WO2020237887A1 - 人脸识别方法、装置、终端及存储介质 - Google Patents
人脸识别方法、装置、终端及存储介质 Download PDFInfo
- Publication number
- WO2020237887A1 WO2020237887A1 PCT/CN2019/103645 CN2019103645W WO2020237887A1 WO 2020237887 A1 WO2020237887 A1 WO 2020237887A1 CN 2019103645 W CN2019103645 W CN 2019103645W WO 2020237887 A1 WO2020237887 A1 WO 2020237887A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- face
- low
- dimensional
- feature vector
- integer code
- Prior art date
Links
Images
Classifications
-
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/74—Image or video pattern matching; Proximity measures in feature spaces
- G06V10/75—Organisation of the matching processes, e.g. simultaneous or sequential comparisons of image or video features; Coarse-fine approaches, e.g. multi-scale approaches; using context analysis; Selection of dictionaries
- G06V10/751—Comparing pixel values or logical combinations thereof, or feature values having positional relevance, e.g. template matching
-
- 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
- This application relates to the field of face recognition technology, and in particular to a face recognition method, device, terminal and storage medium.
- Terminal equipment collects face images in real time through camera devices, and combines them with pre-stored face template images. Match, after the match is successful, the verification is passed.
- the feature vector of the face image to be recognized is first generated, and then the feature vector is compared with the feature vector of each face template image in the face template image library, and the The most similar is the recognition result.
- N is large, such as 100 million or more, the number of comparisons will also become very large, which will make the system
- the response speed is significantly reduced, which affects the efficiency of face recognition.
- This application provides a face recognition method, device, terminal, and storage medium to solve the current 1:N face recognition process, due to the large number of face template images, when comparing with the face image to be recognized one by one , The number of comparisons is huge, which leads to a slowdown in the response speed of the system and affects the efficiency of face recognition.
- This application provides a face recognition method, including the following steps:
- a face recognition device provided by this application includes:
- An extraction module configured to obtain a face image to be recognized, and extract the first high-dimensional feature vector of the face image to be recognized according to a preset deep learning model
- a quantization module configured to quantize the first high-dimensional feature vector into a first low-dimensional integer code using a product quantization algorithm
- the distance measurement module is used to perform similarity matching between the first low-dimensional integer code of the face image to be recognized and the second low-dimensional integer code of the face template image obtained in advance;
- the determining module is configured to determine the target low-dimensional integer code most similar to the first low-dimensional integer code in the second low-dimensional integer code according to the similarity matching result, and assign the person corresponding to the target low-dimensional integer code
- the face template image is used as the target face image.
- the present application provides a terminal, including a memory and a processor, the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor executes a face recognition method A step of;
- the face recognition method includes the following steps:
- This application provides a storage medium on which a computer program is stored, and when the computer program is executed by a processor, a face recognition method is implemented;
- the face recognition method includes the following steps:
- the face recognition method acquires a face image to be recognized, and extracts the first high-dimensional feature vector of the face image to be recognized according to a preset deep learning model; then uses a product quantization algorithm to calculate the first high-dimensional feature vector The dimensional feature vector is quantized into a small number of first low-dimensional integer codes; then the first low-dimensional integer code of the face image to be recognized is matched with the second low-dimensional integer code of the face template image obtained in advance; Finally, the target low-dimensional integer code most similar to the first low-dimensional integer code is determined in the second low-dimensional integer code according to the similarity matching result, and the face template image corresponding to the target low-dimensional integer code is used as In the recognition process of the target face image, due to the product quantization algorithm processing, a large number of high
- FIG. 1 is an implementation environment diagram of a face recognition method provided in an embodiment of the application
- Figure 2 is a flowchart of an embodiment of the applicant's face recognition method
- Figure 3 is a block diagram of an embodiment of the applicant's face recognition device
- Fig. 4 is a block diagram of the internal structure of a terminal in an embodiment of the application.
- FIG. 1 is an implementation environment diagram of a face recognition method provided in an embodiment.
- the implementation environment includes a server 110 and a terminal 120.
- the terminal 120 is connected to a server via a network, and the terminal 120 can be used to collect a face image and upload the face image to the server 110 for face recognition processing.
- the foregoing network may include the Internet, 2G/3G/4G, wi f i, etc.
- the server 110 may be an independent physical server or terminal, or a server cluster composed of multiple physical servers, and may be a cloud server that provides basic cloud computing services such as cloud servers, cloud databases, cloud storage, and CDN.
- the terminal 120 may be a smart phone, a tablet computer, a notebook computer, a desktop computer, a smart speaker, a smart watch, etc., but is not limited thereto.
- the face recognition method may include the following steps:
- S21 Acquire a face image to be recognized, and extract a first high-dimensional feature vector of the face image to be recognized according to a preset deep learning model.
- an image containing a face can be collected through a terminal device, the face part can be recognized from the image, and the face image to be recognized is obtained, and then the key to the face image to be recognized is extracted according to the preset deep learning model
- the feature point and the first high-dimensional feature vector corresponding to each key feature point may include 68 key feature points of a person's face such as the nose, mouth, eyes, and forehead.
- the feature value of the key feature point in the central area of the face can be extracted first to locate the face image, and then according to the central area of the face
- the key feature points sequentially extract the key feature points of other areas of the face image, and finally construct the geometric feature vector based on all the key feature points extracted, thereby mapping into the corresponding first high-dimensional feature vector to improve the first high-dimensional The extraction speed of feature vectors.
- the so-called product quantization is to decompose the feature space into Cartesian products of multiple low-dimensional subspaces, and then quantize each low-dimensional subspace individually.
- each subspace is clustered to obtain kk quantizers, and the Cartesian product of all these quantizers forms a dense division of the entire space, and can ensure that the quantization error is relatively small.
- the product quantization algorithm through the product quantization algorithm, a large number of high-dimensional feature vectors corresponding to the face image to be recognized and the face template image can be quantized into a small number of low-dimensional integer codes.
- S23 Perform similarity matching between the first low-dimensional integer code of the face image to be recognized and the second low-dimensional integer code of the face template image obtained in advance.
- the similarity can be calculated according to the corresponding characters of the first low-dimensional integer code and the second low-dimensional integer code.
- the first low-dimensional integer code is 11101111 and the second low-dimensional integer code is 11011111
- the number of characters in the two low-dimensional integer codes is 8 bits
- the characters that do not match are the third and fourth bits.
- the matched character has 6 bits, then the matching degree between the first low-dimensional integer code and the second low-dimensional integer code is That is 75%.
- the second low-dimensional integer code of the face template image obtained in advance is a low-dimensional integer code set, which includes multiple low-dimensional integer codes.
- the search time in the quantization coding space only depends on the number of quantization codes M, That is, the calculation amount is O(M), which can greatly reduce the calculation amount, so the distance calculation amount between the low-dimensional integer encoding of the face image to be recognized and the low-dimensional integer encoding of the face template image can be greatly reduced.
- the second low-dimensional integer code with the highest matching degree is the target low-dimensional integer encoding
- the face template image corresponding to the target low-dimensional integer encoding is the target face image most similar to the face image to be recognized, thereby completing face recognition.
- the face recognition method acquires a face image to be recognized, and extracts the first high-dimensional feature vector of the face image to be recognized according to a preset deep learning model; and then uses a product quantization algorithm to calculate the first high-dimensional feature vector
- the dimensional feature vector is quantized into a small number of first low-dimensional integer codes; then the first low-dimensional integer code of the face image to be recognized is matched with the second low-dimensional integer code of the face template image obtained in advance; Finally, the target low-dimensional integer code most similar to the first low-dimensional integer code is determined in the second low-dimensional integer code according to the similarity matching result, and the face template image corresponding to the target low-dimensional integer code is used as
- the product quantization algorithm processing a large number of high-dimensional feature vectors corresponding to the face image to be recognized and the face template image can be quantized into a small number of low-dimensional integer codes.
- the method may further include:
- the deep learning model may be a convolutional neural network model.
- Each base layer contains several nodes.
- the nodes between the base layer and the base layer are in a fully connected state, and the connections between the nodes Usually has a weight parameter.
- the weight parameters between nodes are parameter values that can be set arbitrarily.
- a large number of face template images can be input into the deep learning model, and according to the output results of the deep learning model, the weight parameters of the connections between the nodes on the deep learning model are continuously adjusted , Until the best weight parameters are obtained to obtain a trained deep learning model.
- the step of training the deep learning model using the face template image obtained in advance in step S20 may specifically include:
- S201 Input a face template image into a deep learning model, and obtain feature information corresponding to the face template image.
- the face template image when training the deep learning model, can be input into the deep learning model, the face template image is first positioned, and then all the feature information is extracted from the face template image.
- the feature information includes key feature points such as eyes, nose, ears and corresponding feature values.
- the method before inputting the face template image into the deep learning model, the method further includes:
- the face template image can be subjected to BoxBlur blurring processing to filter out the high frequency part in the face template image and retain only the low frequency part, thereby filtering out environmental factors and better retaining feature information.
- environmental factors such as illumination, occlusion, angle, etc. will affect the collection of eye information, resulting in a large deviation between the obtained feature information and its feature value and the true value.
- S202 Calculate the loss of the deep learning model based on the preset loss function and the feature information corresponding to the face template image; wherein the loss function is weighted by the Softmax function and the L-Softmax function.
- the loss function is usually used as its objective function.
- the loss function is used to evaluate the degree of difference between the predicted value of the deep learning model and the true value. The better the loss function, the better the performance of the deep learning model.
- the loss function in this embodiment may be weighted by the Softmax function and the L-Softmax function to improve the accuracy of the deep learning model loss calculation.
- the loss function is:
- the L Softmax is the Softmax function
- the L L-Softmax is the L-Softmax function
- ⁇ is the adjustment factor, and 5 ⁇ 8.
- the preset value can be flexibly set according to needs, and the smaller the preset value is set, the better the effect of the deep learning model obtained by training.
- the method may further include :
- the face template image is acquired, and the second high-dimensional feature vector of the face template image is extracted according to a preset deep learning model.
- a product quantization algorithm is used to quantize the second high-dimensional feature vector into a second low-dimensional integer code.
- This embodiment also needs to obtain a large number of face template images, save them in the face template image library, and extract the second highest value of each face template image from the face template image library according to the preset deep learning model. Then, the second high-dimensional feature vector is quantized into a second low-dimensional integer code by using a product quantization algorithm, so that the similarity of the first low-dimensional integer code and the second low-dimensional integer code can be matched subsequently.
- the face template image library pre-stores the second low-dimensional integer code corresponding to each face template image for subsequent similarity matching.
- the face image to be recognized and the face template image that is frequently queried may be matched first, thereby improving the matching. Efficiency, quickly find the most similar target face image.
- step S24 the step of using the face template image corresponding to the target low-dimensional integer code as the target face image may specifically include:
- S242 Calculate the distance between the first high-dimensional feature vector of the face image to be recognized and the target high-dimensional feature vector.
- the face template image corresponding to the target high-dimensional feature vector is used as the target face image.
- the target low-dimensional integer code corresponding to the target high-dimensional feature vector before quantization is obtained from the second high-dimensional feature vector, and then the target high-dimensional feature vector is calculated.
- the distance between the first high-dimensional feature vector and the target high-dimensional feature vector to perform an accurate 1:N comparison to verify whether the face template image corresponding to the target high-dimensional feature vector is a matching target face image .
- the distance is greater than the preset value, it means that the face template image corresponding to the target high-dimensional feature vector is the target face image, thereby ensuring the accuracy of recognition while accelerating the recognition speed.
- the precise comparison time at this time depends on the vector number L and the vector dimension V of the target high-dimensional feature vector, that is, the calculation amount is O(VL).
- the existing 1:N face recognition method is used from the person to be recognized
- the distance is calculated directly with the second high-dimensional feature vector extracted from each face template image in the face template image library, and finally the closest face template image is found as a comparison
- the amount of calculation required for its identification is O(VN).
- the calculation amount mainly includes: the calculation amount O(UV) of the first low-dimensional integer code of the face image to be recognized is quantized, and the first low-dimensional integer code of the face image to be recognized is searched from the face template image library.
- the calculation amount O(M) of the matched second low-dimensional integer encoding and the calculation amount of calculating the distance between the first high-dimensional feature vector of the face image to be recognized and the target high-dimensional feature vector therefore, the total calculation amount can be reduced to O(UV)+O(M)+O(VL), compared to the calculation before optimization O(VN), because N>>U,M,L, so O(VN)>>O(UV) +O(M)+O(VL), so the amount of calculation is greatly reduced, which can significantly improve the efficiency of face recognition.
- step S242 the step of calculating the distance between the first high-dimensional feature vector of the face image to be recognized and the target high-dimensional feature vector may specifically include:
- the Euclidean distance and the cosine distance are weighted and combined to obtain the distance.
- Euclidean distance is also called Euclidean metric or Euclidean distance. It is the true distance between two points in m-dimensional space. Euclidean distance in two-dimensional space is the straight line between two points. Some distance. Euclidean distance measures the absolute distance of each point in space, which is directly related to the position coordinates of each point; while cosine distance, also known as cosine similarity, is the cosine value of the angle between two vectors in vector space as a measure of two The measure of the size of the difference between individuals, which measures the angle of the space vector, is more reflected in the difference in direction rather than position.
- the Euclidean distance and the cosine distance between the first high-dimensional feature vector of the face image to be recognized and the target high-dimensional feature vector are calculated separately, and the two calculation results are comprehensively analyzed, for example, the Euclidean distance Set different reference ratios for distance and cosine distance. Based on the reference ratio, the Euclidean distance and cosine distance are weighted and combined to obtain a more accurate distance, and then the target high-dimensional feature vector corresponding to the distance is extracted and searched The face template image corresponding to the target high-dimensional feature vector completes face recognition, thereby improving the accuracy of face recognition.
- an embodiment of the present application also provides a face recognition device.
- it includes an extraction module 31, a quantization module 32, a matching module 33 and a determination module 34. among them,
- the extraction module 31 is configured to obtain a face image to be recognized, and extract the first high-dimensional feature vector of the face image to be recognized according to a preset deep learning model.
- an image containing a face can be collected through a terminal device, the face part can be recognized from the image, and the face image to be recognized is obtained, and then the key to the face image to be recognized is extracted according to the preset deep learning model
- the feature point and the first high-dimensional feature vector corresponding to each key feature point may include 68 key feature points of a person's face such as the nose, mouth, eyes, and forehead.
- the feature value of the key feature point in the central area of the face can be extracted first to locate the face image, and then according to the central area of the face
- the key feature points sequentially extract the key feature points of other areas of the face image, and finally construct the geometric feature vector based on all the key feature points extracted, thereby mapping into the corresponding first high-dimensional feature vector to improve the first high-dimensional The extraction speed of feature vectors.
- the quantization module 32 is configured to use a product quantization algorithm to quantize the first high-dimensional feature vector into a first low-dimensional integer code.
- the so-called product quantization is to decompose the feature space into Cartesian products of multiple low-dimensional subspaces, and then quantize each low-dimensional subspace individually.
- each subspace is clustered to obtain kk quantizers, and the Cartesian product of all these quantizers forms a dense division of the entire space, and can ensure that the quantization error is relatively small.
- the product quantization algorithm through the product quantization algorithm, a large number of high-dimensional feature vectors corresponding to the face image to be recognized and the face template image can be quantized into a small number of low-dimensional integer codes.
- the matching module 33 is configured to perform similarity matching between the first low-dimensional integer code of the face image to be recognized and the first low-dimensional integer code of the face template image obtained in advance.
- the similarity can be calculated according to the corresponding characters of the first low-dimensional integer code and the second low-dimensional integer code.
- the first low-dimensional integer code is 11101111 and the second low-dimensional integer code is 11011111
- the number of characters in the two low-dimensional integer codes is 8 bits
- the characters that do not match are the third and fourth bits.
- the matched character has 6 bits, then the matching degree between the first low-dimensional integer code and the second low-dimensional integer code is That is 75%.
- the second low-dimensional integer code of the face template image obtained in advance is a low-dimensional integer code set, which includes multiple low-dimensional integer codes.
- the search time in the quantization coding space only depends on the number of quantization codes M, That is, the calculation amount is O(M), which can greatly reduce the calculation amount, so the distance calculation amount between the low-dimensional integer encoding of the face image to be recognized and the low-dimensional integer encoding of the face template image can be greatly reduced.
- the determining module 34 is configured to determine a target low-dimensional integer code that is most similar to the first low-dimensional integer code in the second low-dimensional integer code according to the similarity matching result, and the target low-dimensional integer code corresponding to the The face template image is used as the target face image.
- the second low-dimensional integer code with the highest matching degree is the target low-dimensional integer code
- the face template image corresponding to the target low-dimensional integer code is a target face image similar to the face image to be recognized, thereby completing face recognition.
- the face recognition device acquires a face image to be recognized, and extracts the first high-dimensional feature vector of the face image to be recognized according to a preset deep learning model; then uses a product quantization algorithm to calculate the first height
- the dimensional feature vector is quantized into a small number of first low-dimensional integer codes; then the first low-dimensional integer code of the face image to be recognized is matched with the second low-dimensional integer code of the face template image obtained in advance;
- the target low-dimensional integer code most similar to the first low-dimensional integer code is determined in the second low-dimensional integer code according to the similarity matching result, and the face template image corresponding to the target low-dimensional integer code is used as
- the product quantization algorithm processing a large number of high-dimensional feature vectors corresponding to the face image to be recognized and the face template image can be quantized into a small number of low-dimensional integer codes.
- the extraction module 31 is further configured to:
- the deep learning model is trained by using the face template image obtained in advance to determine the optimal weight parameter of the connection between the nodes in the deep learning model.
- the extraction module 31 is further configured to:
- the face template image is input into the deep learning model, and feature information corresponding to the face template image is obtained.
- the loss function is:
- the L Softmax is the Softmax function
- the L L-Softmax is the L-Softmax function
- ⁇ is the adjustment factor, and 5 ⁇ 8.
- it further includes:
- a face template image acquisition module configured to acquire a face template image, and extract the second high-dimensional feature vector of the face template image according to a preset deep learning model
- the second high-dimensional feature vector quantization module is configured to use a product quantization algorithm to quantize the second high-dimensional feature vector into a second low-dimensional integer code.
- the determining module 34 is further configured to:
- the face template image corresponding to the target high-dimensional feature vector is used as the target face image.
- the determining module 34 is further configured to:
- the Euclidean distance and the cosine distance are weighted and combined to obtain the distance.
- a terminal provided by the present application includes a memory and a processor, and computer-readable instructions are stored in the memory.
- the processor executes any of the above The steps of the face recognition method described.
- the terminal is a computer device, as shown in FIG. 4.
- the computer equipment described in this embodiment may be equipment such as servers, personal computers, and network equipment.
- the computer equipment includes a processor 402, a memory 403, an input unit 404, a display unit 405 and other devices.
- the memory 403 may be used to store a computer program 401 and various functional modules, and the processor 402 runs the computer program 401 stored in the memory 403 to execute various functional applications and data processing of the device.
- the memory may be internal memory or external memory, or include both internal memory and external memory.
- the internal memory may include read only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, or random access memory.
- ROM read only memory
- PROM programmable ROM
- EPROM electrically programmable ROM
- EEPROM electrically erasable programmable ROM
- flash memory or random access memory.
- External storage can include hard disk, floppy disk, ZIP disk, U disk, tape, etc.
- the memory disclosed in this application includes but is not limited to these types of memory.
- the memory disclosed in this application is only an example and not a limitation.
- the input unit 404 is used for receiving input of signals and receiving keywords input by the user.
- the input unit 404 may include a touch panel and other input devices.
- the touch panel can collect the user's touch operations on or near it (for example, the user uses any suitable objects or accessories such as fingers, stylus, etc., to operate on the touch panel or near the touch panel), and according to preset
- the program drives the corresponding connection device; other input devices may include, but are not limited to, one or more of a physical keyboard, function keys (such as playback control buttons, switch buttons, etc.), trackball, mouse, and joystick.
- the display unit 405 can be used to display information input by the user or information provided to the user and various menus of the computer device.
- the display unit 405 can take the form of a liquid crystal display, an organic light emitting diode, or the like.
- the processor 402 is the control center of the computer equipment. It uses various interfaces and lines to connect the various parts of the entire computer. It executes by running or executing the software programs and/or modules stored in the memory 402 and calling the data stored in the memory. Various functions and processing data.
- the computer device includes: one or more processors 402, a memory 403, and one or more computer programs 401, wherein the one or more computer programs 401 are stored in the memory 403 and configured to Executed by the one or more processors 402, the one or more computer programs 401 are configured to execute a face recognition method, wherein the face recognition method includes the following steps: acquiring a face image to be recognized , Extracting the first high-dimensional feature vector of the face image to be recognized according to a preset deep learning model; using a product quantization algorithm to quantize the first high-dimensional feature vector into a first low-dimensional integer code; The first low-dimensional integer code of the face image is matched with the second low-dimensional integer code of the face template image obtained in advance; according to the similarity matching result, the second low-dimensional integer code is determined to be the same as the first The low-dimensional integer code is the most similar target low-dimensional integer code, and the face template image corresponding to the target low-dimensional integer code is used as the target face image.
- the face recognition method includes the
- this application also proposes a storage medium storing computer-readable instructions.
- the one or more processors execute a face A recognition method, wherein the face recognition method includes the following steps: acquiring a face image to be recognized, extracting a first high-dimensional feature vector of the face image to be recognized according to a preset deep learning model; using a product quantization algorithm to Quantizing the first high-dimensional feature vector into a first low-dimensional integer code; performing similarity matching between the first low-dimensional integer code of the face image to be recognized and the second low-dimensional integer code of the face template image obtained in advance; Determine the target low-dimensional integer code most similar to the first low-dimensional integer code in the second low-dimensional integer code according to the similarity matching result, and use the face template image corresponding to the target low-dimensional integer code as the target Face image.
- the storage medium may be ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, optical data storage device, etc.
- the aforementioned storage medium may be a non-volatile storage medium such as a magnetic disk, an optical disc, a read-only memory (Read-Only Memory, ROM), or a random access memory (Random Access Memory, RAM), etc.
- the face recognition method, device, terminal, and storage medium provided in this application acquire the face image to be recognized, and extract the first high-dimensional feature vector of the face image to be recognized according to a preset deep learning model; then use the product to quantify The algorithm quantizes the first high-dimensional feature vector into a small number of first low-dimensional integer codes; then the first low-dimensional integer code of the face image to be recognized is combined with the second low-dimensional code of the face template image obtained in advance Integer coding performs similarity matching; finally, according to the similarity matching result, the target low-dimensional integer code most similar to the first low-dimensional integer code is determined in the second low-dimensional integer code, and the target low-dimensional integer code is encoded The corresponding face template image is used as the target face image.
- a large number of high-dimensional feature vectors corresponding to the face image to be recognized and the face template image can be quantified into a relatively large number.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Multimedia (AREA)
- Artificial Intelligence (AREA)
- General Health & Medical Sciences (AREA)
- Oral & Maxillofacial Surgery (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Human Computer Interaction (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computing Systems (AREA)
- Databases & Information Systems (AREA)
- Medical Informatics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
Abstract
本申请提供了一种人脸识别方法、装置、终端及存储介质。该人脸识别方法包括:获取待识别人脸图像,根据预设的深度学习模型提取待识别人脸图像的第一高维特征向量;利用乘积量化算法将第一高维特征向量量化为第一低维整数编码;将待识别人脸图像的第一低维整数编码与预先得到的人脸模板图像的第二低维整数编码进行相似度匹配;根据相似度匹配结果在所述第二低维整数编码中确定与所述第一低维整数编码最相似的目标低维整数编码,将目标低维整数编码对应的人脸模板图像作为目标人脸图像。本申请在人脸识别过程中,利用低维整数编码进行相似度比较,从而减少两者之间的比对次数,大大加快了系统的响应速度,提升人脸识别效率。
Description
本申请要求于2019年5月29日提交中国专利局、申请号为201910457256.8,发明名称为“人脸识别方法、装置、终端及存储介质”的中国专利申请的优先权,其全部内容通过引用结合在本申请中。
本申请涉及人脸识别技术领域,尤其涉及一种人脸识别方法、装置、终端及存储介质。
随着人脸识别技术的发展,人脸识别被广泛用于安全监控、安全支付、屏幕解锁及考勤等领域,终端设备通过摄像装置实时采集人脸图像,并与预先存储的人脸模板图像进行匹配,在匹配成功后,则验证通过。
具体的,在1:N人脸识别中,首先会生成待识别人脸图像的特征向量,然后将该特征向量与人脸模板图像库中每个人脸模板图像的特征向量进行相似度比较,取最相似的作为识别结果。发明人意识到,在此过程中,特征向量的两两比对速度本身是比较快的,但是当N很大,比如1亿以上时,比对次数也会变得非常巨大,会使系统的响应速度明显降低,从而影响人脸识别效率。
发明内容
本申请提供一种人脸识别方法、装置、终端及存储介质,以解决当前1:N人脸识别过程中,由于人脸模板图像数量庞大,在与待识别人脸图像进行一一比对时,比对次数巨大,导致系统的响应速度降低,影响人脸识别效率的问题。
为解决上述问题,本申请采用如下技术方案:
本申请提供一种人脸识别方法,包括如下步骤:
获取待识别人脸图像,根据预设的深度学习模型提取所述待识别人脸图像的第一高维特征向量;
利用乘积量化算法将所述第一高维特征向量量化为第一低维整数编码;
将待识别人脸图像的第一低维整数编码与预先得到的人脸模板图像的第二低维整数编码进行相似度匹配;
根据相似度匹配结果在所述第二低维整数编码中确定与所述第一低维整数编码最相似的目标低维整数编码,将所述目标低维整数编码对应的人脸模板图像作为目标人脸图像。
本申请提供的一种人脸识别装置,包括:
提取模块,用于获取待识别人脸图像,根据预设的深度学习模型提取所述待识别人脸图像的第一高维特征向量;
量化模块,用于利用乘积量化算法将所述第一高维特征向量量化为第一低维整数编码;
距离度量模块,用于将待识别人脸图像的第一低维整数编码与预先得到的人脸模板图像的第二低维整数编码进行相似度匹配;
确定模块,用于根据相似度匹配结果在所述第二低维整数编码中确定与所述第一低维整数编码最相似的目标低维整数编码,将所述目标低维整数编码对应的人脸模板图像作为目标人脸图像。
本申请提供一种终端,包括存储器和处理器,所述存储器中存储有计算机可读指令,所述计算机可读指令被所述处理器执行时,使得所述处理器执行一种人脸识别方法的步骤;
其中,所述人脸识别方法包括以下步骤:
获取待识别人脸图像,根据预设的深度学习模型提取所述待识别人脸图像的第一高维特征向量;
利用乘积量化算法将所述第一高维特征向量量化为第一低维整数编码;
将待识别人脸图像的第一低维整数编码与预先得到的人脸模板图像的第二低维整数编码进行相似度匹配;
根据相似度匹配结果在所述第二低维整数编码中确定与所述第一低维整数编码最相似的目标低维整数编码,将所述目标低维整数编码对应的人脸模板图像作为目标人脸图像。
本申请提供一种存储介质,其上存储有计算机程序,所述计算机程序被处理器执行时,实现一种人脸识别方法;
其中,所述人脸识别方法包括以下步骤:
获取待识别人脸图像,根据预设的深度学习模型提取所述待识别人脸图像的第一高维特征向量;
利用乘积量化算法将所述第一高维特征向量量化为第一低维整数编码;
将待识别人脸图像的第一低维整数编码与预先得到的人脸模板图像的第二低维整数编码进行相似度匹配;
根据相似度匹配结果在所述第二低维整数编码中确定与所述第一低维整数编码最相似的目标低维整数编码,将所述目标低维整数编码对应的人脸模板图像作为目标人脸图像。本申请提供的人脸识别方法,通过获取待识别人脸图像,并根据预设的深度学习模型提取待识别人脸图像的第一高维特征向量;然后利用乘积量化算法将所述第一高维特征向量量化为数量较少的第一低维整数编码;再将待识别人脸图像的第一低维整数编码与预先得到的人脸模板图像的第二低维整数编码进行相似度匹配;最后根据相似度匹配结果在所述第二低维整数编码中确定与所述第一低维整数编码最相似的目标低维整数编码,将所述目标低维整数编码对应的人脸模板图像作为目标人脸图像,在识别过程中,由于经过乘积量化算法处理后,可将数量庞大的待识别人脸图像和人脸模板图像对应的高维特征向量量化为数量较少的低维整数编码,利用低维整数编码进行相似度比较,从而减少两者之间的比对次数,大大加快了系统的响应速度,提升人脸识别效率。
图1为本申请一个实施例中提供的人脸识别方法的实施环境图;
图2为本申请人脸识别方法一种实施例流程图;
图3为本申请人脸识别装置一种实施例模块框图;
图4为本申请一个实施例中终端的内部结构框图。
为了使本技术领域的人员更好地理解本申请方案,下面将结合本申请实施例中的附图,对本申请实施例中的技术方案进行清楚、完整地描述。
在本申请的说明书和权利要求书及上述附图中的描述的一些流程中,包 含了按照特定顺序出现的多个操作,但是应该清楚了解,这些操作可以不按照其在本文中出现的顺序来执行或并行执行,操作的序号如S11、S12等,仅仅是用于区分开各个不同的操作,序号本身不代表任何的执行顺序。另外,这些流程可以包括更多或更少的操作,并且这些操作可以按顺序执行或并行执行。需要说明的是,本文中的“第一”、“第二”等描述,是用于区分不同的消息、设备、模块等,不代表先后顺序,也不限定“第一”和“第二”是不同的类型。
本领域普通技术人员可以理解,除非特意声明,这里使用的单数形式“一”、“一个”、“所述”和“该”也可包括复数形式。应该进一步理解的是,本申请的说明书中使用的措辞“包括”是指存在所述特征、整数、步骤、操作、元件和/或组件,但是并不排除存在或添加一个或多个其他特征、整数、步骤、操作、元件、组件和/或它们的组。应该理解,当我们称元件被“连接”或“耦接”到另一元件时,它可以直接连接或耦接到其他元件,或者也可以存在中间元件。此外,这里使用的“连接”或“耦接”可以包括无线连接或无线耦接。这里使用的措辞“和/或”包括一个或更多个相关联的列出项的全部或任一单元和全部组合。
本领域普通技术人员可以理解,除非另外定义,这里使用的所有术语(包括技术术语和科学术语),具有与本申请所属领域中的普通技术人员的一般理解相同的意义。还应该理解的是,诸如通用字典中定义的那些术语,应该被理解为具有与现有技术的上下文中的意义一致的意义,并且除非像这里一样被特定定义,否则不会用理想化或过于正式的含义来解释。
下面将结合本申请实施例中的附图,对本申请实施例中的技术方案进行清楚、完整地描述,其中自始至终相同或类似的标号表示相同或类似的元件或具有相同或类似功能的元件。显然,所描述的实施例仅仅是本申请一部分实施例,而不是全部的实施例。基于本申请中的实施例,本领域技术人员在没有作出创造性劳动前提下所获得的所有其他实施例,都属于本申请保护的范围。
图1为一个实施例中提供的人脸识别方法的实施环境图,如图1所示,在该实施环境中,包括服务器110、终端120。终端120通过网络与服务器连接,所述终端120可用于采集人脸图像,并将人脸图像上传至服务器110进 行人脸识别处理。其中,上述网络可以包括因特网、2G/3G/4G、wi f i等。
需要说明的是,服务器110可以是独立的物理服务器或终端,也可以是多个物理服务器构成的服务器集群,可以是提供云服务器、云数据库、云存储和CDN等基础云计算服务的云服务器。
终端120可以是智能手机、平板电脑、笔记本电脑、台式计算机、智能音箱、智能手表等,但并不局限于此。
请参阅图2,本申请提供了一种人脸识别方法,以解决当前1:N人脸识别过程中,由于人脸模板图像数量庞大,在与待识别人脸图像进行一一比对时,比对次数巨大,导致系统的响应速度降低,影响人脸识别效率的问题。其中一种实施方式中,所述人脸识别方法可包括如下步骤:
S21、获取待识别人脸图像,根据预设的深度学习模型提取所述待识别人脸图像的第一高维特征向量。
在进行人脸识别前,可通过终端设备采集包含人脸的图像,从图像中识别出人脸部分,得到待识别人脸图像,然后根据预设的深度学习模型提取待识别人脸图像的关键特征点及每一关键特征点对应的第一高维特征向量。其中,所述关键特征点可包括人的鼻子、嘴巴、眼睛、额头等68个人脸关键特征点。
在利用深度学习模型提取待识别人脸图像的第一高维特征向量时,可先提取人脸中心区域的关键特征点的特征值,以对人脸图像进行定位,然后根据人脸中心区域的关键特征点依次提取出人脸图像其他区域的关键特征点,最后根据提取出的所有关键特征点进行几何特征向量构造,从而映射成相对应的第一高维特征向量,以提高第一高维特征向量的提取速度。
S22、利用乘积量化算法将所述第一高维特征向量量化为第一低维整数编码。
所谓乘积量化,即将特征空间分解为多个低维子空间的笛卡尔乘积,然后单独地对每一个低维子空间进行量化。在训练阶段,每一个子空间经过聚类后得到kk个量化器,所有这些量化器的笛卡尔乘积构成了一个对全空间的密集划分,并且能够保证量化误差比较小。在本实施例中,通过乘积量化算法,可分别将数量庞大的待识别人脸图像和人脸模板图像对应的高维特征向量量化为数量较少的低维整数编码。
S23、将待识别人脸图像的第一低维整数编码与预先得到的人脸模板图像的第二低维整数编码进行相似度匹配。
在本实施例中,将待识别人脸图像与人脸模板图像进行比对时,可根据第一低维整数编码与第二低维整数编码的对应字符计算相似度,对应字符相同的数量越多,则匹配度越高。例如,当第一低维整数编码为11101111,第二低维整数编码为11011111时,两个低维整数编码的字符数量都为8位,不匹配的字符为第三位和第四位,相匹配的字符有6位,则该第一低维整数编码与第二低维整数编码的匹配度为
即75%。其中,预先得到的人脸模板图像的第二低维整数编码是一个低维整数编码集合,其包括多个低维整数编码。
本实施例中,由于待识别人脸图像和人脸模板图像的低维整数编码的数量有限,且在比对过程中,其量化编码空间中的搜索时间只取决于量化编码的个数M,即计算量为O(M),可大大减少计算量,所以待识别人脸图像的低维整数编码与人脸模板图像的低维整数编码之间的距离计算量可大大减少。
S24、根据相似度匹配结果在所述第二低维整数编码中确定与所述第一低维整数编码最相似的目标低维整数编码,将所述目标低维整数编码对应的人脸模板图像作为目标人脸图像。
在本实施例中,当待识别人脸图像的第一低维整数编码与预先得到的人脸模板图像的第二低维整数编码的匹配度最高时,则该匹配度最高的第二低维整数编码为目标低维整数编码,且该目标低维整数编码对应的人脸模板图像为与待识别人脸图像最相似的目标人脸图像,从而完成人脸识别。
本申请提供的人脸识别方法,通过获取待识人脸图像,并根据预设的深度学习模型提取待识别人脸图像的第一高维特征向量;然后利用乘积量化算法将所述第一高维特征向量量化为数量较少的第一低维整数编码;再将待识别人脸图像的第一低维整数编码与预先得到的人脸模板图像的第二低维整数编码进行相似度匹配;最后根据相似度匹配结果在所述第二低维整数编码中确定与所述第一低维整数编码最相似的目标低维整数编码,将所述目标低维整数编码对应的人脸模板图像作为目标人脸图像,在识别过程中,由于经过乘积量化算法处理后,可将数量庞大的待识别人脸图像和人脸模板图像对应的高维特征向量量化为数量较少的低维整数编码,利用低维整数编码进行相似度比较,从而减少两者之间的比对次数,大大加快了系统的响应速度,提 升人脸识别效率。
在一实施例中,步骤S21的根据预设的深度学习模型提取所述待识别人脸图像的第一高维特征向量之前,还可包括:
S20、利用预先得到的人脸模板图像对所述深度学习模型进行训练,以确定所述深度学习模型中各节点之间的连接的最佳权重参数。
在本实施例中,所述深度学习模型可为卷积神经网络模型,其每一个基层都包含若干个节点,基层与基层之间的节点处于一种全连接的状态,且节点之间的连接通常具有一个权重参数。在深度学习模型训练之前,节点之间的权重参数为随意设置的参数值。在对深度学习模型进行训练时,可以将海量的人脸模板图像输入深度学习模型中,根据深度学习模型的输出结果,不断对该深度学习模型各基层上节点之间的连接的权重参数进行调整,直至得到最佳权重参数,以获得训练合格的深度学习模型。
在一实施例中,步骤S20的利用预先得到的人脸模板图像对所述深度学习模型进行训练的步骤,可具体包括:
S201、将人脸模板图像输入深度学习模型中,获取所述人脸模板图像对应的特征信息。
在本实施例中,对深度学习模型进行训练时,可将人脸模板图像输入深度学习模型中,先对人脸模板图像进行定位,然后从人脸模板图像中提取出所有特征信息。其中,所述特征信息包括诸如眼睛、鼻子、耳朵的关键特征点及相应的特征值。
在一实施例中,在将人脸模板图像输入深度学习模型之前,还包括:
对人脸模板图像进行BoxBlur模糊处理。
本实施例可通过将人脸模板图像进行BoxBlur模糊处理,以滤除人脸模板图像中的高频部分并只保留低频部分,从而滤除环境因素,以较好的保留特征信息。例如,光照、遮挡、角度等环境因素会影响眼睛信息的采集,使得到的特征信息及其特征值与真实值之间存在较大偏差。
S202、基于预设的损失函数和所述人脸模板图像对应的特征信息,计算所述深度学习模型的损失;其中,所述损失函数由Softmax函数和L-Softmax函数加权构成。
通常机器学习的每一个算法中都会有一个目标函数,算法的求解过程即 是对这个目标函数不断优化的过程。在分类或者回归问题中,通常使用损失函数作为其目标函数。损失函数用来评价深度学习模型的预测值和真实值不一样的程度,损失函数越好,则深度学习模型的性能也越好。本实施例的损失函数可由Softmax函数和L-Softmax函数加权构成,以提高深度学习模型损失计算的准确性。
在一实施例中,所述损失函数为:
其中,所述L
Softmax为Softmax函数,所述L
L-Softmax为L-Softmax函数,∝为调节因子,5≤∝≤8。从而通过将Softmax函数和L-Softmax函数设置一定的比重,达到最佳的损失计算结果。
S203、当损失高于预设值时,调整所述深度学习模型中各节点之间的连接的权重参数,对深度学习模型重新训练,直至损失低于或等于预设值时,得到最佳权重参数及其对应的深度学习模型。
计算深度学习模型的损失后,判断其损失是否大于预设值,若是,则调整所述深度学习模型中各节点之间的连接的权重参数,对深度学习模型重新训练,然后计算其损失,直至损失低于或等于预设值时,获得深度学习模型中各节点之间的连接的最佳权重参数,从而得到得到最佳权重参数对应的深度学习模型,即得到训练合格的深度学习模型。其中,所述预设值可根据需要灵活设定,预设值设定得越小,则训练得到的深度学习模型的效果越好。
在一实施例中,在步骤S23中,所述将待识别人脸图像的第一低维整数编码与预先得到的人脸模板图像的第二低维整数编码进行相似度匹配之前,还可包括:
获取人脸模板图像,根据预设的深度学习模型提取所述人脸模板图像的第二高维特征向量。
利用乘积量化算法将所述第二高维特征向量量化为第二低维整数编码。
本实施例还需获取海量的人脸模板图像,将其保存在人脸模板图像库中,并根据预设的深度学习模型从人脸模板图像库中提取出各人脸模板图像的第二高维特征向量,然后利用乘积量化算法将所述第二高维特征向量量化为第二低维整数编码,以便于后续进行第一低维整数编码与第二低维整数编码的相似度匹配。其中,人脸模板图像库中预先存储有每一人脸模板图像对应的 第二低维整数编码,用于后续相似度匹配。
在一实施例中,由于目标人脸图像的被查找频率往往比较高,在相似度匹配时,可将待识别人脸图像与被查询频率较高的人脸模板图像优先进行匹配,从而提高匹配效率,快速找到最相似的目标人脸图像。
在一实施例中,步骤S24中,所述将所述目标低维整数编码对应的人脸模板图像作为目标人脸图像的步骤,可具体包括:
S241、在所述第二高维特征向量中获取所述目标低维整数编码对应量化前的目标高维特征向量。
S242、计算待识别人脸图像的第一高维特征向量与目标高维特征向量的距离。
当所述距离大于预设值时,将所述目标高维特征向量对应的人脸模板图像作为目标人脸图像。
在本实施例中,当得到最相似的目标低维整数编码后,在第二高维特征向量中获取该目标低维整数编码对应量化前目标高维特征向量,然后计算待识别人脸图像的第一高维特征向量与所述目标高维特征向量的距离,以进行精确的1:N比对,从而验证该目标高维特征向量对应的人脸模板图像是否为相匹配的目标人脸图像。当距离大于预设值时,则表示该目标高维特征向量对应的人脸模板图像为目标人脸图像,从而在加快识别速度的情况下保证识别的精准性。此时的精确比对时间取决于目标高维特征向量的向量数L及向量维数V,即计算量为O(VL)。
具体地,假设待识别的人脸图像的第一高维特征向量维数为V维,人脸模板图像的数量为N,采用现有的1:N人脸识别方法中,从待识别的人脸图像中提取出第一高维特征向量之后,与人脸模板图像库中每个人脸模板图像提取的第二高维特征向量直接两两求距离,最后寻找距离最近的人脸模板图像作为比对结果,其识别所需计算量为O(VN)。而通过使用乘积量化算法把高维特征向量量化为低维整数编码后,由于低维整数编码的数量较少,在将高维特征向量间的距离计算转换为低维整数编码间的比对后,其计算量主要包括:待识别人脸图像量化为第一低维整数编码的计算量O(UV)、从人脸模板图像库中查找与待识别人脸图像的第一低维整数编码相匹配的第二低维整数编码的计算量O(M)及计算待识别人脸图像的第一高维特征向量与目标高维 特征向量的距离的计算量,因此,总计算量可缩少为O(UV)+O(M)+O(VL),相比于未优化前的计算量O(VN),由于N>>U,M,L,所以O(VN)>>O(UV)+O(M)+O(VL),因此计算量大大减少,可显著提高人脸识别效率。
在一实施例中,步骤S242中,所述计算待识别人脸图像的第一高维特征向量与目标高维特征向量的距离的步骤,可具体包括:
计算待识别人脸图像的第一高维特征向量与目标高维特征向量的欧氏距离和余弦距离。
将所述欧氏距离和余弦距离进行加权组合,得到所述距离。
欧氏距离也称为欧几里得度量、欧几里得距离,它是在m维空间中两个点之间的真实距离,在二维空间中的欧氏距离就是两点之间的直线段距离。欧氏距离衡量的是空间各点的绝对距离,跟各个点所在的位置坐标直接相关;而余弦距离,也称为余弦相似度,是用向量空间中两个向量夹角的余弦值作为衡量两个个体间差异的大小的度量,其衡量的是空间向量的夹角,更加体现在方向上的差异,而不是位置。因此,本实施例通过分别计算待识别人脸图像的第一高维特征向量与目标高维特征向量的欧氏距离和余弦距离,并将两个计算结果进行综合分析,例如,分别对欧氏距离和余弦距离设置不同的参考比例,基于该参考比例将所述欧氏距离和余弦距离进行加权组合,从而得到准确性更高的距离,然后提取该距离对应的目标高维特征向量,并查找该目标高维特征向量对应的人脸模板图像,完成人脸识别,从而提高人脸识别的精准度。
请参考图3,本申请的实施例还提供一种人脸识别装置,一种本实施例中,包括提取模块31、量化模块32、匹配模块33及确定模块34。其中,
提取模块31,用于获取待识别人脸图像,根据预设的深度学习模型提取所述待识别人脸图像的第一高维特征向量。
在进行人脸识别前,可通过终端设备采集包含人脸的图像,从图像中识别出人脸部分,得到待识别人脸图像,然后根据预设的深度学习模型提取待识别人脸图像的关键特征点及每一关键特征点对应的第一高维特征向量。其中,所述关键特征点可包括人的鼻子、嘴巴、眼睛、额头等68个人脸关键特征点。
在利用深度学习模型提取待识别人脸图像的第一高维特征向量时,可先提取人脸中心区域的关键特征点的特征值,以对人脸图像进行定位,然后根据人脸中心区域的关键特征点依次提取出人脸图像其他区域的关键特征点,最后根据提取出的所有关键特征点进行几何特征向量构造,从而映射成相对应的第一高维特征向量,以提高第一高维特征向量的提取速度。
量化模块32,用于利用乘积量化算法将所述第一高维特征向量量化为第一低维整数编码。
所谓乘积量化,即将特征空间分解为多个低维子空间的笛卡尔乘积,然后单独地对每一个低维子空间进行量化。在训练阶段,每一个子空间经过聚类后得到kk个量化器,所有这些量化器的笛卡尔乘积构成了一个对全空间的密集划分,并且能够保证量化误差比较小。在本实施例中,通过乘积量化算法,可分别将数量庞大的待识别人脸图像和人脸模板图像对应的高维特征向量量化为数量较少的低维整数编码。
匹配模块33,用于将待识别人脸图像的第一低维整数编码与预先得到的人脸模板图像的第一低维整数编码进行相似度匹配。
在本实施例中,将待识别人脸图像与人脸模板图像进行比对时,可根据第一低维整数编码与第二低维整数编码的对应字符计算相似度,对应字符相同的数量越多,则匹配度越高。例如,当第一低维整数编码为11101111,第二低维整数编码为11011111时,两个低维整数编码的字符数量都为8位,不匹配的字符为第三位和第四位,相匹配的字符有6位,则该第一低维整数编码与第二低维整数编码的匹配度为
即75%。其中,预先得到的人脸模板图像的第二低维整数编码是一个低维整数编码集合,其包括多个低维整数编码。
本实施例中,由于待识别人脸图像和人脸模板图像的低维整数编码的数量有限,且在比对过程中,其量化编码空间中的搜索时间只取决于量化编码的个数M,即计算量为O(M),可大大减少计算量,所以待识别人脸图像的低维整数编码与人脸模板图像的低维整数编码之间的距离计算量可大大减少。
确定模块34,用于根据相似度匹配结果在所述第二低维整数编码中确定与所述第一低维整数编码最相似的目标低维整数编码,将所述目标低维整数编码对应的人脸模板图像作为目标人脸图像。
在本实施例中,当待识别人脸图像的第一低维整数编码与预先得到的人 脸模板图像的第二低维整数编码的匹配度最高时,则该匹配度最高的第二低维整数编码为目标低维整数编码,且该目标低维整数编码对应的人脸模板图像为与待识别人脸图像相似的目标人脸图像,从而完成人脸识别。
本申请提供的人脸识别装置,通过获取待识别人脸图像,并根据预设的深度学习模型提取待识别人脸图像的第一高维特征向量;然后利用乘积量化算法将所述第一高维特征向量量化为数量较少的第一低维整数编码;再将待识别人脸图像的第一低维整数编码与预先得到的人脸模板图像的第二低维整数编码进行相似度匹配;最后根据相似度匹配结果在所述第二低维整数编码中确定与所述第一低维整数编码最相似的目标低维整数编码,将所述目标低维整数编码对应的人脸模板图像作为目标人脸图像,在识别过程中,由于经过乘积量化算法处理后,可将数量庞大的待识别人脸图像和人脸模板图像对应的高维特征向量量化为数量较少的低维整数编码,利用低维整数编码进行相似度比较,从而减少两者之间的比对次数,大大加快了系统的响应速度,提升人脸识别效率。
在一实施例中,所述提取模块31还被配置为:
利用预先得到的人脸模板图像对所述深度学习模型进行训练,以确定所述深度学习模型中各节点之间的连接的最佳权重参数。
在一实施例中,所述提取模块31还被配置为:
将人脸模板图像输入深度学习模型中,获取所述人脸模板图像对应的特征信息。
基于预设的损失函数和所述人脸模板图像对应的特征信息,计算所述深度学习模型的损失;其中,所述损失函数由Softmax函数和L-Softmax函数加权构成。
当损失高于预设值时,调整所述深度学习模型中各节点之间的连接的权重参数,对深度学习模型重新训练,直至损失低于或等于预设值时,得到最佳权重参数及其对应的深度学习模型。
在一实施例中,所述损失函数为:
其中,所述L
Softmax为Softmax函数,所述L
L-Softmax为L-Softmax函数,∝为调节因子,5≤∝≤8。
在一实施例中,还包括:
人脸模板图像获取模块,用于获取人脸模板图像,根据预设的深度学习模型提取所述人脸模板图像的第二高维特征向量;
第二高维特征向量量化模块,用于利用乘积量化算法将所述第二高维特征向量量化为第二低维整数编码。
在一实施例中,所述确定模块34还被配置为:
在所述第二高维特征向量中获取所述目标低维整数编码对应量化前的目标高维特征向量。
计算待识别人脸图像的第一高维特征向量与目标高维特征向量的距离。
当所述距离大于预设值时,将所述目标高维特征向量对应的人脸模板图像作为目标人脸图像。
在一实施例中,所述确定模块34还被配置为:
计算待识别人脸图像的第一高维特征向量与目标高维特征向量的欧氏距离和余弦距离。
将所述欧氏距离和余弦距离进行加权组合,得到所述距离。
关于上述实施例中的装置,其中各个模块执行操作的具体方式已经在有关该方法的实施例中进行了详细描述,此处将不做详细阐述说明。
本申请提供的一种终端,包括存储器和处理器,所述存储器中存储有计算机可读指令,所述计算机可读指令被所述处理器执行时,使得所述处理器执行如上任一项所述的人脸识别方法的步骤。
在一实施例中,所述终端为一种计算机设备,如图4所示。本实施例所述的计算机设备可以是服务器、个人计算机以及网络设备等设备。所述计算机设备包括处理器402、存储器403、输入单元404以及显示单元405等器件。本领域技术人员可以理解,图4示出的设备结构器件并不构成对所有设备的限定,可以包括比图示更多或更少的部件,或者组合某些部件。存储器403可用于存储计算机程序401以及各功能模块,处理器402运行存储在存储器403的计算机程序401,从而执行设备的各种功能应用以及数据处理。存储器可以是内存储器或外存储器,或者包括内存储器和外存储器两者。内存储器可以包括只读存储器(ROM)、可编程ROM(PROM)、电可编程ROM(EPROM)、电可擦写可编程ROM(EEPROM)、快闪存储器、或者随机存储器。外存储器 可以包括硬盘、软盘、ZIP盘、U盘、磁带等。本申请所公开的存储器包括但不限于这些类型的存储器。本申请所公开的存储器只作为例子而非作为限定。
输入单元404用于接收信号的输入,以及接收用户输入的关键字。输入单元404可包括触控面板以及其它输入设备。触控面板可收集用户在其上或附近的触摸操作(比如用户使用手指、触笔等任何适合的物体或附件在触控面板上或在触控面板附近的操作),并根据预先设定的程序驱动相应的连接装置;其它输入设备可以包括但不限于物理键盘、功能键(比如播放控制按键、开关按键等)、轨迹球、鼠标、操作杆等中的一种或多种。显示单元405可用于显示用户输入的信息或提供给用户的信息以及计算机设备的各种菜单。显示单元405可采用液晶显示器、有机发光二极管等形式。处理器402是计算机设备的控制中心,利用各种接口和线路连接整个电脑的各个部分,通过运行或执行存储在存储器402内的软件程序和/或模块,以及调用存储在存储器内的数据,执行各种功能和处理数据。
作为一个实施例,所述计算机设备包括:一个或多个处理器402,存储器403,一个或多个计算机程序401,其中所述一个或多个计算机程序401被存储在存储器403中并被配置为由所述一个或多个处理器402执行,所述一个或多个计算机程序401配置用于执行一种人脸识别方法,其中,所述人脸识别方法包括以下步骤:获取待识别人脸图像,根据预设的深度学习模型提取所述待识别人脸图像的第一高维特征向量;利用乘积量化算法将所述第一高维特征向量量化为第一低维整数编码;将待识别人脸图像的第一低维整数编码与预先得到的人脸模板图像的第二低维整数编码进行相似度匹配;根据相似度匹配结果在所述第二低维整数编码中确定与所述第一低维整数编码最相似的目标低维整数编码,将所述目标低维整数编码对应的人脸模板图像作为目标人脸图像。
在一个实施例中,本申请还提出了一种存储有计算机可读指令的存储介质,该计算机可读指令被一个或多个处理器执行时,使得一个或多个处理器执行一种人脸识别方法,其中,所述人脸识别方法包括以下步骤:获取待识别人脸图像,根据预设的深度学习模型提取所述待识别人脸图像的第一高维特征向量;利用乘积量化算法将所述第一高维特征向量量化为第一低维整数编码;将待识别人脸图像的第一低维整数编码与预先得到的人脸模板图像的 第二低维整数编码进行相似度匹配;根据相似度匹配结果在所述第二低维整数编码中确定与所述第一低维整数编码最相似的目标低维整数编码,将所述目标低维整数编码对应的人脸模板图像作为目标人脸图像。例如,所述存储介质可以是ROM、随机存取存储器(RAM)、CD-ROM、磁带、软盘和光数据存储设备等。
本领域普通技术人员可以理解实现上述实施例方法中的全部或部分流程,是可以通过计算机程序来指令相关的硬件来完成,该计算机程序可存储于一存储介质中,该程序在执行时,可包括如上述各方法的实施例的流程。其中,前述的存储介质可为磁碟、光盘、只读存储记忆体(Read-Only Memory,ROM)等非易失性存储介质,或随机存储记忆体(Random Access Memory,RAM)等。
综合上述实施例可知,本申请最大的有益效果在于:
本申请提供的人脸识别方法、装置、终端及存储介质,通过获取待识别人脸图像,并根据预设的深度学习模型提取待识别人脸图像的第一高维特征向量;然后利用乘积量化算法将所述第一高维特征向量量化为数量较少的第一低维整数编码;再将待识别人脸图像的第一低维整数编码与预先得到的人脸模板图像的第二低维整数编码进行相似度匹配;最后根据相似度匹配结果在所述第二低维整数编码中确定与所述第一低维整数编码最相似的目标低维整数编码,将所述目标低维整数编码对应的人脸模板图像作为目标人脸图像,在识别过程中,由于经过乘积量化算法处理后,可将数量庞大的待识别人脸图像和人脸模板图像对应的高维特征向量量化为数量较少的低维整数编码,利用低维整数编码进行相似度比较,从而减少两者之间的比对次数,大大加快了系统的响应速度,提升人脸识别效率。
以上所述实施例的各技术特征可以进行任意的组合,为使描述简洁,未对上述实施例中的各个技术特征所有可能的组合都进行描述,然而,只要这些技术特征的组合不存在矛盾,都应当认为是本说明书记载的范围。
以上所述实施例仅表达了本申请的几种实施方式,其描述较为具体和详细,但并不能因此而理解为对本申请专利范围的限制。应当指出的是,对于本领域的普通技术人员来说,在不脱离本申请构思的前提下,还可以做出若 干变形和改进,这些都属于本申请的保护范围。因此,本申请专利的保护范围应以所附权利要求为准。
Claims (20)
- 一种人脸识别方法,包括:获取待识别人脸图像,根据预设的深度学习模型提取所述待识别人脸图像的第一高维特征向量;利用乘积量化算法将所述第一高维特征向量量化为第一低维整数编码;将待识别人脸图像的第一低维整数编码与预先得到的人脸模板图像的第二低维整数编码进行相似度匹配;根据相似度匹配结果在所述第二低维整数编码中确定与所述第一低维整数编码最相似的目标低维整数编码,将所述目标低维整数编码对应的人脸模板图像作为目标人脸图像。
- 根据权利要求1所述的人脸识别方法,所述根据预设的深度学习模型提取所述待识别人脸图像的第一高维特征向量之前,还包括:利用预先得到的人脸模板图像对所述深度学习模型进行训练,以确定所述深度学习模型中各节点之间的连接的最佳权重参数。
- 根据权利要求2所述的人脸识别方法,所述利用预先得到的人脸模板图像对所述深度学习模型进行训练的步骤,包括:将人脸模板图像输入深度学习模型中,获取所述人脸模板图像对应的特征信息;基于预设的损失函数和所述人脸模板图像对应的特征信息,计算所述深度学习模型的损失;其中,所述损失函数由Softmax函数和L-Softmax函数加权构成;当损失高于预设值时,调整所述深度学习模型中各节点之间的连接的权重参数,对深度学习模型重新训练,直至损失低于或等于预设值时,得到最佳权重参数及其对应的深度学习模型。
- 根据权利要求1所述的人脸识别方法,所述将待识别人脸图像的第一低 维整数编码与预先得到的人脸模板图像的第二低维整数编码进行相似度匹配之前,还包括:获取人脸模板图像,根据预设的深度学习模型提取所述人脸模板图像的第二高维特征向量;利用乘积量化算法将所述第二高维特征向量量化为第二低维整数编码。
- 根据权利要求5所述的人脸识别方法,所述将所述目标低维整数编码对应的人脸模板图像作为目标人脸图像的步骤,包括:在所述第二高维特征向量中获取所述目标低维整数编码对应量化前的目标高维特征向量;计算待识别人脸图像的第一高维特征向量与目标高维特征向量的距离;当所述距离大于预设值时,将所述目标高维特征向量对应的人脸模板图像作为目标人脸图像。
- 根据权利要求6所述的人脸识别方法,所述计算待识别人脸图像的第一高维特征向量与目标高维特征向量的距离的步骤,包括:计算待识别人脸图像的第一高维特征向量与目标高维特征向量的欧氏距离和余弦距离;将所述欧氏距离和余弦距离进行加权组合,得到所述距离。
- 一种人脸识别装置,包括:提取模块,用于获取待识别人脸图像,根据预设的深度学习模型提取所述待识别人脸图像的第一高维特征向量;量化模块,用于利用乘积量化算法将所述第一高维特征向量量化为第一低维整数编码;匹配模块,用于将待识别人脸图像的第一低维整数编码与预先得到的人脸模板图像的第二低维整数编码进行相似度匹配;确定模块,用于根据相似度匹配结果在所述第二低维整数编码中确定与所述第一低维整数编码最相似的目标低维整数编码,将所述目标低维整数编码对应的人脸模板图像作为目标人脸图像。
- 一种终端,包括存储器和处理器,所述存储器中存储有计算机可读指令,所述计算机可读指令被所述处理器执行时,使得所述处理器执行一种人脸识别方法的步骤;其中,所述人脸识别方法包括以下步骤:获取待识别人脸图像,根据预设的深度学习模型提取所述待识别人脸图像的第一高维特征向量;利用乘积量化算法将所述第一高维特征向量量化为第一低维整数编码;将待识别人脸图像的第一低维整数编码与预先得到的人脸模板图像的第二低维整数编码进行相似度匹配;根据相似度匹配结果在所述第二低维整数编码中确定与所述第一低维整数编码最相似的目标低维整数编码,将所述目标低维整数编码对应的人脸模板图像作为目标人脸图像。
- 根据权利要求9所述的终端,所述根据预设的深度学习模型提取所述待识别人脸图像的第一高维特征向量之前,还包括:利用预先得到的人脸模板图像对所述深度学习模型进行训练,以确定所述深度学习模型中各节点之间的连接的最佳权重参数。
- 根据权利要求10所述的终端,所述利用预先得到的人脸模板图像对所述深度学习模型进行训练的步骤,包括:将人脸模板图像输入深度学习模型中,获取所述人脸模板图像对应的特征信息;基于预设的损失函数和所述人脸模板图像对应的特征信息,计算所述深度学习模型的损失;其中,所述损失函数由Softmax函数和L-Softmax函数加权构成;当损失高于预设值时,调整所述深度学习模型中各节点之间的连接的权重参数,对深度学习模型重新训练,直至损失低于或等于预设值时,得到最佳权重参数及其对应的深度学习模型。
- 根据权利要求9所述的终端,所述将待识别人脸图像的第一低维整数编码与预先得到的人脸模板图像的第二低维整数编码进行相似度匹配之前,还包括:获取人脸模板图像,根据预设的深度学习模型提取所述人脸模板图像的第二高维特征向量;利用乘积量化算法将所述第二高维特征向量量化为第二低维整数编码。
- 根据权利要求13所述的终端,所述将所述目标低维整数编码对应的人脸模板图像作为目标人脸图像的步骤,包括:在所述第二高维特征向量中获取所述目标低维整数编码对应量化前的目标高维特征向量;计算待识别人脸图像的第一高维特征向量与目标高维特征向量的距离;当所述距离大于预设值时,将所述目标高维特征向量对应的人脸模板图像作为目标人脸图像。
- 根据权利要求14所述的终端,所述计算待识别人脸图像的第一高维特征向量与目标高维特征向量的距离的步骤,包括:计算待识别人脸图像的第一高维特征向量与目标高维特征向量的欧氏距离和余弦距离;将所述欧氏距离和余弦距离进行加权组合,得到所述距离。
- 一种非易失性存储介质,其上存储有计算机程序,所述计算机程序被处理器执行时,实现一种人脸识别方法;其中,所述人脸识别方法包括以下步骤:获取待识别人脸图像,根据预设的深度学习模型提取所述待识别人脸图像的第一高维特征向量;利用乘积量化算法将所述第一高维特征向量量化为第一低维整数编码;将待识别人脸图像的第一低维整数编码与预先得到的人脸模板图像的第二低维整数编码进行相似度匹配;根据相似度匹配结果在所述第二低维整数编码中确定与所述第一低维整数编码最相似的目标低维整数编码,将所述目标低维整数编码对应的人脸模板图像作为目标人脸图像。
- 根据权利要求16所述的非易失性存储介质,所述根据预设的深度学习模型提取所述待识别人脸图像的第一高维特征向量之前,还包括:利用预先得到的人脸模板图像对所述深度学习模型进行训练,以确定所述深度学习模型中各节点之间的连接的最佳权重参数。
- 根据权利要求17所述的非易失性存储介质,所述利用预先得到的人脸模板图像对所述深度学习模型进行训练的步骤,包括:将人脸模板图像输入深度学习模型中,获取所述人脸模板图像对应的特征信息;基于预设的损失函数和所述人脸模板图像对应的特征信息,计算所述深度学习模型的损失;其中,所述损失函数由Softmax函数和L-Softmax函数加权构成;当损失高于预设值时,调整所述深度学习模型中各节点之间的连接的权重参数,对深度学习模型重新训练,直至损失低于或等于预设值时,得到最佳权重参数及其对应的深度学习模型。
- 根据权利要求16所述的非易失性存储介质,所述将待识别人脸图像的第一低维整数编码与预先得到的人脸模板图像的第二低维整数编码进行相似度匹配之前,还包括:获取人脸模板图像,根据预设的深度学习模型提取所述人脸模板图像的第二高维特征向量;利用乘积量化算法将所述第二高维特征向量量化为第二低维整数编码。
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910457256.8 | 2019-05-29 | ||
CN201910457256.8A CN110298249A (zh) | 2019-05-29 | 2019-05-29 | 人脸识别方法、装置、终端及存储介质 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2020237887A1 true WO2020237887A1 (zh) | 2020-12-03 |
Family
ID=68027274
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/CN2019/103645 WO2020237887A1 (zh) | 2019-05-29 | 2019-08-30 | 人脸识别方法、装置、终端及存储介质 |
Country Status (2)
Country | Link |
---|---|
CN (1) | CN110298249A (zh) |
WO (1) | WO2020237887A1 (zh) |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112488062A (zh) * | 2020-12-18 | 2021-03-12 | 浙江大华技术股份有限公司 | 一种图像识别方法、装置、设备及介质 |
CN112668632A (zh) * | 2020-12-25 | 2021-04-16 | 浙江大华技术股份有限公司 | 一种数据处理方法、装置、计算机设备及存储介质 |
CN113408348A (zh) * | 2021-05-14 | 2021-09-17 | 桂林电子科技大学 | 一种基于视频的人脸识别方法、装置及存储介质 |
CN113743176A (zh) * | 2021-01-29 | 2021-12-03 | 北京沃东天骏信息技术有限公司 | 一种图像识别方法、设备和计算机可读存储介质 |
CN113868291A (zh) * | 2021-10-21 | 2021-12-31 | 深圳云天励飞技术股份有限公司 | 一种最近邻搜索方法、装置、终端和存储介质 |
CN114267065A (zh) * | 2021-12-23 | 2022-04-01 | 广州津虹网络传媒有限公司 | 人脸关键点修正方法及其装置、设备、介质 |
CN114842544A (zh) * | 2022-07-04 | 2022-08-02 | 江苏布罗信息技术有限公司 | 一种适用于面瘫患者的智能化人脸识别方法和系统 |
Families Citing this family (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111091080A (zh) * | 2019-12-06 | 2020-05-01 | 贵州电网有限责任公司 | 人脸识别方法及系统 |
CN111198963A (zh) * | 2019-12-11 | 2020-05-26 | 智慧眼科技股份有限公司 | 基于平均特征的目标检索方法、装置及其相关设备 |
CN111222485A (zh) * | 2020-01-15 | 2020-06-02 | 北京明略软件系统有限公司 | 3d人脸识别方法、装置、电子设备和存储介质 |
CN111368644B (zh) * | 2020-02-14 | 2024-01-05 | 深圳市商汤科技有限公司 | 图像处理方法、装置、电子设备及存储介质 |
CN111325156B (zh) * | 2020-02-24 | 2023-08-11 | 北京沃东天骏信息技术有限公司 | 人脸识别方法、装置、设备和存储介质 |
CN111968152B (zh) * | 2020-07-15 | 2023-10-17 | 桂林远望智能通信科技有限公司 | 一种动态身份识别方法及装置 |
CN112560635B (zh) * | 2020-12-10 | 2024-03-26 | 深圳云天励飞技术股份有限公司 | 人脸匹配加速方法、装置、电子设备及存储介质 |
CN113221820B (zh) * | 2021-05-28 | 2022-07-19 | 杭州网易智企科技有限公司 | 一种对象识别方法、装置、设备及介质 |
CN113807164A (zh) * | 2021-07-29 | 2021-12-17 | 四川天翼网络服务有限公司 | 一种基于余弦损失函数的人脸识别方法 |
CN113743239A (zh) * | 2021-08-12 | 2021-12-03 | 青岛图灵科技有限公司 | 行人重识别方法、装置及电子设备 |
CN113496222B (zh) * | 2021-09-09 | 2021-12-07 | 南方电网数字电网研究院有限公司 | 基于无标签图表数据的模式识别方法、装置和计算机设备 |
CN113971805A (zh) * | 2021-12-22 | 2022-01-25 | 深圳市迪博企业风险管理技术有限公司 | 一种结合机器视觉和语义分析的智能阅卷评分方法 |
CN117558057B (zh) * | 2024-01-12 | 2024-04-16 | 清华大学深圳国际研究生院 | 一种人脸识别方法 |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20120213419A1 (en) * | 2011-02-22 | 2012-08-23 | Postech Academy-Industry Foundation | Pattern recognition method and apparatus using local binary pattern codes, and recording medium thereof |
CN103150561A (zh) * | 2013-03-19 | 2013-06-12 | 华为技术有限公司 | 人脸识别方法和设备 |
CN109426781A (zh) * | 2017-08-29 | 2019-03-05 | 阿里巴巴集团控股有限公司 | 人脸识别数据库的构建方法、人脸识别方法、装置及设备 |
Family Cites Families (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US10120879B2 (en) * | 2013-11-29 | 2018-11-06 | Canon Kabushiki Kaisha | Scalable attribute-driven image retrieval and re-ranking |
CN103745235B (zh) * | 2013-12-18 | 2017-07-04 | 小米科技有限责任公司 | 人脸识别方法、装置及终端设备 |
EP3371712A1 (en) * | 2015-11-06 | 2018-09-12 | Thomson Licensing | Method and apparatus for generating codebooks for efficient search |
CN105956518A (zh) * | 2016-04-21 | 2016-09-21 | 腾讯科技(深圳)有限公司 | 一种人脸识别方法、装置和系统 |
CN109190505A (zh) * | 2018-08-11 | 2019-01-11 | 石修英 | 基于视觉理解的图像识别方法 |
-
2019
- 2019-05-29 CN CN201910457256.8A patent/CN110298249A/zh active Pending
- 2019-08-30 WO PCT/CN2019/103645 patent/WO2020237887A1/zh active Application Filing
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20120213419A1 (en) * | 2011-02-22 | 2012-08-23 | Postech Academy-Industry Foundation | Pattern recognition method and apparatus using local binary pattern codes, and recording medium thereof |
CN103150561A (zh) * | 2013-03-19 | 2013-06-12 | 华为技术有限公司 | 人脸识别方法和设备 |
CN109426781A (zh) * | 2017-08-29 | 2019-03-05 | 阿里巴巴集团控股有限公司 | 人脸识别数据库的构建方法、人脸识别方法、装置及设备 |
Non-Patent Citations (2)
Title |
---|
TAO, JIN ET AL.: "Approximate Nearest Neighbor Search Based on Product Quantization", JOURNAL OF COMPUTER APPLICATIONS, no. 38S2, 25 December 2018 (2018-12-25), ISSN: 1001-9081, DOI: 20200219112012A * |
WANG, DAYONG ET AL.: "Face Search at Scale: 80 Million Gallery", MSU TECHNICAL REPORT MSU-CSE-15-11, ARXIV:1507.07242V2 [CS.CV], 28 July 2015 (2015-07-28), XP055757741, DOI: 20200219114915A * |
Cited By (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112488062A (zh) * | 2020-12-18 | 2021-03-12 | 浙江大华技术股份有限公司 | 一种图像识别方法、装置、设备及介质 |
CN112488062B (zh) * | 2020-12-18 | 2023-04-07 | 浙江大华技术股份有限公司 | 一种图像识别方法、装置、设备及介质 |
CN112668632A (zh) * | 2020-12-25 | 2021-04-16 | 浙江大华技术股份有限公司 | 一种数据处理方法、装置、计算机设备及存储介质 |
CN112668632B (zh) * | 2020-12-25 | 2022-04-08 | 浙江大华技术股份有限公司 | 一种数据处理方法、装置、计算机设备及存储介质 |
CN113743176A (zh) * | 2021-01-29 | 2021-12-03 | 北京沃东天骏信息技术有限公司 | 一种图像识别方法、设备和计算机可读存储介质 |
CN113408348A (zh) * | 2021-05-14 | 2021-09-17 | 桂林电子科技大学 | 一种基于视频的人脸识别方法、装置及存储介质 |
CN113868291A (zh) * | 2021-10-21 | 2021-12-31 | 深圳云天励飞技术股份有限公司 | 一种最近邻搜索方法、装置、终端和存储介质 |
CN114267065A (zh) * | 2021-12-23 | 2022-04-01 | 广州津虹网络传媒有限公司 | 人脸关键点修正方法及其装置、设备、介质 |
CN114842544A (zh) * | 2022-07-04 | 2022-08-02 | 江苏布罗信息技术有限公司 | 一种适用于面瘫患者的智能化人脸识别方法和系统 |
CN114842544B (zh) * | 2022-07-04 | 2022-09-06 | 江苏布罗信息技术有限公司 | 一种适用于面瘫患者的智能化人脸识别方法和系统 |
Also Published As
Publication number | Publication date |
---|---|
CN110298249A (zh) | 2019-10-01 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2020237887A1 (zh) | 人脸识别方法、装置、终端及存储介质 | |
CN109117803B (zh) | 人脸图像的聚类方法、装置、服务器及存储介质 | |
Wang et al. | Robust subspace clustering for multi-view data by exploiting correlation consensus | |
WO2019119505A1 (zh) | 人脸识别的方法和装置、计算机装置及存储介质 | |
WO2020143321A1 (zh) | 一种基于变分自编码器的训练样本数据扩充方法、存储介质及计算机设备 | |
US9274607B2 (en) | Authenticating a user using hand gesture | |
US10192098B2 (en) | Palm print image matching techniques | |
Xie et al. | m-SNE: Multiview stochastic neighbor embedding | |
US11080316B1 (en) | Context-inclusive face clustering | |
CN108875487B (zh) | 行人重识别网络的训练及基于其的行人重识别 | |
CN112784778B (zh) | 生成模型并识别年龄和性别的方法、装置、设备和介质 | |
KR101912748B1 (ko) | 확장성을 고려한 특징 기술자 생성 및 특징 기술자를 이용한 정합 장치 및 방법 | |
CN110414550B (zh) | 人脸识别模型的训练方法、装置、系统和计算机可读介质 | |
WO2014082496A1 (zh) | 客户端特征的识别方法、装置及存储介质 | |
US11874866B2 (en) | Multiscale quantization for fast similarity search | |
WO2022213857A1 (zh) | 动作识别方法和装置 | |
WO2020134819A1 (zh) | 一种搜索人脸的方法及相关装置 | |
US11809990B2 (en) | Method apparatus and system for generating a neural network and storage medium storing instructions | |
Boufounos et al. | Representation and coding of signal geometry | |
JP2021168195A (ja) | 顔画像の分類方法、装置、電子機器及び記憶媒体 | |
Giansiracusa et al. | Persistent homology machine learning for fingerprint classification | |
CN112668482A (zh) | 人脸识别训练方法、装置、计算机设备及存储介质 | |
CN117854156B (zh) | 一种特征提取模型的训练方法和相关装置 | |
WO2020199498A1 (zh) | 指静脉比对方法、装置、计算机设备及存储介质 | |
WO2020258507A1 (zh) | X光片的分类方法、装置、终端及存储介质 |
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: 19931457 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: 19931457 Country of ref document: EP Kind code of ref document: A1 |