display.
  • the face information is embedded and displayed, the detected face is highlighted by the gesture information, and the character name is also displayed on the picture by using the freetype (font) display Chinese method.
  • the images are grouped by the labeled characters to display thumbnails of all the pictures containing the corresponding characters, wherein the grouping can include single and multi-person pictures, and the face information is embedded in the display, thereby highlighting the selected characters, and Unmarking is available in the thumbnail display.
  • the terminal provided by this embodiment improves the labeling efficiency by means of source image display, embedded display and/or thumbnail display.
  • the acquiring module, the labeling module, the forming module, the sorting module, and the display module in the terminal provided by the embodiment of the present invention, and each unit included in each module may be implemented by a processor in the terminal;
  • the logic circuit is implemented; in a specific embodiment, the processor may be a central processing unit (CPU), a microprocessor (MPU), a digital signal processor (DSP), or a field programmable gate array (FPGA).
  • the processor may be a central processing unit (CPU), a microprocessor (MPU), a digital signal processor (DSP), or a field programmable gate array (FPGA).
  • CPU central processing unit
  • MPU microprocessor
  • DSP digital signal processor
  • FPGA field programmable gate array
  • the character labeling method described above is implemented in the form of a software function module and sold or used as an independent product, it may also be stored in a computer readable storage medium.
  • the technical solution of the embodiment of the present invention The portion of the quality or contribution to the prior art may be embodied in the form of a software product stored in a storage medium, including instructions for causing a computer device (which may be a personal computer, a server) Or, a network device, etc.) performs all or part of the methods described in various embodiments of the present invention.
  • the foregoing storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read only memory (ROM), a magnetic disk, or an optical disk.
  • program codes such as a USB flash drive, a mobile hard disk, a read only memory (ROM), a magnetic disk, or an optical disk.
  • the embodiment of the present invention further provides a computer storage medium, where the computer storage medium stores computer executable instructions, and the computer executable instructions are used to execute the character labeling method provided in the embodiments of the present invention.
  • the character of the marked person in the picture of the character is obtained by combining the skin color filtering and the face detection; and the corresponding character is identified by using the feature similarity of the character according to the obtained character feature of the marked character,
  • the corresponding characters in the unlabeled person picture are clustered and labeled. Therefore, the technical solution provided by the embodiment of the present invention has the advantages of high accuracy and high efficiency.

Landscapes

  • Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Oral & Maxillofacial Surgery (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • General Health & Medical Sciences (AREA)
  • Human Computer Interaction (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Multimedia (AREA)
  • Theoretical Computer Science (AREA)
  • Image Analysis (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

本发明实施例公开了一种人物标注方法，通过结合肤色过滤和人脸检测以获取已标注人物图片中已标注人物的人物特征；根据获取的已标注人物的人物特征，利用人物特征相似度识别对应的人物，以对未标注人物图片中对应的人物进行聚类标注。本发明实施例还提供一种终端和存储介质。

Description

人物标注方法和终端、存储介质 技术领域
本发明涉及通讯领域，尤其涉及人物标注方法和终端、存储介质。
背景技术
随着科技的飞速发展，人脸检测在各方面得到广泛应用，然而关于人脸的标注，尤其是对人物标注，往往只能依靠人手逐一标注，或者是利用文本从人脸图片上下文中获得人物信息，前者费时费力，效率低下，而后者的准确度对上下文的依赖性大，往往达不到理想的效果。因此，如何设计一种准确度高、且标注效率高的人物标注方法，是一个亟待解决的问题。
发明内容
本发明实施例为解决上述的问题之一而提供一种人物标注方法和终端、存储介质，旨在解决人物标注准确度低、且标注效率低的问题。
本发明实施例提供的技术方案如下：
第一方面，本发明实施例提供一种人物标注方法，所述人物标注方法包括：
结合肤色过滤和人脸检测以获取已标注人物图片中已标注人物的人物特征；
根据获取的已标注人物的人物特征，利用人物特征相似度识别对应的人物，以对未标注人物图片中对应的人物进行聚类标注。
在本发明的一种实施例中，所述结合肤色过滤和人脸检测以获取已标注人物图片中已标注人物的人物特征的步骤之前包括：
通过检索本地文件，获取人物图片；集合所有人物图片，形成人物图 片集合。
在本发明的一种实施例中，所述结合肤色过滤和人脸检测以获取已标注人物图片中已标注人物的人物特征的步骤包括：
结合肤色过滤和人脸检测识别人物图片中手动标注的人物；
根据识别的手动标注的人物，获取手动标注人物的人物特征。
在本发明的一种实施例中，根据获取的已标注人物的人物特征，利用人物特征相似度识别对应的人物，以对未标注人物图片中对应的人物进行聚类标注的步骤之后包括：
对聚类标注后仍未标注的人物图片中对应的人物进行迭代标注和/或推荐排序。
在本发明的一种实施例中，根据获取的已标注人物的人物特征，利用人物特征相似度识别对应的人物，以对未标注人物图片中对应的人物进行聚类标注的步骤之后包括：
通过源图显示、嵌入显示和/或缩略图显示的方式，对聚类标注的人物图片进行显示。
第二方面，本发明实施例进一步提供一种终端，所述终端包括：
获取模块，配置为结合肤色过滤和人脸检测以获取已标注人物图片中已标注人物的人物特征；
标注模块，配置为根据获取的已标注人物的人物特征，利用人物特征相似度识别对应的人物，以对未标注人物图片中对应的人物进行聚类标注。
在本发明的一种实施例中，所述终端还包括：
形成模块，配置为通过检索本地文件，获取人物图片；集合所有人物图片，形成人物图片集合。
在本发明的一种实施例中，所述获取模块包括：
识别单元，配置为结合肤色过滤和人脸检测识别人物图片中手动标注 的人物；
获取单元，配置为根据识别的手动标注的人物，获取手动标注人物的人物特征。
在本发明的一种实施例中，所述终端还包括：
排序模块，配置为对聚类标注后仍未标注的人物图片中对应的人物进行迭代标注和/或推荐排序。
在本发明的一种实施例中，所述终端还包括：
显示模块，配置为通过源图显示、嵌入显示和/或缩略图显示的方式，对聚类标注的人物图片进行显示。
第三方面，本发明实施例提供一种计算机存储介质，所述计算机存储介质中存储有计算机可执行指令，该计算机可执行指令用于执行本发明第一方面实施例提供的人物标注方法。
本发明实施例提供的人物标注方法和终端、存储介质，通过结合肤色过滤和人脸检测以获取已标注人物图片中已标注人物的人物特征；根据获取的已标注人物的人物特征，利用人物特征相似度识别对应的人物，以对未标注人物图片中对应的人物进行聚类标注。本发明标注准确度和效率高。
附图说明
图1为本发明人物标注方法第一实施例的流程示意图；
图2为本发明人物标注方法第二实施例的流程示意图；
图3为图1中所述结合肤色过滤和人脸检测以获取已标注人物图片中已标注人物的人物特征的步骤的细化流程示意图；
图4为本发明人物标注方法第三实施例的流程示意图；
图5为本发明人物标注方法第四实施例的流程示意图；
图6为本发明终端第一实施例的功能模块示意图；
图7为本发明终端第二实施例的功能模块示意图；
图8为图6中所述获取模块的功能模块示意图；
图9为本发明终端第三实施例的功能模块示意图；
图10为本发明终端第四实施例的功能模块示意图；
本发明目的的实现、功能特点及优点将结合实施例，参照附图做进一步说明。
具体实施方式
应当理解，此处所描述的具体实施例仅仅用以解释本发明的技术方案，并不用于限定本发明的保护范围。
本发明实施例提供一种人物标注方法，参照图1，在第一实施例中，所述人物标注方法包括：
步骤S100、结合肤色过滤和人脸检测以结合肤色过滤和人脸检测以获取已标注人物图片中已标注人物的人物特征。
在本实施例中，已标注人物图片中标注的人物可以为手动标注，也可以以为终端自动标注，还可以是利用若干次人工标注的结果，对未标注的人脸进行自动标注。其中，人工标注可以是多图同时标注，也可以是通过人工手工修改或移除标注；自动标注也可以是同时对单个人物或是多个人物进行单图或多图标注。本实施例中，采用人脸信息对人物图片的人物特征进行检测和识别，并结合肤色过滤和人脸检测技术。人脸检测是基于V-J人脸检测器的框架实现的。传统的V-J人脸检测器通过扫描大量检测框，提取Haar特征，再进行Adaboost算法快速滤掉非人脸的图片框，而在本实施例中，根据人脸中一定包含大部分肤色区域的先验知识，对每个人脸检测框先进行肤色检测，通过肤色检测结果可以快速有效地协助判断是否为人脸，从而作为前一级滤波，过滤掉大部分区域。具体肤色检测时通过大量训练图片得到每个像素的肤色似然度，计算检测框的平均肤色似然度，当大于整幅图像平均肤色似然度时将被保留的图片送入下级分类器，否则滤 掉。然后遍历检测当前本地路径下的所有图片，保存人脸数据，创建关于每个人脸的一个数据单元。此数据单元含有该人脸的人物信息、序号信息、特征向量信息、来源信息、坐标和姿态信息。创建时相关人物信息为空，后续的标注操作会改变这个信息项，来源信息即该人脸的源图路径为读取值，序号信息为设计值，按序递加，坐标信息为人脸检测的结果值，特征向量和姿态信息则为计算结果值。最终特征向量的维度为200维，其中100维是降维后的LBP特征，LBP(Local Binary Pattern，局部二值模式)100维是降维后的HOG(Histogram of Oriented Gradient，方向梯度直方图)特征。HOG特征首先利用SDM算法通过进行特征点定位，定位出左右眼，左右嘴角和鼻尖共五个特征点，通过左右嘴角的坐标可以计算出嘴巴中心的位置，利用嘴巴中心的位置和左右眼三个点作为对齐的标准，通过仿射变化将此三点映射到100*100大小图像上的固定的位置，进行在该100*100大小的人脸图像生提取LBP特征和HOG特征，然后分别利用预先训练好的PCA和LDA降维矩阵分别对LBP和HOG特征降维值100维，分别进行模归一化，串联成200维的特征向量。通过定位好的五个人脸特征点，以及通用三维人脸模型中的对应点的三维坐标，可以反求出对应的旋转映射矩阵，利用此旋转映射矩阵对一个标准的正面人脸框(边长相等，且无旋转)进行变换，便可以得到与人脸姿态相同的三维人脸框，然后可以通过透视投影得到具有明显三维视觉效果的人脸框。
步骤S200、根据获取的已标注人物的人物特征，利用人物特征相似度识别对应的人物，以对未标注人物图片中对应的人物进行聚类标注。
在本实施中，利用人物特征相似度识别对应的人物，根据识别的对应人物，对未标注人物图片中对应的人物采用聚类标注，其中，聚类标注使用K-均值聚类方式，预设一个k值，对所有人物进行聚类，聚类结果分为k类，取每个类的聚类中心进行显示和提示标注，可以读取预设的人物信息 进行选择标注，终端维护一个保存预设人物的文件，这时载入这个文件，就可以在文件提供的人物中选择一个人物进行标注，也允许输入标注。完成这k个聚类中心的人工标注后，触发自动标注，对每个聚类集合，集合内的所有图片和聚类中心图片进行相似度计算，相似度大于阈值的图片和聚类中心采用同样的标注人物。可以对剩余的未标注的图片再次按K-均值聚类，重复上述标注操作，直到放弃循环或者完成所有人物图片的标注。
本实施例提供的人物标注方法，通过结合肤色过滤和人脸检测以获取已标注人物图片中已标注人物的人物特征；根据获取的已标注人物的人物特征，利用人物特征相似度识别对应的人物，以对未标注人物图片中对应的人物进行聚类标注，准确度和标注效率高。
图2为本发明人物标注方法第二实施例的流程示意图，如图2所示，在第一实施例的基础上，所述步骤S100之前包括：
步骤S100A、通过检索本地文件，获取人物图片；集合所有人物图片，形成人物图片集合。
终端检索本地文件，指定一个含有图片的本地路径，获取图片文件，进行人脸检测，获取人物图片，对所有的人物图片进行集合，形成人物图片集合，将所有的人物图片保存在人物图片集合中。
本实施例提供的人物标注方法，通过人脸检测和图片集合，形成人物图片集合，以利于快速标注，大大提升了标注效率。
图3为图1中所述步骤S100的细化流程示意图，如图3所示，所述步骤S100包括：
步骤S110、结合肤色过滤和人脸检测识别人物图片中手动标注的人物。
终端根据拾取人的触屏动作，将触屏信号转换为标示信号，从而结合肤色过滤和人脸检测识别人物图片中手动标注的人物。
步骤S120、根据识别的手动标注的人物，获取手动标注人物的人物特 征。
终端根据识别的手动标注的人物，通过人脸识别等手段，获取手动标注人物的特征，如眼膜、脸部形态特征。
图4为本发明人物标注方法第三实施例的流程示意图，如图4所示，在第一实施例的基础上，第三实施例提供的人物标注方法，所述步骤S200之后包括：
步骤S300、对聚类标注后仍未标注的人物图片中对应的人物进行迭代标注和/或推荐排序。
结束聚类标注环节后，终端可以开始迭代标注，选取一个或多个同人物的图片，完成该人物的标注后，触发终端的推荐排序，根据当前的标注人物，对未标注的人脸图片进行关于当前人物的综合排序，按相似度从高到低排序显示，从而便于人工的进一步标注，选取靠前的人脸进行标注，节约查找时间。在所述标注操作模块中，输入时有选项推荐，推荐项为当前选择的要标注的人脸与不同人物的相似度排序结果。选项推荐的实施方式是：将当前被选中的人脸图片和所有已标注图片进行比较，按相似度排序比较结果，将比较结果对应的人物作为选项推荐人物。然后触发第一轮迭代，对已标注人物的图片和所有未标注人物图片计算相似度，对相似度高于阈值的未标注人物图片自动进行标注。待上一轮标注完之后，可以继续选择人工标注或者迭代，由于在上一轮中，有可能自动标注了一些图片，因此迭代的输入图片改变了，再次迭代可能会有更多其他图片被标注。而针对那些多次迭代依然不能被自动标注的，可以人工补充标注，然后采用迭代标注和推荐排序重复交叉的方式，直至完成所有人脸图片的人物标注。上述的相似度使用的是余弦相似度，余弦相似度就是两个特征向量夹角θ的余弦值，
Figure PCTCN2015073337-appb-000001
在特征向量模归一化的情况下，两个向量的余弦相 似度和其内积成比例。推荐排序的具体实施方式可描述如下，设所有人脸图片的集合为O，已标注为当前人物的图片集A，未标注图片集为B，标注为其他人物的图片集为C。在B中取一个图片X和A中所有图片计算相似度，取最大值作为图片X对A的相似度PX。之后再将X与C中所有图片计算相似度，如果存在相似度PC大于PX，则认为图片X更接近非A的人物，故将PX减1。再完成B中所有图片后，对PX降序排序，只有那些没有被减1的图片会排在前面。这些就是推荐的结果。
本实施例提供的人物标注方法，对聚类标注后仍未标注的人物图片中对应的人物进行迭代标注和/或推荐排序，从而提高标注的效率。
图5为本发明人物标注方法第四实施例的流程示意图，如图5所示，在第一实施例的基础上，第四实施例提供的人物标注方法，所述步骤S200之后包括：
步骤S400、通过源图显示、嵌入显示和/或缩略图显示的方式，对聚类标注的人物图片进行显示。
终端根据识别的对应的人物，对所有图片集合的人物图片按照识别的人物特征，分类进行储存，例如将同一人物的人物图片放在同一文件夹下进行存储，并统计分类后人物图片的数量，进行分类显示，所述显示的方式可以是按源图显示，也可以是人脸信息嵌入显示，还可以是缩略图显示。其中人脸信息嵌入显示时，利用姿态信息突出标识检测到的人脸，利用了freetype(字体)显示中文的方法把人物姓名也显示在图片上。利用标注的人物对图片进行分组，显示含有对应人物的所有图片的缩略图，其中，分组中可以含单人和多人图片，并将人脸信息嵌入显示，从而突出显示选中的人物，并在缩略图显示中提供取消标注的功能。
本实施例提供的人物标注方法，通过源图显示、嵌入显示和/或缩略图显示的方式，提高了标注效率。
图6为本发明终端第一实施例的流程示意图，如图6所示，在第一实施例中，所述终端包括：
获取模块10，配置为结合肤色过滤和人脸检测以获取已标注人物图片中已标注人物的人物特征；
标注模块20，配置为根据获取的已标注人物的人物特征，利用人物特征相似度识别对应的人物，以对未标注人物图片中对应的人物进行聚类标注。
在本实施例中，已标注人物图片中标注的人物可以为手动标注，也可以以为终端自动标注，还可以是利用若干次人工标注的结果，对未标注的人脸进行自动标注。其中，人工标注可以是多图同时标注，也可以是通过人工手工修改或移除标注；自动标注也可以是同时对单个人物或是多个人物进行单图或多图标注。本实施例中，终端的获取模块10采用人脸信息对人物图片的人物特征进行检测和识别，并结合肤色过滤和人脸检测。人脸检测基于V-J人脸检测器的框架实现的。传统的V-J人脸检测器通过扫描大量检测框，提取Haar特征，再进行Adaboost算法快速滤掉非人脸的图片框，而在本实施例中，获取模块10根据人脸中一定包含大部分肤色区域的先验知识，对每个人脸检测框先进行肤色检测，通过肤色检测结果可以快速有效地协助判断是否为人脸，从而作为前一级滤波，过滤掉大部分区域。具体肤色检测时通过大量训练图片得到每个像素的肤色似然度，计算检测框的平均肤色似然度，当大于整幅图像平均肤色似然度时将被保留的图片送入下级分类器，否则滤掉。然后遍历检测当前本地路径下的所有图片，保存人脸数据，创建关于每个人脸的一个数据单元。此数据单元含有该人脸的人物信息、序号信息、特征向量信息、来源信息、坐标和姿态信息。创建时相关人物信息为空，后续的标注操作会改变这个信息项，来源信息即该人脸的源图路径为读取值，序号信息为设计值，按序递加，坐标信息为 人脸检测的结果值，特征向量和姿态信息则为计算结果值。最终特征向量的维度为200维，其中100维是降维后的LBP特征，LBP(Local BinaryPattern，局部二值模式)100维是降维后的HOG(Histogram of OrientedGradient，方向梯度直方图)特征。HOG特征首先利用SDM算法通过进行特征点定位，定位出左右眼，左右嘴角和鼻尖共五个特征点，通过左右嘴角的坐标可以计算出嘴巴中心的位置，利用嘴巴中心的位置和左右眼三个点作为对齐的标准，通过仿射变化将此三点映射到100*100大小图像上的固定的位置，进行在该100*100大小的人脸图像生提取LBP特征和HOG特征，然后分别利用预先训练好的PCA和LDA降维矩阵分别对LBP和HOG特征降维值100维，分别进行模归一化，串联成200维的特征向量。通过定位好的五个人脸特征点，以及通用三维人脸模型中的对应点的三维坐标，可以反求出对应的旋转映射矩阵，利用此旋转映射矩阵对一个标准的正面人脸框(边长相等，且无旋转)进行变换，便可以得到与人脸姿态相同的三维人脸框，然后可以通过透视投影得到具有明显三维视觉效果的人脸框。
在本实施中，终端的标注模块20利用人物特征相似度识别对应的人物，根据识别的对应人物，对未标注人物图片中对应的人物采用聚类标注，其中，聚类标注使用K-均值聚类方式，预设一个k值，对所有人物进行聚类，聚类结果分为k类，取每个类的聚类中心进行显示和提示标注，可以读取预设的人物信息进行选择标注，终端维护一个保存预设人物的文件，这时载入这个文件，就可以在文件提供的人物中选择一个人物进行标注，也允许输入标注。完成这k个聚类中心的人工标注后，触发自动标注，对每个聚类集合，集合内的所有图片和聚类中心图片进行相似度计算，相似度大于阈值的图片和聚类中心采用同样的标注人物。可以对剩余的未标注的图片再次按K-均值聚类，重复上述标注操作，直到放弃循环或者完成所有人物图片的标注。
本实施例提供的终端，通过结合肤色过滤和人脸检测以获取已标注人物图片中已标注人物的人物特征；根据获取的已标注人物的人物特征，利用人物特征相似度识别对应的人物，以对未标注人物图片中对应的人物进行聚类标注，准确度和标注效率高。
图7为本发明终端第二实施例的流程示意图，如图7所示，在第一实施例的基础上，所述终端还包括：
形成模块30，配置为通过检索本地文件，获取人物图片；集合所有人物图片，形成人物图片集合。
终端的形成模块30检索本地文件，指定一个含有图片的本地路径，获取图片文件，进行人脸检测，获取人物图片，对所有的人物图片进行集合，形成人物图片集合，将所有的人物图片保存在人物图片集合中。
本实施例提供的终端，通过人脸检测和图片集合，形成人物图片集合，以利于快速标注，大大提升了标注效率。
图8为图6中所述获取模块的功能模块示意图，如图8所示，所述获取模块10包括：
识别单元11，配置为结合肤色过滤和人脸检测识别人物图片中手动标注的人物；
获取单元12，配置为根据识别的手动标注的人物，获取手动标注人物的人物特征。
终端的识别单元11根据拾取人的触屏动作，将触屏信号转换为标示信号，从而结合肤色过滤和人脸检测识别人物图片中手动标注的人物。
终端的获取单元12根据识别的手动标注的人物，通过人脸识别等手段，获取手动标注人物的特征，如眼膜、脸部形态特征。
图9为本发明终端第三实施例的流程示意图，如图9所示，在第一实施例的基础上，所述终端还包括：
排序模块30，配置为对聚类标注后仍未标注的人物图片中对应的人物进行迭代标注和/或推荐排序。
终端的排序模块30结束聚类标注环节后，终端可以开始迭代标注，选取一个或多个同人物的图片，完成该人物的标注后，触发终端的推荐排序，根据当前的标注人物，对未标注的人脸图片进行关于当前人物的综合排序，按相似度从高到低排序显示，从而便于人工的进一步标注，选取靠前的人脸进行标注，节约查找时间。在所述标注操作模块中，输入时有选项推荐，推荐项为当前选择的要标注的人脸与不同人物的相似度排序结果。选项推荐的实施方式是：将当前被选中的人脸图片和所有已标注图片进行比较，按相似度排序比较结果，将比较结果对应的人物作为选项推荐人物。然后触发第一轮迭代，对已标注人物的图片和所有未标注人物图片计算相似度，对相似度高于阈值的未标注人物图片自动进行标注。待上一轮标注完之后，可以继续选择人工标注或者迭代，由于在上一轮中，有可能自动标注了一些图片，因此迭代的输入图片改变了，再次迭代可能会有更多其他图片被标注。而针对那些多次迭代依然不能被自动标注的，可以人工补充标注，然后采用迭代标注和推荐排序重复交叉的方式，直至完成所有人脸图片的人物标注。上述的相似度使用的是余弦相似度，余弦相似度就是两个特征向量夹角θ的余弦值，
Figure PCTCN2015073337-appb-000002
在特征向量模归一化的情况下，两个向量的余弦相似度和其内积成比例。推荐排序的具体实施方式可描述如下，设所有人脸图片的集合为O，已标注为当前人物的图片集A，未标注图片集为B，标注为其他人物的图片集为C。在B中取一个图片X和A中所有图片计算相似度，取最大值作为图片X对A的相似度PX。之后再将X与C中所有图片计算相似度，如果存在相似度PC大于PX，则认为图片X更接近非A的人物，故将PX减1。再完成B中所有图片后，对PX降序排序，只有那些没有被减1的图片会排在前面。这些就是推荐的结果。
本实施例提供的终端，对聚类标注后仍未标注的人物图片中对应的人物进行迭代标注和/或推荐排序，从而提高标注的效率。
图10为本发明终端第四实施例的流程示意图，如图10所示，在第一实施例的基础上，所述终端还包括：
显示模块40，配置为通过源图显示、嵌入显示和/或缩略图显示的方式，对聚类标注的人物图片进行显示。
终端的显示模块40根据识别的对应的人物，对所有图片集合的人物图片按照识别的人物特征，分类进行储存，例如将同一人物的人物图片放在同一文件夹下进行存储，并统计分类后人物图片的数量，进行分类显示，所述显示的方式可以是按源图显示，也可以是人脸信息嵌入显示，还可以是缩略图显示。其中人脸信息嵌入显示时，利用姿态信息突出标识检测到的人脸，利用了freetype(字体)显示中文的方法把人物姓名也显示在图片上。利用标注的人物对图片进行分组，显示含有对应人物的所有图片的缩略图，其中，分组中可以含单人和多人图片，并将人脸信息嵌入显示，从而突出显示选中的人物，并在缩略图显示中提供取消标注的功能。
本实施例提供的终端，通过源图显示、嵌入显示和/或缩略图显示的方式，提高了标注效率。
本发明实施例提供的终端中的获取模块、标注模块、形成模块、排序模块和显示模块，以及各模块各自所包括的各单元都可以通过终端中的处理器来实现；当然也可通过具体的逻辑电路实现；在具体实施例的过程中，处理器可以为中央处理器(CPU)、微处理器(MPU)、数字信号处理器(DSP)或现场可编程门阵列(FPGA)等。
需要说明的是，本发明实施例中，如果以软件功能模块的形式实现上述的人物标注方法，并作为独立的产品销售或使用时，也可以存储在一个计算机可读取存储介质中。基于这样的理解，本发明实施例的技术方案本 质上或者说对现有技术做出贡献的部分可以以软件产品的形式体现出来，该计算机软件产品存储在一个存储介质中，包括若干指令用以使得一台计算机设备(可以是个人计算机、服务器、或者网络设备等)执行本发明各个实施例所述方法的全部或部分。而前述的存储介质包括：U盘、移动硬盘、只读存储器(ROM，Read Only Memory)、磁碟或者光盘等各种可以存储程序代码的介质。这样，本发明实施例不限制于任何特定的硬件和软件结合。
相应地，本发明实施例再提供一种计算机存储介质，所述计算机存储介质中存储有计算机可执行指令，该计算机可执行指令用于执行本发明各实施例中提供的人物标注方法。
以上仅为本发明的优选实施例，并非因此限制本发明的专利范围，凡是利用本发明说明书及附图内容所作的等效结构或等效流程变换，或直接或间接运用在其他相关的技术领域，均同理包括在本发明的专利保护范围内。
工业实用性
本发明实施例中，通过结合肤色过滤和人脸检测以获取已标注人物图片中已标注人物的人物特征；根据获取的已标注人物的人物特征，利用人物特征相似度识别对应的人物，以对未标注人物图片中对应的人物进行聚类标注，如此，本发明实施例提供的技术方案，具有标注准确度和效率高的优点。

Claims (11)

  1. 一种人物标注方法，所述人物标注方法包括：
    结合肤色过滤和人脸检测以结合肤色过滤和人脸检测以获取已标注人物图片中已标注人物的人物特征；
    根据获取的已标注人物的人物特征，利用人物特征相似度识别对应的人物，以对未标注人物图片中对应的人物进行聚类标注。
  2. 如权利要求1所述的人物标注方法，其中，所述结合肤色过滤和人脸检测以获取已标注人物图片中已标注人物的人物特征的步骤之前包括：
    通过检索本地文件，获取人物图片；集合所有人物图片，形成人物图片集合。
  3. 如权利要求2所述的人物标注方法，其中，所述结合肤色过滤和人脸检测以结合肤色过滤和人脸检测以获取已标注人物图片中已标注人物的人物特征的步骤包括：
    结合肤色过滤和人脸检测结合肤色过滤和人脸检测识别人物图片中手动标注的人物；
    根据识别的手动标注的人物，获取手动标注人物的人物特征。
  4. 如权利要求1所述的人物标注方法，其中，所述根据获取的已标注人物的人物特征，利用人物特征相似度识别对应的人物，以对未标注人物图片中对应的人物进行聚类标注的步骤之后包括：
    对聚类标注后仍未标注的人物图片中对应的人物进行迭代标注和/或推荐排序。
  5. 如权利要求1至4任一项所述的人物标注方法，其中，所述根据获取的已标注人物的人物特征，利用人物特征相似度识别对应的人物，以对未标注人物图片中对应的人物进行聚类标注的步骤之后包括：
    通过源图显示、嵌入显示和/或缩略图显示的方式，对聚类标注的人物图片进行显示。
  6. 一种终端，所述终端包括：
    获取模块，配置为结合肤色过滤和人脸检测以获取已标注人物图片中已标注人物的人物特征；
    标注模块，配置为根据获取的已标注人物的人物特征，利用人物特征相似度识别对应的人物，以对未标注人物图片中对应的人物进行聚类标注。
  7. 如权利要求6所述的终端，其中，所述终端还包括：
    形成模块，配置为通过检索本地文件，获取人物图片；集合所有人物图片，形成人物图片集合。
  8. 如权利要求6所述的终端，其中，所述获取模块包括：
    识别单元，配置为结合肤色过滤和人脸检测识别人物图片中手动标注的人物；
    获取单元，配置为根据识别的手动标注的人物，获取手动标注人物的人物特征。
  9. 如权利要求6至8任一项所述的终端，其中，所述终端还包括：
    排序模块，配置为对聚类标注后仍未标注的人物图片中对应的人物进行迭代标注和/或推荐排序。
  10. 如权利要求6至8任一项所述的终端，其中，所述终端还包括：
    显示模块，配置为通过源图显示、嵌入显示和/或缩略图显示的方式，对聚类标注的人物图片进行显示。
  11. 一种计算机存储介质，所述计算机存储介质中存储有计算机可执行指令，该计算机可执行指令用于执行权利要求1至5任一项所述的人物标注方法。
PCT/CN2015/073337 2014-12-31 2015-02-27 人物标注方法和终端、存储介质 WO2016106966A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201410851972.1 2014-12-31
CN201410851972.1A CN105809096A (zh) 2014-12-31 2014-12-31 人物标注方法和终端

Publications (1)

Publication Number Publication Date
WO2016106966A1 true WO2016106966A1 (zh) 2016-07-07

Family

ID=56284052

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2015/073337 WO2016106966A1 (zh) 2014-12-31 2015-02-27 人物标注方法和终端、存储介质

Country Status (2)

Country Link
CN (1) CN105809096A (zh)
WO (1) WO2016106966A1 (zh)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111381743A (zh) * 2018-12-29 2020-07-07 杭州光启人工智能研究院 数据打标方法、计算机装置及计算机可读存储介质
CN112863493A (zh) * 2021-01-14 2021-05-28 北京天行汇通信息技术有限公司 语音数据标注方法和装置，电子设备
CN113657173A (zh) * 2021-07-20 2021-11-16 北京搜狗科技发展有限公司 一种数据处理方法、装置和用于数据处理的装置

Families Citing this family (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107391703B (zh) * 2017-07-28 2019-11-15 北京理工大学 图像库的建立方法及系统、图像库和图像分类方法
CN107657269A (zh) * 2017-08-24 2018-02-02 百度在线网络技术（北京）有限公司 一种用于训练图片提纯模型的方法和装置
CN108229384A (zh) * 2017-12-29 2018-06-29 广州图语信息科技有限公司 一种利用连续性结构的人脸聚类方法、装置及用户终端
CN109658572B (zh) 2018-12-21 2020-09-15 上海商汤智能科技有限公司 图像处理方法及装置、电子设备和存储介质
CN112766296B (zh) * 2019-11-06 2023-04-07 济南信通达电气科技有限公司 输电线路安全隐患目标检测模型训练方法及装置
CN111353059A (zh) * 2020-03-02 2020-06-30 腾讯科技（深圳）有限公司 图片处理方法和装置、计算机可读的存储介质及电子装置
CN111695628B (zh) * 2020-06-11 2023-05-05 北京百度网讯科技有限公司 关键点标注方法、装置、电子设备及存储介质

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101344922A (zh) * 2008-08-27 2009-01-14 华为技术有限公司 一种人脸检测方法及其装置
CN103839076A (zh) * 2014-02-25 2014-06-04 中国科学院自动化研究所 一种基于光照特征的网络敏感图像识别方法
CN104133875A (zh) * 2014-07-24 2014-11-05 北京中视广信科技有限公司 一种基于人脸的视频标注方法和视频检索方法

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN100550038C (zh) * 2007-12-26 2009-10-14 腾讯科技（深圳）有限公司 图片内容识别方法及识别系统
CN101795400B (zh) * 2010-03-16 2013-03-27 上海复控华龙微系统技术有限公司 一种婴幼儿主动跟踪监护的方法及其实现系统
CN103473275A (zh) * 2013-08-23 2013-12-25 中山大学 一种采用多特征融合的图像自动标注方法和系统
CN104217008B (zh) * 2014-09-17 2018-03-13 中国科学院自动化研究所 互联网人物视频交互式标注方法及系统

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101344922A (zh) * 2008-08-27 2009-01-14 华为技术有限公司 一种人脸检测方法及其装置
CN103839076A (zh) * 2014-02-25 2014-06-04 中国科学院自动化研究所 一种基于光照特征的网络敏感图像识别方法
CN104133875A (zh) * 2014-07-24 2014-11-05 北京中视广信科技有限公司 一种基于人脸的视频标注方法和视频检索方法

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111381743A (zh) * 2018-12-29 2020-07-07 杭州光启人工智能研究院 数据打标方法、计算机装置及计算机可读存储介质
CN111381743B (zh) * 2018-12-29 2022-07-12 深圳光启高等理工研究院 数据打标方法、计算机装置及计算机可读存储介质
CN112863493A (zh) * 2021-01-14 2021-05-28 北京天行汇通信息技术有限公司 语音数据标注方法和装置，电子设备
CN113657173A (zh) * 2021-07-20 2021-11-16 北京搜狗科技发展有限公司 一种数据处理方法、装置和用于数据处理的装置
CN113657173B (zh) * 2021-07-20 2024-05-24 北京搜狗科技发展有限公司 一种数据处理方法、装置和用于数据处理的装置

Also Published As

Publication number Publication date
CN105809096A (zh) 2016-07-27

Similar Documents

Publication Publication Date Title
WO2016106966A1 (zh) 人物标注方法和终端、存储介质
Song et al. Predicting occupation via human clothing and contexts
Balntas et al. Pose guided RGBD feature learning for 3D object pose estimation
AU2014321165B2 (en) Image searching method and apparatus
CN105631039B (zh) 一种图片浏览方法
Johnson et al. Learning effective human pose estimation from inaccurate annotation
Koestinger et al. Annotated facial landmarks in the wild: A large-scale, real-world database for facial landmark localization
Cheng et al. Global contrast based salient region detection
US9317781B2 (en) Multiple cluster instance learning for image classification
CN108520226B (zh) 一种基于躯体分解和显著性检测的行人重识别方法
WO2017016240A1 (zh) 一种钞票冠字号识别方法
Varol et al. Toward retail product recognition on grocery shelves
WO2021012494A1 (zh) 基于深度学习的人脸识别方法、装置及计算机可读存储介质
US20140314313A1 (en) Visual clothing retrieval
US20100226584A1 (en) Method of Grouping Images by Face
WO2019061658A1 (zh) 眼镜定位方法、装置及存储介质
WO2016190814A1 (en) Method and system for facial recognition
CN110222582B (zh) 一种图像处理方法和相机
CN111178195A (zh) 人脸表情识别方法、装置及计算机可读存储介质
Tsitsoulis et al. A methodology for extracting standing human bodies from single images
CN111666976A (zh) 基于属性信息的特征融合方法、装置和存储介质
Kumar et al. On-the-fly hand detection training with application in egocentric action recognition
Manivannan et al. Hep-2 specimen classification using multi-resolution local patterns and SVM
Xue et al. Composite sketch recognition using multi-scale HOG features and semantic attributes
CN113780116A (zh) 发票分类方法、装置、计算机设备和存储介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 15874659

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 15874659

Country of ref document: EP

Kind code of ref document: A1