WO2021143237A1 - 一种动态人脸聚类方法、装置、设备和存储介质 - Google Patents
一种动态人脸聚类方法、装置、设备和存储介质 Download PDFInfo
- Publication number
- WO2021143237A1 WO2021143237A1 PCT/CN2020/121660 CN2020121660W WO2021143237A1 WO 2021143237 A1 WO2021143237 A1 WO 2021143237A1 CN 2020121660 W CN2020121660 W CN 2020121660W WO 2021143237 A1 WO2021143237 A1 WO 2021143237A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- face
- cluster
- file
- pictures
- picture
- Prior art date
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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/55—Clustering; Classification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
Definitions
- the embodiments of the present application relate to the technical field of face recognition, and in particular, to a dynamic face clustering method, device, device, and storage medium.
- Face clustering refers to grouping faces according to their identities. Generally, face clustering is done by comparing all the faces in the set in pairs, and then according to the similarity value obtained by the comparison, they will belong to the same identity. People are divided into a group to achieve clustering.
- the embodiments of the present application provide a dynamic face clustering method, device, equipment, and storage medium, which improve the efficiency of face clustering.
- an embodiment of the present application provides a dynamic face clustering method, including:
- the face pictures in each cluster are filtered, and multiple face pictures are selected from each cluster to create files respectively;
- the file corresponding to the newly added face picture is determined, and the newly added face picture is added to the cluster corresponding to the file.
- an embodiment of the present application provides a dynamic face clustering device, including a static clustering module, a face screening module, a file merging module, and a picture archiving module, wherein:
- the static clustering module is used to perform static clustering processing on batches of face images to obtain multiple clusters including neighboring face sets;
- the face filtering module is used to filter the face pictures in each cluster according to the average similarity of each face picture in each cluster relative to other face pictures in the cluster, and select multiple face pictures from each cluster Create files separately for face pictures;
- the file merging module is used to merge the clusters that meet the neighbor merging conditions according to the neighbor similarity of the files, and re-establish the file based on the merged clusters;
- the picture archiving module is used to determine the file corresponding to the newly added face picture according to the average similarity between the newly added face picture and the face picture in each file, and add the newly added face picture to the The file corresponds to the cluster.
- an embodiment of the present application provides a computer device, including: a memory and one or more processors;
- the memory is used to store one or more programs
- the one or more processors When the one or more programs are executed by the one or more processors, the one or more processors implement the dynamic face clustering method as described in the first aspect.
- embodiments of the present application provide a storage medium containing computer-executable instructions, which are used to execute the dynamic face clustering described in the first aspect when the computer-executable instructions are executed by a computer processor. method.
- a batch of face images is statically clustered to obtain multiple clusters including a collection of neighbors' faces, according to the average similarity of each face image in each cluster with respect to other face images in the cluster ,
- To filter the face pictures in each cluster select multiple face pictures from each cluster to create files separately, merge the clusters that meet the neighbor merging conditions according to the neighbor similarity of the files, and based on the merged clusters Re-create the file, determine the file corresponding to the newly-added face picture according to the average similarity between the newly-added face picture and the face picture in each file, and add the newly-added face picture to the file correspondence
- matching through real-time input pictures is more in line with the scene where the camera continuously collects new pictures in the application. When matching, it is only compared with some pictures in the file, and the comparison is reduced. The number of pictures facilitates the expansion of the data set size.
- FIG. 1 is a flowchart of a dynamic face clustering method provided by an embodiment of the present application
- FIG. 2 is a flowchart of another dynamic face clustering method provided by an embodiment of the present application.
- FIG. 3 is a schematic structural diagram of a residual network provided by an embodiment of the present application.
- FIG. 4 is a diagram of the internal structure of a residual network provided by an embodiment of the present application.
- Fig. 5 is a structural diagram of a face feature extractor provided by an embodiment of the present application.
- Fig. 6 is a flowchart of another dynamic face clustering method provided by an embodiment of the present application.
- FIG. 7 is a flowchart of another dynamic face clustering method provided by an embodiment of the present application.
- FIG. 8 is a structural block diagram of a dynamic face clustering device provided by an embodiment of the present application.
- Fig. 9 is a schematic structural diagram of a computer device provided by an embodiment of the present application.
- Fig. 1 shows a flow chart of a dynamic face clustering method provided by an embodiment of this application.
- the method can be applied to face clustering.
- the dynamic face clustering method provided by the embodiment of this application can be composed of dynamic face clustering.
- the clustering device is implemented, and the dynamic face clustering device can be implemented by hardware and/or software, and integrated in a computer device.
- the dynamic face clustering method includes:
- S101 Perform static clustering processing on batches of face pictures to obtain multiple clusters including neighbor face sets.
- the batch of face pictures may be pictures taken through a camera, pictures stored inside a computer device or pictures downloaded from the network.
- the neighbor’s face should be understood as: After calculating the distance between face picture A and other faces, it will determine multiple face pictures ⁇ B1, B2, B3... ⁇ that are relatively small from itself, and compare people according to the distance between faces. Face pictures ⁇ B1, B2, B3... ⁇ are sorted, and a threshold is set. Face pictures with a distance less than the threshold are identified as neighbor faces of face picture A.
- a neural network is first established to learn and train a face data set, and a trained neural network is obtained.
- the neural network can be used to obtain a face feature vector from an input face picture.
- the neural network obtains the face feature vectors of batches of face pictures, determines the vector distance between the face pictures based on the face feature vectors, and determines the neighbor face set of each photo according to the vector distance, and divides each face picture
- the neighbor face sets of are determined as a cluster respectively.
- the face data set can be a public data set commonly used in the field of face recognition, such as the LFW data set.
- the face data set is established for studying the problem of face recognition in an unrestricted environment and contains more than 13,000 face images. All are collected on the Internet, and each face is marked with a person's name. Among them, about 1,680 people contain more than two faces. Others such as IJB-B, CASIA-Webface, and VGG-Face can also be used for static clustering processing, and this solution is not limited.
- S102 According to the average similarity of each face picture in each cluster relative to other face pictures in the cluster, filter the face pictures in each cluster, and select multiple face pictures from each cluster to establish respectively file.
- the similarity of each face picture in each cluster relative to other face pictures in the same cluster is calculated, and the similarity of each face picture relative to other face pictures is summed, and then calculated The average similarity of each face picture.
- the face pictures whose average similarity does not meet the requirements are deleted from the corresponding clusters, reducing the situation that there are multiple different faces in one file in the subsequently created file.
- the judgment of whether the average similarity meets the requirements can be made by comparing the average similarity with a preset lower threshold of similarity.
- multiple face pictures are randomly selected from each cluster, and files are created based on the selected face pictures in each cluster.
- the number of face pictures selected from each cluster can be set according to actual needs, such as 3, 5, or 10 faces. This solution is described by taking three face pictures randomly selected from each cluster as an example.
- Each file created contains three face images whose average similarity meets the requirements.
- the neighbor face pictures of the multiple face pictures in each file are determined.
- the neighbors of the three face pictures in each file are determined.
- the face picture is described as an example.
- the neighbor similarity of each file relative to other files is calculated, and the neighbor similarity meets the neighbor merging condition (for example, the neighbor similarity reaches a preset
- the clusters of the two files corresponding to the merge threshold are merged into one cluster, delete these two files, and select multiple (take three as an example) face pictures from the merged clusters to recreate the files to reduce the appearance A situation where a face exists in multiple files at the same time.
- S104 Determine the file corresponding to the newly added face picture according to the average similarity between the newly added face picture and the face pictures in each file, and add the newly added face picture to the cluster corresponding to the file middle.
- the newly-added face picture may be a picture taken by a camera, a picture stored inside a computer device or a picture downloaded from the network.
- the average similarity between the newly-added face picture and multiple face pictures in each file (three pictures are taken as an example) in each file is calculated separately . Compare the average similarity of the new face image corresponding to each file, determine the file corresponding to the maximum average similarity, and add the new face image to the cluster corresponding to the file to complete the new person Dynamic cluster matching of face images.
- static clustering is used to process a batch of face images first, build files based on the clusters obtained by static clustering, and then combine a single new face image with the created one.
- the multiple face pictures in the file are matched, so that the newly added face picture is added to a cluster corresponding to the matched file, and the dynamic cluster matching of the newly added face picture is completed.
- the application scenario of the new picture is only compared with multiple face pictures in the file, which greatly reduces the number of face pictures to be compared, and makes the data scale under dynamic clustering easier to expand.
- FIG. 2 is a flowchart of another dynamic face clustering method provided by an embodiment of the application.
- the dynamic face clustering method is a specific embodiment of the above-mentioned dynamic face clustering method. As shown in Figure 2, the dynamic face clustering method includes:
- the face data set used for training may be a public data set commonly used in the field of face recognition, such as the LFW data set.
- the face data set is established to study the problem of face recognition in an unrestricted environment, including more than 13,000 face images were collected on the Internet, and each face was tagged with a name. Among them, about 1,680 people contained more than two faces.
- Others such as IJB-B, CASIA-Webface, and VGG-Face can also be used to train the residual network, and this scheme is not limited.
- a specific residual network is first constructed.
- the residual network is shown in Figure 3.
- Figure 3 is a schematic structural diagram of a residual network provided by an embodiment of this application, using a public face data set
- the specific residual network is learned and trained to obtain a trained residual network, and the trained residual network can be used to perform face classification tasks.
- the specific residual network consists of input (input), N ResNet blocks, a fully connected layer, and softmax (normalization layer).
- the internal structure of the ResNet block is shown in Figure 4.
- 4 is a diagram of the internal structure of a residual network provided by an embodiment of this application.
- conv(1*1) represents the use of the convolutional layer and the size of the convolution kernel is 1*1
- the BN layer is used for batch normalization
- Relu is the commonly used neural network activation function
- the symbol "+" represents the execution vector The addition operation.
- the fully connected layer uses 1024 neural network nodes.
- S202 Process the residual network to obtain a face feature extractor, and input a face picture to be classified into the face feature extractor to obtain a face feature vector corresponding to each face picture.
- the normalization layer of the residual network is removed to obtain a face feature extractor, as shown in FIG. 5, which is a structural diagram of a face feature extractor provided by an embodiment of the application.
- input corresponds to the input face picture
- the fully connected layer has 1024 nodes, that is, a vector of 1024 values is output for each input picture as the face feature vector corresponding to the face picture.
- S203 Calculate the vector distance between each face feature vector and other face feature vectors, and determine the neighbor face set of each face picture according to the vector distance.
- a and b represent two different face pictures, and a i and b i are the face feature vectors corresponding to each picture.
- the above formula not only considers the directional similarity of the face feature vectors, but also considers the person The difference between the facial feature vector values makes the vector distance measurement result more reasonable. It should be noted that this solution can also use other existing vector distance calculation formulas, but the calculation effect is not as good as the above formulas.
- the process of determining the neighbor face set of each face picture according to the vector distance may be: according to the formula
- the vector distance is normalized, and the face pictures that are less than the first preset threshold in the processing result are determined as the neighbor face set.
- the first preset threshold includes 0.25 (the first preset threshold can be performed according to actual calculation needs. Adjustment), where N represents the number of samples, which is a positive integer greater than 1.
- S204 Determine the neighbor face set of each face picture as a cluster respectively, and merge the clusters whose similarity between the clusters meets the merging condition between the clusters.
- the inter-cluster merging condition may be that the similarity between the clusters is greater than the second preset threshold. For example, according to the formula
- the second preset threshold includes 0.7, where A and B represent two different clusters.
- represents the number of elements in set A
- represents the number of elements in set B
- represents the number of elements in the intersection of set A and set B.
- the cluster is initialized, that is, the neighbor face set of each face picture is determined to be a cluster.
- these can be separated individually.
- the clusters form a cluster list.
- the specific merging process can be: take a cluster from the cluster list, calculate the similarity between the cluster and other clusters in the cluster list, merge if the merging conditions are met, and calculate the merged cluster and the cluster list.
- the similarity between clusters in other clusters is deduced by analogy until all clusters in the cluster list have been traversed. Take out the second cluster in the cluster list.
- the cluster has been merged, take out the next cluster in the cluster list until the unmerged cluster is taken out, and then calculate the similarity between clusters and other clusters in the cluster list. It determines whether the merging condition is satisfied, and if it is satisfied, the merging is performed. The merging steps are repeated until the number of clusters in a round of iteration is reduced by less than 5% when it is not iterated, and the clustering is determined to be completed.
- S205 According to the average similarity of each face picture in each cluster relative to other face pictures in the cluster, filter the face pictures in each cluster, and select multiple face pictures from each cluster to establish respectively file.
- S207 Determine the file corresponding to the newly added face picture according to the average similarity between the newly added face picture and the face pictures in each file, and add the newly added face picture to the cluster corresponding to the file middle.
- the face features extracted by the residual network are driven by data, without human prior experience, and the residual network can easily find the characteristics of the data, while the artificial Defining characteristics cannot be done.
- the artificially defined features are limited, and the more and more refined the defined features, the more energy is consumed.
- For the residual network it is only necessary to increase the number of nodes to efficiently obtain more features.
- the advantage of the clustering method in this scheme is that the amount of calculation is small, the convergence speed is fast in the iterative process, and the result accuracy is high.
- the initialization in this scheme is based on each sample as the center, and the neighbor faces are selected.
- this method initializes N (number of samples) centers, and the subsequent process will gradually reduce the number of clusters. The reason is that in the initial process, the number of people in the face set cannot be determined, and no prior experience is introduced.
- an element can appear in multiple clusters at the beginning, find N overlapping regions of clusters, and decide whether they can be combined according to the overlapping regions.
- a single new face image is matched with multiple face images in the created file, so that the new face image is added to a cluster corresponding to the matched file.
- the dynamic cluster matching of the newly added face image is completed.
- FIG. 6 is a flowchart of another dynamic face clustering method provided by an embodiment of the application.
- the dynamic face clustering method is a specific embodiment of the above-mentioned dynamic face clustering method. As shown in Figure 6, the dynamic face clustering method includes:
- S301 Perform static clustering processing on batches of face pictures to obtain multiple clusters including neighbor face sets.
- S302 Calculate the average similarity of each face picture in each cluster with respect to other face pictures in the cluster.
- the vector distance between the face feature vector of each face picture in each cluster and the face feature vectors of other face pictures in the cluster is calculated according to the following formula:
- a and b respectively represent two different face pictures in the same cluster, and a i and b i are the face feature vectors corresponding to each picture.
- the above formula not only considers the directional similarity of the face feature vectors, The difference between the value of the facial feature vector is also considered to make the vector distance measurement result more reasonable. It should be noted that this solution can also use other existing vector distance calculation formulas, but the calculation effect is not as good as the above formulas.
- N represents the number of samples, which is a positive integer greater than 1.
- the value obtained by subtracting the normalized vector distance from 1 is used as the similarity of the face image to other face images in the cluster, and the face image is calculated for the cluster
- the mean value of the similarity of all other face pictures in the cluster is obtained, and the average similarity of the face picture in the cluster relative to other face pictures in the cluster is obtained.
- Traverse all face pictures in each cluster to obtain the average similarity of each face picture in each cluster relative to other face pictures in the cluster.
- S303 Sort the average similarity in each cluster, and determine the average similarity corresponding to the upper quartile and the lower quartile.
- the average similarity in each cluster is sorted according to the order of the average similarity from small to large , And find the average similarity S3 and S1 corresponding to the upper quartile Q3 and the lower quartile Q1.
- the quartile is also called the quartile point, which refers to the value in statistics that arranges all values from small to large and divides them into four equal parts. It is the 25% and 75% values of a group of data after sorting.
- the quartile is to divide all the data into four parts by 3 points, each part contains 25% of the data.
- the middle quartile is the median, and the quartile usually refers to The value at the 25% position (called the lower quartile) and the value at the 75% position (called the upper quartile).
- the average similarity S3 corresponding to the upper quartile Q3 in the same cluster is calculated.
- S306 Use the maximum value of the lower edge and the preset lower limit in each cluster as the filtering threshold, and delete from each cluster the face pictures whose average similarity is less than the corresponding filtering threshold.
- a face picture whose average similarity with respect to other face pictures in the cluster is less than the deadline value is considered to be a misclassified face, and this type of face picture can be directly deleted from the cluster.
- For each cluster after determining the screening threshold, compare the corresponding average similarity of each face photo in the cluster with the corresponding screening threshold of the cluster, and compare the face photos with the average similarity less than the screening threshold from this Delete from the cluster to filter each cluster.
- S307 Select multiple face pictures from each cluster after filtering to create files respectively.
- the multiple face pictures include three face pictures (which can be adjusted according to actual calculation needs). After completing the screening of face pictures, three face pictures are randomly selected from each cluster, and files are created based on the selected face pictures in each cluster.
- S308 Determine the neighbor faces of the multiple face pictures in each file, and calculate the neighbor similarity of each file.
- the neighbor faces of the three face pictures in each file are determined.
- the determination of the neighbor's face can be determined according to step S203, which will not be repeated in this embodiment.
- all neighbor faces in the cluster are formed into a neighbor face set, and the neighbor similarity between each two files is further calculated based on the neighbor face set.
- the neighbor similarity of each file is calculated according to the following formula:
- a and B are the neighbor face sets of the two files respectively
- a ⁇ B is the intersection of the two neighbor face sets of A and B
- Count() is used to return the number of set elements
- min is a function used to take the smallest of two numbers.
- the merging threshold includes 0.5 (which can be adjusted according to actual calculation needs).
- the neighbor similarity between the two files is compared with the merge threshold. When the neighbor similarity between any two files is greater than the merge threshold, the clusters corresponding to the two files are merged into one cluster, and the two files are deleted. Files.
- S310 Select multiple face pictures based on the merged clusters to recreate a file.
- multiple face images are selected from the merged clusters (in this embodiment, three face images are selected as an example) to recreate a file, and the newly created file is added to step S308 Continue to determine whether it needs to be merged with other files.
- S311 Calculate the average similarity between the newly added face picture and the multiple face pictures in each file.
- the average similarity can be similar according to the calculation of the average similarity in step S302, which will not be repeated here.
- S312 Determine the file corresponding to the newly added face picture that has the highest average similarity and reaches the archiving threshold, and add the newly added face picture to the cluster corresponding to the file.
- the archiving threshold includes 0.70 (which can be adjusted according to actual calculation needs). After determining the average similarity of the newly added face image with respect to all files, compare the average similarity of the newly added face image to each file, determine the file corresponding to the maximum average similarity, and compare the maximum similarity The degree is compared with the archiving threshold, and when the maximum similarity is greater than or equal to the archiving threshold, the newly added face image is added to the cluster corresponding to the file to complete the dynamic cluster matching of the newly added face image.
- static clustering is used to process a batch of face images first, build files based on the clusters obtained by static clustering, and then combine a single new face image with the created one.
- the three face pictures in the file are matched, so that the newly added face picture is added to the cluster of the matched file pairs, and the dynamic cluster matching of the newly added face picture is completed.
- FIG. 7 is a flow chart of another dynamic face clustering method provided by an embodiment of the application.
- the dynamic face clustering method is a specific embodiment of the above-mentioned dynamic face clustering method. As shown in Figure 7, the dynamic face clustering method includes:
- S401 Perform static clustering processing on batches of face pictures to obtain multiple clusters including neighbor face sets.
- S402 Determine the number of face pictures in each cluster, and add face pictures of the clusters whose number of pictures does not meet the archiving requirements into the remaining picture set.
- the number of face pictures in each cluster is determined, and it is determined whether the number of pictures in each cluster is greater than or equal to the number required for file creation (in this embodiment, the number of images is greater than or equal to 3 as the file creation requirement ), if the number of pictures meets the file-building requirement, the corresponding cluster is used for the file-building operation in step S403, and if the number of pictures does not meet the file-building requirement, the face pictures in the corresponding cluster are added to the remaining picture set.
- S403 According to the average similarity of each face picture in each cluster relative to other face pictures in the cluster, filter the face pictures in each cluster, and select multiple face pictures from each cluster to establish respectively file.
- step S405 Determine whether there is a file corresponding to the newly added face picture. If yes, skip to step S406, otherwise, skip to step S407.
- the average similarity can be similar according to the calculation of the average similarity in step S302, which will not be repeated here.
- the average similarity of the newly added face image After determining the average similarity of the newly added face image with respect to all files, compare the average similarity of the newly added face image to each file, determine the file corresponding to the maximum average similarity, and compare the maximum similarity The degree of similarity is compared with the archiving threshold (taking 0.7 as an example, which can be adjusted according to the actual situation).
- the maximum similarity is greater than or equal to the archiving threshold, the newly-added face picture is considered to correspond to the file, and it jumps to layout S406 . If the maximum similarity is less than the archiving threshold, it is considered that there is no file corresponding to the newly added face picture, and step S407 is jumped to.
- the newly added face image is added to the cluster corresponding to the file to complete the dynamic cluster matching of the newly added face image.
- the newly added face picture is added to the remaining picture set.
- S408 Perform clustering processing on the remaining picture sets in which the number of face pictures reaches the number threshold, to obtain multiple clusters including neighbor face sets.
- the number of face pictures in the remaining picture set is monitored, and when the number of pictures reaches the number threshold (which can be determined according to the actual situation), the remaining picture set is clustered to obtain a plurality of neighbors including neighbors. Cluster of human faces collection.
- the clustering process for the remaining picture sets is similar to steps S201-S204, and will not be repeated here.
- step S409 Confirm whether the number of face pictures in each cluster meets the filing requirements. If the file creation requirements are met, skip to step S410, otherwise, skip to step S411.
- step S402 after clustering the remaining picture sets to obtain multiple clusters, refer to step S402 to determine whether the number of face pictures in these clusters meets the filing requirements.
- S410 Create files based on clusters whose number of pictures meets the file building requirements.
- the file is created based on the clusters whose number of pictures meets the archiving requirement.
- the face pictures in the corresponding cluster are added to the remaining picture set again, and the next time the number of face pictures in the remaining picture set reaches the number threshold, then participate again Static clustering processing.
- static clustering is used to process a batch of face images first, build files based on the clusters obtained by static clustering, and then combine a single new face image with the created one.
- the three face pictures in the file are matched, so that the newly added face picture is added to a cluster corresponding to the matched file, and the dynamic cluster matching of the newly added face picture is completed.
- dynamic clustering when dynamic clustering is matched, it is only compared with the three face images in the file, which greatly reduces the number of face images to be compared, making the data scale under dynamic clustering easier to expand.
- Put the remaining face pictures into the remaining picture set cluster the remaining picture sets when the number of pictures reaches the number threshold, and build files based on the processing results to improve the accuracy of dynamic processing.
- FIG. 8 is a structural block diagram of a dynamic face clustering device provided by an embodiment of the present application.
- the dynamic face clustering device is used to execute the dynamic face clustering method provided in the above embodiment, and has functional modules corresponding to the execution method. And beneficial effects.
- the device specifically includes a static clustering module 81, a face screening module 82, a file merging module 83, and a picture archiving module 84, among which:
- the static clustering module 81 is used to perform static clustering processing on batches of face pictures to obtain multiple clusters including neighbor face sets;
- the face filtering module 82 is used to filter the face pictures in each cluster according to the average similarity of each face picture in each cluster relative to other face pictures in the cluster, and select multiple face pictures from each cluster. Create files for each face picture;
- the file merging module 83 is used for merging clusters that meet the neighbor merging conditions according to the neighbor similarity of the files, and re-establishing the files based on the merged clusters;
- the picture archiving module 84 is configured to determine the file corresponding to the newly added face picture according to the average similarity between the newly added face picture and the face picture in each file, and add the newly added face picture to all the files. In the cluster corresponding to the file.
- static clustering is used to process a batch of face images first, build files based on the clusters obtained by static clustering, and then combine a single new face image with the created one.
- the multiple face pictures in the file are matched, so that the newly added face picture is added to the cluster of the matched file pair, and the dynamic cluster matching of the newly added face picture is completed.
- the application scenario of the new picture is only compared with multiple face pictures in the file, which greatly reduces the number of face pictures to be compared, and makes the data scale under dynamic clustering easier to expand.
- the static clustering module 81 is specifically configured to:
- the neighbor face set of each face picture is determined as a cluster, and the clusters whose similarity between the clusters meets the merging condition between the clusters are merged.
- the face screening module 82 is specifically configured to:
- the file merging module 83 is specifically configured to:
- the picture archiving module 84 is specifically configured to:
- FIG. 9 is a schematic structural diagram of a computer device provided by an embodiment of the present application.
- the computer device includes an input device 93, an output device 94, a memory 92, and one or more processors 91; the memory 92 , Used to store one or more programs; when the one or more programs are executed by the one or more processors 91, the one or more processors 91 implement the dynamic face provided in the above-mentioned embodiment Clustering method.
- the input device 93, the output device 94, the memory 92, and the processor 91 can be connected by a bus or in other ways. In Fig. 9, the connection by a bus is taken as an example.
- the memory 92 as a storage medium readable by a computing device, can be used to store software programs, computer-executable programs, and modules, such as program instructions/modules corresponding to the dynamic face clustering method described in any embodiment of this application (for example, dynamic The static clustering module 81, the face screening module 82, the file merging module 83 and the picture archiving module 84 in the face clustering device).
- the memory 92 may mainly include a program storage area and a data storage area.
- the program storage area may store an operating system and an application program required by at least one function; the data storage area may store data created according to the use of the device, and the like.
- the memory 92 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one magnetic disk storage device, a flash memory device, or other non-volatile solid-state storage devices.
- the memory 92 may further include a memory remotely provided with respect to the processor 91, and these remote memories may be connected to the device through a network. Examples of the aforementioned networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.
- the input device 93 can be used to receive inputted numeric or character information, and generate key signal input related to user settings and function control of the device.
- the output device 94 may include a display device such as a display screen.
- the processor 91 executes various functional applications and data processing of the device by running the software programs, instructions, and modules stored in the memory 92, that is, realizes the above-mentioned dynamic face clustering method.
- the dynamic face clustering apparatus and computer equipment provided above can be used to execute the dynamic face clustering method provided in the above embodiments, and have corresponding functions and beneficial effects.
- the embodiment of the present application also provides a storage medium containing computer-executable instructions, when the computer-executable instructions are executed by a computer processor, they are used to execute the dynamic face clustering method provided in the above-mentioned embodiments.
- Clustering methods include:
- the face pictures in each cluster are filtered, and multiple face pictures are selected from each cluster to create files respectively;
- the file corresponding to the newly added face picture is determined, and the newly added face picture is added to the cluster corresponding to the file.
- Storage medium any of various types of storage devices or storage devices.
- the term "storage medium” is intended to include: installation media such as CD-ROM, floppy disk or tape device; computer system memory or random access memory such as DRAM, DDR RAM, SRAM, EDO RAM, Rambus RAM, etc. ; Non-volatile memory, such as flash memory, magnetic media (such as hard disk or optical storage); registers or other similar types of memory elements.
- the storage medium may further include other types of memory or a combination thereof.
- the storage medium may be located in the first computer system in which the program is executed, or may be located in a different second computer system connected to the first computer system through a network (such as the Internet).
- the second computer system can provide the program instructions to the first computer for execution.
- storage media may include two or more storage media that may reside in different locations (for example, in different computer systems connected through a network).
- the storage medium may store program instructions (for example, embodied as a computer program) executable by one or more processors.
- the storage medium containing computer-executable instructions provided by the embodiments of the present application is not limited to the above-mentioned dynamic face clustering method, and can also execute the dynamic face clustering methods provided in any embodiment of the present application. Related operations in the face clustering method.
- the dynamic face clustering apparatus, device and storage medium provided in the above embodiments can implement the dynamic face clustering method provided in any embodiment of this application.
- the dynamic face clustering method provided by any embodiment can implement the dynamic face clustering method provided in any embodiment of this application.
- the dynamic face clustering method provided by any embodiment.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Health & Medical Sciences (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Oral & Maxillofacial Surgery (AREA)
- Bioinformatics & Computational Biology (AREA)
- Probability & Statistics with Applications (AREA)
- General Health & Medical Sciences (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Databases & Information Systems (AREA)
- Image Analysis (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
Description
Claims (11)
- 一种动态人脸聚类方法,其特征在于,包括:对批量人脸图片进行静态聚类处理,得到多个包括邻居人脸集合的簇;根据每个簇中每张人脸图片相对于簇中其他人脸图片的平均相似度,对每个簇中的人脸图片进行筛选,从每个簇中选取多张人脸图片分别建立档案;根据档案的邻居相似度对满足邻居合并条件的簇进行合并,并基于合并后的簇重新建立档案;根据新增人脸图片与每个档案中的人脸图片的平均相似度,确定所述新增人脸图片所对应的档案,将所述新增人脸图片加入所述档案对应的簇中。
- 根据权利要求1所述的动态人脸聚类方法,其特征在于,所述对批量人脸图片进行静态聚类处理,得到多个包括邻居人脸集合的簇,包括:通过人脸数据集进行训练得到训练后的残差网络;对所述残差网络进行处理得到人脸特征提取器,将待分类的人脸图片输入所述人脸特征提取器得到每张人脸图片对应的人脸特征向量;计算每个人脸特征向量和其它人脸特征向量的向量距离,依据所述向量距离确定每张人脸图片的邻居人脸集合;将每张人脸图片的邻居人脸集合分别确定为一个簇,将簇间相似度满足簇间合并条件的簇进行合并。
- 根据权利要求1所述的动态人脸聚类方法,其特征在于,所述根据每个簇中每张人脸图片相对于簇中其他人脸图片的平均相似度,对每个簇中的人脸图片进行筛选,从每个簇中选取多张人脸图片分别建立档案,包括:计算每个簇中每张人脸图片相对于簇中其他人脸图片的平均相似度;将每个簇中的平均相似度进行排序,确定上四分位数和下四分位数对应的平均相似度;根据每个簇的上四分位数和下四分位数对应的平均相似度的差,得到每个簇的相似度公差;根据每个簇中下四分位数对应的平均相似度与两倍相似度公差的差,得到每个簇的下边缘;将每个簇中下边缘和预设下限中的最大值作为筛选阈值,从每个簇中删除平均相似度小于相应筛选阈值的人脸图片;从筛选后的每个簇中选取多张人脸图片分别建立档案。
- 根据权利要求1所述的动态人脸聚类方法,其特征在于,所述根据档案的 邻居相似度对满足邻居合并条件的簇进行合并,并基于合并后的簇重新建立档案,包括:确定每个档案中多张人脸图片的邻居人脸,计算每个档案的邻居相似度;将邻居相似度达到合并阈值的档案对应的簇进行合并;基于合并后的簇选取多张人脸图片重新建立档案。
- 根据权利要求1所述的动态人脸聚类方法,其特征在于,所述根据新增人脸图片与每个档案中的人脸图片的平均相似度,确定所述新增人脸图片所对应的档案,将所述新增人脸图片加入所述档案对应的簇中,包括:计算新增人脸图片与每个档案中多张人脸图片的平均相似度;确定平均相似度最高且达到归档阈值的所述新增人脸图片所对应的档案;将所述新增人脸图片加入所述档案对应的簇中。
- 根据权利要求1-6任一项所述的动态人脸聚类方法,其特征在于,所述根据每个簇中每张人脸图片相对于簇中其他人脸图片的平均相似度,对每个簇中的人脸图片进行筛选,从每个簇中选取多张人脸图片分别建立档案之前,还包括:确定每个簇中人脸图片的图片数量,将图片数量不满足建档要求的簇的人脸图片加入剩余图片集;所述根据每个簇中每张人脸图片相对于簇中其他人脸图片的平均相似度,对每个簇中的人脸图片进行筛选,从每个簇中选取多张人脸图片分别建立档案之后,还包括:将人脸图片的数量达到数量阈值的剩余图片集进行聚类处理,得到多个包括邻居人脸集合的簇;确认每个簇中人脸图片的图片数量是否满足建档要求;若满足建档要求,则基于图片数量满足建档要求的簇建立档案;若不满足建档要求,则将图片数量不满足建档要求的簇的人脸图片加入剩余图片集。
- 根据权利要求7所述的动态人脸聚类方法,其特征在于,所述根据新增人脸图片与每个档案中的人脸图片的平均相似度,确定所述新增人脸图片所对应的档案,将所述新增人脸图片加入所述档案对应的簇中,包括:根据新增人脸图片与每个档案中的人脸图片的平均相似度,确定是否存在与所述新增人脸图片对应的档案;若存在对应的档案,则将所述新增人脸图片加入所述档案对应的簇中;若未存在对应的档案,则将所述新增人脸图片加入剩余图片集。
- 一种动态人脸聚类装置,其特征在于,包括静态聚类模块、人脸筛选模块、档案合并模块和图片归档模块,其中:静态聚类模块,用于对批量人脸图片进行静态聚类处理,得到多个包括邻居人脸集合的簇;人脸筛选模块,用于根据每个簇中每张人脸图片相对于簇中其他人脸图片的平均相似度,对每个簇中的人脸图片进行筛选,从每个簇中选取多张人脸图片分别建立档案;档案合并模块,用于根据档案的邻居相似度对满足邻居合并条件的簇进行合并,并基于合并后的簇重新建立档案;图片归档模块,用于根据新增人脸图片与每个档案中的人脸图片的平均相似度,确定所述新增人脸图片所对应的档案,将所述新增人脸图片加入所述档案对应的簇中。
- 一种计算机设备,其特征在于,包括:存储器以及一个或多个处理器;所述存储器,用于存储一个或多个程序;当所述一个或多个程序被所述一个或多个处理器执行,使得所述一个或多个处理器实现如权利要求1-8任一所述的动态人脸聚类方法。
- 一种包含计算机可执行指令的存储介质,其特征在于,所述计算机可执行指令在由计算机处理器执行时用于执行如权利要求1-8任一所述的动态人脸聚类方法。
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010040705.1A CN111242040B (zh) | 2020-01-15 | 2020-01-15 | 一种动态人脸聚类方法、装置、设备和存储介质 |
CN202010040705.1 | 2020-01-15 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2021143237A1 true WO2021143237A1 (zh) | 2021-07-22 |
Family
ID=70868504
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/CN2020/121660 WO2021143237A1 (zh) | 2020-01-15 | 2020-10-16 | 一种动态人脸聚类方法、装置、设备和存储介质 |
Country Status (2)
Country | Link |
---|---|
CN (1) | CN111242040B (zh) |
WO (1) | WO2021143237A1 (zh) |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113569704A (zh) * | 2021-07-23 | 2021-10-29 | 上海明略人工智能(集团)有限公司 | 分割点判断方法、系统、存储介质及电子设备 |
CN114201658A (zh) * | 2022-02-16 | 2022-03-18 | 广东图友软件科技有限公司 | 一种基于人脸识别的档案快速检索方法 |
CN114639143A (zh) * | 2022-03-07 | 2022-06-17 | 北京百度网讯科技有限公司 | 基于人工智能的人像归档方法、设备及存储介质 |
WO2023124134A1 (zh) * | 2021-12-28 | 2023-07-06 | 上海商汤智能科技有限公司 | 档案处理方法、装置、电子设备、计算机存储介质和程序 |
CN117367023A (zh) * | 2023-10-25 | 2024-01-09 | 广东鑫焱智能设备科技有限公司 | 一种冷藏柜能耗控制方法、系统、设备及存储介质 |
Families Citing this family (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111242040B (zh) * | 2020-01-15 | 2022-08-02 | 佳都科技集团股份有限公司 | 一种动态人脸聚类方法、装置、设备和存储介质 |
CN112001414A (zh) * | 2020-07-14 | 2020-11-27 | 浙江大华技术股份有限公司 | 聚类方法、设备及计算机存储介质 |
CN112257801B (zh) * | 2020-10-30 | 2022-04-29 | 浙江商汤科技开发有限公司 | 图像的增量聚类方法、装置、电子设备及存储介质 |
CN112270290B (zh) * | 2020-11-10 | 2022-08-02 | 佳都科技集团股份有限公司 | 人脸图像动态聚类方法、装置、电子设备和存储介质 |
CN112948613B (zh) * | 2020-12-01 | 2023-09-15 | 上海云从企业发展有限公司 | 图像增量聚类方法、系统、介质及装置 |
CN112417197B (zh) * | 2020-12-02 | 2022-02-25 | 云从科技集团股份有限公司 | 一种排序方法、装置、机器可读介质及设备 |
CN112446362B (zh) * | 2020-12-16 | 2022-07-22 | 上海芯翌智能科技有限公司 | 人脸图片档案处理方法及设备 |
CN112668482B (zh) * | 2020-12-29 | 2023-11-21 | 中国平安人寿保险股份有限公司 | 人脸识别训练方法、装置、计算机设备及存储介质 |
CN112765388A (zh) * | 2021-01-29 | 2021-05-07 | 云从科技集团股份有限公司 | 一种目标数据标注方法、系统、设备及介质 |
CN114022934B (zh) * | 2021-11-04 | 2023-06-27 | 清华大学 | 一种基于多数原则的实时人像聚档方法、系统和介质 |
CN115661495B (zh) * | 2022-09-28 | 2023-04-18 | 中国测绘科学研究院 | 一种紧凑划分及多层次合并策略的大规模SfM方法 |
Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US6092072A (en) * | 1998-04-07 | 2000-07-18 | Lucent Technologies, Inc. | Programmed medium for clustering large databases |
CN109145844A (zh) * | 2018-08-29 | 2019-01-04 | 北京旷视科技有限公司 | 用于城市安防监控的档案管理方法、装置以及电子设备 |
CN110059657A (zh) * | 2019-04-25 | 2019-07-26 | 北京旷视科技有限公司 | 档案处理方法、装置、电子设备及计算机可读存储介质 |
CN110163135A (zh) * | 2019-05-10 | 2019-08-23 | 杭州商警云智能科技有限公司 | 一种基于动态算法的一人一档人脸聚类的方法及系统 |
CN110232373A (zh) * | 2019-08-12 | 2019-09-13 | 佳都新太科技股份有限公司 | 人脸聚类方法、装置、设备和存储介质 |
CN110414429A (zh) * | 2019-07-29 | 2019-11-05 | 佳都新太科技股份有限公司 | 人脸聚类方法、装置、设备和存储介质 |
CN110555144A (zh) * | 2019-09-06 | 2019-12-10 | 北京奇艺世纪科技有限公司 | 一种流聚类的方法、装置、电子设备及介质 |
CN111242040A (zh) * | 2020-01-15 | 2020-06-05 | 佳都新太科技股份有限公司 | 一种动态人脸聚类方法、装置、设备和存储介质 |
Family Cites Families (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2012064169A2 (es) * | 2010-11-10 | 2012-05-18 | Rolando Rafael Hernandez Cisneros | Metodo para la deteccion y clasificacion de grupos de microcalcificacion en mamografias digitales |
WO2012105085A1 (ja) * | 2011-01-31 | 2012-08-09 | Necソフト株式会社 | 画像認証装置、画像認証方法、プログラムおよび記録媒体 |
US8666984B2 (en) * | 2011-03-18 | 2014-03-04 | Microsoft Corporation | Unsupervised message clustering |
CN109783685B (zh) * | 2018-12-28 | 2022-03-15 | 上海依图网络科技有限公司 | 一种查询方法及装置 |
CN109710780B (zh) * | 2018-12-28 | 2022-03-15 | 上海依图网络科技有限公司 | 一种归档方法及装置 |
CN109815370A (zh) * | 2018-12-28 | 2019-05-28 | 上海依图网络科技有限公司 | 一种归档方法及装置 |
CN110232331B (zh) * | 2019-05-23 | 2022-09-27 | 深圳大学 | 一种在线人脸聚类的方法及系统 |
CN110222177A (zh) * | 2019-05-24 | 2019-09-10 | 雷恩友力数据科技南京有限公司 | 一种基于k-均值聚类算法的初始聚类中心确定方法及装置 |
CN110413815B (zh) * | 2019-07-26 | 2022-05-24 | 北京旷视科技有限公司 | 人像聚类清洗方法和装置 |
-
2020
- 2020-01-15 CN CN202010040705.1A patent/CN111242040B/zh active Active
- 2020-10-16 WO PCT/CN2020/121660 patent/WO2021143237A1/zh active Application Filing
Patent Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US6092072A (en) * | 1998-04-07 | 2000-07-18 | Lucent Technologies, Inc. | Programmed medium for clustering large databases |
CN109145844A (zh) * | 2018-08-29 | 2019-01-04 | 北京旷视科技有限公司 | 用于城市安防监控的档案管理方法、装置以及电子设备 |
CN110059657A (zh) * | 2019-04-25 | 2019-07-26 | 北京旷视科技有限公司 | 档案处理方法、装置、电子设备及计算机可读存储介质 |
CN110163135A (zh) * | 2019-05-10 | 2019-08-23 | 杭州商警云智能科技有限公司 | 一种基于动态算法的一人一档人脸聚类的方法及系统 |
CN110414429A (zh) * | 2019-07-29 | 2019-11-05 | 佳都新太科技股份有限公司 | 人脸聚类方法、装置、设备和存储介质 |
CN110232373A (zh) * | 2019-08-12 | 2019-09-13 | 佳都新太科技股份有限公司 | 人脸聚类方法、装置、设备和存储介质 |
CN110555144A (zh) * | 2019-09-06 | 2019-12-10 | 北京奇艺世纪科技有限公司 | 一种流聚类的方法、装置、电子设备及介质 |
CN111242040A (zh) * | 2020-01-15 | 2020-06-05 | 佳都新太科技股份有限公司 | 一种动态人脸聚类方法、装置、设备和存储介质 |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113569704A (zh) * | 2021-07-23 | 2021-10-29 | 上海明略人工智能(集团)有限公司 | 分割点判断方法、系统、存储介质及电子设备 |
CN113569704B (zh) * | 2021-07-23 | 2023-12-12 | 上海明略人工智能(集团)有限公司 | 分割点判断方法、系统、存储介质及电子设备 |
WO2023124134A1 (zh) * | 2021-12-28 | 2023-07-06 | 上海商汤智能科技有限公司 | 档案处理方法、装置、电子设备、计算机存储介质和程序 |
CN114201658A (zh) * | 2022-02-16 | 2022-03-18 | 广东图友软件科技有限公司 | 一种基于人脸识别的档案快速检索方法 |
CN114201658B (zh) * | 2022-02-16 | 2022-04-26 | 广东图友软件科技有限公司 | 一种基于人脸识别的档案快速检索方法 |
CN114639143A (zh) * | 2022-03-07 | 2022-06-17 | 北京百度网讯科技有限公司 | 基于人工智能的人像归档方法、设备及存储介质 |
CN114639143B (zh) * | 2022-03-07 | 2024-04-16 | 北京百度网讯科技有限公司 | 基于人工智能的人像归档方法、设备及存储介质 |
CN117367023A (zh) * | 2023-10-25 | 2024-01-09 | 广东鑫焱智能设备科技有限公司 | 一种冷藏柜能耗控制方法、系统、设备及存储介质 |
Also Published As
Publication number | Publication date |
---|---|
CN111242040B (zh) | 2022-08-02 |
CN111242040A (zh) | 2020-06-05 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2021143237A1 (zh) | 一种动态人脸聚类方法、装置、设备和存储介质 | |
WO2021027193A1 (zh) | 人脸聚类方法、装置、设备和存储介质 | |
Zhang et al. | Learning two-view correspondences and geometry using order-aware network | |
CN107766850B (zh) | 基于结合人脸属性信息的人脸识别方法 | |
CN109241982B (zh) | 基于深浅层卷积神经网络的目标检测方法 | |
WO2018171109A1 (zh) | 基于卷积神经网络的视频动作检测方法 | |
WO2020192736A1 (zh) | 物体识别方法及装置 | |
CN107358262B (zh) | 一种高分辨率图像的分类方法及分类装置 | |
WO2021057056A1 (zh) | 神经网络架构搜索方法、图像处理方法、装置和存储介质 | |
WO2023206944A1 (zh) | 一种语义分割方法、装置、计算机设备和存储介质 | |
CN108710893B (zh) | 一种基于特征融合的数字图像相机源模型分类方法 | |
WO2021103731A1 (zh) | 一种语义分割方法、模型训练方法及装置 | |
Chen et al. | Corse-to-fine road extraction based on local Dirichlet mixture models and multiscale-high-order deep learning | |
CN111797846B (zh) | 一种基于特征金字塔网络的反馈式目标检测方法 | |
CN116312782B (zh) | 一种融合影像基因数据的空间转录组spot区域聚类方法 | |
CN111738341B (zh) | 一种分布式大规模人脸聚类方法及装置 | |
CN110929099A (zh) | 一种基于多任务学习的短视频帧语义提取方法及系统 | |
CN112861659A (zh) | 一种图像模型训练方法、装置及电子设备、存储介质 | |
CN111339862A (zh) | 一种基于通道注意力机制的遥感场景分类方法及装置 | |
CN112508181A (zh) | 一种基于多通道机制的图池化方法 | |
Gyawali et al. | Age range estimation using MTCNN and vgg-face model | |
CN111242176A (zh) | 计算机视觉任务的处理方法、装置及电子系统 | |
WO2022063321A1 (zh) | 图像处理方法、装置、设备及存储介质 | |
CN111652329B (zh) | 一种图像分类方法、装置、存储介质和电子设备 | |
WO2021057690A1 (zh) | 构建神经网络的方法与装置、及图像处理方法与装置 |
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: 20914219 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: 20914219 Country of ref document: EP Kind code of ref document: A1 |
|
32PN | Ep: public notification in the ep bulletin as address of the adressee cannot be established |
Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 31.01.2023) |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 20914219 Country of ref document: EP Kind code of ref document: A1 |