WO2020098225A1 - 关键点检测方法及装置、电子设备和存储介质 - Google Patents

关键点检测方法及装置、电子设备和存储介质 Download PDF

Info

Publication number
WO2020098225A1
WO2020098225A1 PCT/CN2019/083721 CN2019083721W WO2020098225A1 WO 2020098225 A1 WO2020098225 A1 WO 2020098225A1 CN 2019083721 W CN2019083721 W CN 2019083721W WO 2020098225 A1 WO2020098225 A1 WO 2020098225A1
Authority
WO
WIPO (PCT)
Prior art keywords
feature
feature map
processing
map
maps
Prior art date
Application number
PCT/CN2019/083721
Other languages
English (en)
French (fr)
Inventor
杨昆霖
田茂清
伊帅
Original Assignee
北京市商汤科技开发有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 北京市商汤科技开发有限公司 filed Critical 北京市商汤科技开发有限公司
Priority to KR1020207012580A priority Critical patent/KR102394354B1/ko
Priority to JP2020518758A priority patent/JP6944051B2/ja
Priority to SG11202003818YA priority patent/SG11202003818YA/en
Priority to US16/855,630 priority patent/US20200250462A1/en
Publication of WO2020098225A1 publication Critical patent/WO2020098225A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/21Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/24Classification techniques
    • G06F18/241Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2413Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/25Fusion techniques
    • G06F18/253Fusion techniques of extracted features
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/045Combinations of networks
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00Arrangements for image or video recognition or understanding
    • G06V10/70Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/764Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00Arrangements for image or video recognition or understanding
    • G06V10/70Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/80Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level
    • G06V10/806Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level of extracted features
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00Arrangements for image or video recognition or understanding
    • G06V10/70Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V2201/00Indexing scheme relating to image or video recognition or understanding
    • G06V2201/03Recognition of patterns in medical or anatomical images
    • G06V2201/033Recognition of patterns in medical or anatomical images of skeletal patterns
    • YGENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02TCLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T10/00Road transport of goods or passengers
    • Y02T10/10Internal combustion engine [ICE] based vehicles
    • Y02T10/40Engine management systems

Definitions

  • the present disclosure relates to the field of computer vision technology, and in particular, to a key point detection method and device, electronic equipment, and storage medium.
  • Human key point detection is to detect the position information of key points such as joints or facial features from the human body image, so as to describe the posture of the human body through the position information of these key points.
  • the existing technology can usually use a neural network to obtain the multi-scale features of the image, and finally used to predict the position of the key point of the human body.
  • we found that using this method we cannot fully mine and use multi-scale features, and the detection accuracy of key points is low.
  • Embodiments of the present disclosure provide an effective key point detection method and device, electronic equipment, and storage medium that improve key point detection accuracy.
  • a key point detection method which includes:
  • first feature maps for multiple scales of the input image and the scale of each first feature map has a multiple relationship; use the first pyramid neural network to perform a forward processing on each of the first feature maps to obtain A second feature map with one-to-one correspondence of feature maps, wherein the second feature map has the same scale as the first feature map with one-to-one correspondence; a second pyramid neural network is used to invert each of the second feature maps Processing to obtain a third feature map corresponding to each of the second feature maps, wherein the third feature map and the one-to-one corresponding second feature map have the same scale; for each of the third features The map performs feature fusion processing, and uses the feature map after feature fusion processing to obtain the positions of key points in the input image.
  • the obtaining the first feature map of multiple scales for the input image includes: adjusting the input image to a first image of a preset specification; and inputting the first image to a residual
  • the neural network performs downsampling processing at different sampling frequencies on the first image to obtain multiple first feature maps with different scales.
  • the forward processing includes first convolution processing and first linear interpolation processing
  • the reverse processing includes second convolution processing and second linear interpolation processing
  • the forward processing of each of the first feature maps by using the first pyramid neural network to obtain a second feature map corresponding to each of the first feature maps includes: wherein a first check convolution FIG C 1 ... C n first feature of C n in FIG performs convolution processing, wherein FIG obtain a second first feature corresponding to FIG C n F. n, where n represents a first FIG characteristic number, and n is an integer greater than 1; the second feature map F n performs linear interpolation processing to obtain the second feature map F n corresponding to the first intermediate characteristic graph F 'n, wherein the first intermediate feature FIG. F 'n is the same scale dimensions of the first feature of the FIG.
  • the second feature map F i is composed of the second intermediate feature map C ′ i and the first An intermediate feature map F ′ i + 1 is obtained by performing superposition processing.
  • the first intermediate feature map F ′ i is obtained by linear interpolation from the corresponding second feature map F i
  • the second intermediate feature map C ′ i is The scale of an intermediate feature map F ′ i + 1 is the same, where i is an integer greater than or equal to 1 and less than n.
  • the inverse processing of each of the second feature maps by using the second pyramid neural network to obtain a third feature map corresponding to each of the second feature maps includes: wherein the second check convolutional FIG three F 1 ... F. m second feature map F 1 of the convoluting process, obtaining a second characteristic diagram of FIG third feature F 1 corresponding to R 1, where m represents a second The number of feature maps, and m is an integer greater than 1; the fourth convolution kernel is used to convolve the second feature maps F 2 ... F m to obtain the corresponding third intermediate feature maps F ′′ 2 ... F ′′ m , where the scale of the third intermediate feature map is the same as the scale of the corresponding second feature map;
  • the performing feature fusion processing on each of the third feature maps, and using the feature maps after feature fusion processing to obtain the position of each key point in the input image includes: Perform feature fusion processing on the three feature maps to obtain a fourth feature map: obtain the positions of key points in the input image based on the fourth feature map.
  • performing feature fusion processing on each third feature map to obtain a fourth feature map includes: using linear interpolation to adjust each third feature map to a feature map with the same scale; The feature maps with the same scale are connected to obtain the fourth feature map.
  • the method before performing feature fusion processing on each third feature map to obtain a fourth feature map, the method further includes: inputting the first group of third feature maps into different bottleneck block structures respectively Perform convolution processing to obtain an updated third feature map, each of the bottleneck block structures includes a different number of convolution modules, wherein the third feature map includes a first set of third feature maps and a second A set of third feature maps, each of the first set of third feature maps and the second set of third feature maps includes at least one third feature map.
  • performing feature fusion processing on each third feature map to obtain a fourth feature map includes: using linear interpolation, the updated third feature map and the third feature map Two sets of third feature maps are adjusted to feature maps with the same scale; the feature maps with the same scale are connected to obtain the fourth feature map.
  • the obtaining the position of each key point in the input image based on the fourth feature map includes: performing a dimensionality reduction process on the fourth feature map using a fifth convolution kernel; The fourth feature map after the dimension processing determines the positions of the key points of the input image.
  • the obtaining the position of each key point in the input image based on the fourth feature map includes: performing a dimensionality reduction process on the fourth feature map using a fifth convolution kernel; using a volume
  • the block attention module performs purification processing on the features in the fourth feature map after dimensionality reduction processing to obtain a purified feature map; and uses the purified feature map to determine the position of the key point of the input image.
  • the method further includes training the first pyramid neural network using a training image data set, which includes: using the first pyramid neural network to perform a first feature corresponding to each image in the training image data set The image is subjected to the forward processing to obtain a second feature map corresponding to each image in the training image data set; each second feature map is used to determine the identified key points; and the first loss of the key points is obtained according to the first loss function ; Using the first loss to reversely adjust each convolution kernel in the first pyramid neural network until the number of trainings reaches the set first number threshold.
  • the method further includes training the second pyramid neural network using a training image data set, which includes using the second pyramid neural network to output training image data output from the first pyramid neural network to the first pyramid neural network Perform the reverse processing on the second feature map corresponding to each image to obtain a third feature map corresponding to each image in the training image data set; use each third feature map to determine the identified key points; obtain according to the second loss function The second loss of each identified key point; use the second loss to reversely adjust the convolution kernel in the second pyramid neural network until the number of trainings reaches the set second number threshold; or, use the second The loss reversely adjusts the convolution kernel in the first pyramid network and the convolution kernel in the second pyramid neural network until the number of trainings reaches the set second number threshold.
  • performing the feature fusion processing on each of the third feature maps through a feature extraction network and performing the feature fusion processing on each of the third feature maps through a feature extraction network
  • the method further included: training the feature extraction network using the training image data set, which includes using a feature extraction network to output a third feature map corresponding to each image in the training image data set from the second pyramid neural network Perform the feature fusion process, and use the feature map after the feature fusion process to identify the key points of each image in the training image data set; obtain the third loss of each key point according to the third loss function; use the third loss value Reversely adjust the parameters of the feature extraction network until the number of training times reaches the set third-time threshold; or, use the third loss function to reversely adjust the convolution kernel parameters and the first in the first pyramid neural network The convolution kernel parameters in the two-pyramid neural network and the parameters of the feature extraction network until the training times reach the set third times threshold.
  • a key point detection apparatus which includes: a multi-scale feature acquisition module configured to obtain first feature maps for multiple scales of an input image, each of the first feature maps The scale is in a multiple relationship; the forward processing module is configured to forward process each of the first feature maps using the first pyramid neural network to obtain a second feature map corresponding to each of the first feature maps, wherein, The second feature map has the same scale as the first feature map in one-to-one correspondence; the inverse processing module is configured to perform inverse processing on each of the second feature maps using the second pyramid neural network to obtain A third feature map corresponding to the second feature map in one-to-one correspondence, wherein the third feature map has the same scale as the second feature map in one-to-one correspondence; a key point detection module is configured for each third The feature map is subjected to feature fusion processing, and the feature map after feature fusion processing is used to obtain the position of each key point in the input image.
  • the multi-scale feature acquisition module is configured to adjust the input image to a first image of a preset specification, and input the first image to a residual neural network.
  • the image is down-sampled at different sampling frequencies to obtain multiple first feature maps of different scales.
  • the forward processing includes first convolution processing and first linear interpolation processing
  • the reverse processing includes second convolution processing and second linear interpolation processing
  • the forward processing module is configured to perform a convolution process on the first feature map C n in the first feature map C 1 ... C n using the first convolution kernel to obtain FIG feature a second feature corresponding to C n F n in FIG, where n represents the number of a first characteristic diagram, and n is an integer greater than 1; and F n performs linear interpolation processing on the second characteristic diagram obtained with a second The first intermediate feature map F ′ n corresponding to the feature map F n , wherein the scale of the first intermediate feature map F ′ n is the same as the scale of the first feature map C n-1 ; and the second feature map is used to check the first feature map FIG respective first feature other than C n C 1 ...
  • C n- 1 performs convolution processing, respectively, to obtain C 1 ... C n-1-one correspondence of the second intermediate first feature characteristic diagram C of FIG. ' 1 ... C ' n-1 , wherein the scale of the second intermediate feature map is the same as the scale of the first feature map corresponding to it; and based on the second feature map F n and each of the first Two intermediate feature maps C ' 1 ... C' n-1 to obtain a second feature map F 1 ... F n-1 and a first intermediate feature map F ′ 1 ...
  • the second feature map F i is obtained by superimposing the second intermediate feature map C ′ i and the first intermediate feature map F ′ i + 1 , and the first intermediate feature map F ′ i is formed by the corresponding second feature
  • the graph F i is obtained by linear interpolation, and the second intermediate feature map C ′ i has the same scale as the first intermediate feature map F ′ i + 1 , where i is an integer greater than or equal to 1 and less than n.
  • the reverse processing module is configured to perform a convolution process on the second feature map F 1 in the second feature maps F 1 ... F m using a third convolution kernel to obtain 1 corresponding to the third characteristic feature of Figure II in FIG. F R 1, wherein m represents the number of the second characteristic diagram, and m is an integer greater than 1; and using a second feature matching fourth convolution F 2 ... F m FIG. Perform a convolution process to obtain the corresponding third intermediate feature maps F ′′ 2 ...
  • FIG third collation performed convolution processing to obtain the third characteristic corresponds to FIG fourth intermediate R 1 wherein FIG R '1; and FIG characterized by each of the third intermediate F "2 ... F" m and a fourth Intermediate feature map R ' 1 to obtain a third feature map R 2 ... R m and a fourth intermediate feature map R' 2 ...
  • the key point detection module is configured to perform feature fusion processing on each third feature map to obtain a fourth feature map, and obtain each key in the input image based on the fourth feature map The location of the point.
  • the key point detection module is configured to use linear interpolation to adjust each third feature map to a feature map with the same scale, and connect the feature maps with the same scale to obtain The fourth characteristic diagram is described.
  • the device further includes: an optimization module configured to input the first set of third feature maps to different bottleneck block structures for convolution processing to obtain updated third features, respectively Figures, each of the bottleneck block structures includes a different number of convolution modules, wherein the third feature map includes a first set of third feature maps and a second set of third feature maps, the first set of third Both the feature map and the second set of third feature maps include at least one third feature map.
  • an optimization module configured to input the first set of third feature maps to different bottleneck block structures for convolution processing to obtain updated third features, respectively Figures, each of the bottleneck block structures includes a different number of convolution modules, wherein the third feature map includes a first set of third feature maps and a second set of third feature maps, the first set of third Both the feature map and the second set of third feature maps include at least one third feature map.
  • the keypoint detection module is further configured to adjust each of the updated third feature map and the second set of third feature maps to the same scale using linear interpolation Feature map, and connect the feature maps with the same scale to obtain the fourth feature map.
  • the key point detection module is further configured to perform dimensionality reduction processing on the fourth feature map using a fifth convolution kernel, and determine the key of the input image using the fourth feature map after the dimensionality reduction processing The location of the point.
  • the keypoint detection module is further configured to perform a dimensionality reduction process on the fourth feature map using a fifth convolution kernel, and use a convolutional block attention module to perform the dimensionality reduction on the fourth feature
  • the features in the figure are purified to obtain a purified feature map, and the purified feature map is used to determine the position of the key point of the input image.
  • the forward processing module is further configured to train the first pyramid neural network using a training image data set, which includes: using the first pyramid neural network to correspond to each image in the training image data set The first feature map of the is subjected to the forward processing to obtain a second feature map corresponding to each image in the training image data set; the second feature map is used to determine the identified key points; the key points are obtained according to the first loss function The first loss; using the first loss to reversely adjust each convolution kernel in the first pyramid neural network until the training times reach the set first times threshold.
  • the reverse processing module is further configured to train the second pyramid neural network using a training image data set, which includes: using the second pyramid neural network to output the first pyramid neural network Perform the reverse processing on the second feature map corresponding to each image in the training image data set to obtain a third feature map corresponding to each image in the training image data set; use each third feature map to determine the identified key points;
  • the second loss function obtains the second loss of each identified key point; the second loss is used to reversely adjust the convolution kernel in the second pyramid neural network until the number of trainings reaches the set second number threshold; or, use The second loss reversely adjusts the convolution kernel in the first pyramid network and the convolution kernel in the second pyramid neural network until the number of training times reaches the set second number threshold.
  • the key point detection module is further configured to perform the feature fusion process on each of the third feature maps through a feature extraction network, and execute the Before performing feature fusion processing on the third feature map, the training image data set is used to train the feature extraction network, which includes: using the feature extraction network to output the second pyramid neural network with respect to each image corresponding to each image in the training image data set.
  • the three feature maps perform the feature fusion processing, and use the feature maps after feature fusion processing to identify the key points of each image in the training image data set; obtain the third loss of each key point according to the third loss function;
  • the three loss values reversely adjust the parameters of the feature extraction network until the training times reach the set third times threshold; or, use the third loss function to reversely adjust the convolution kernel in the first pyramid neural network
  • an electronic device including: a processor; a memory for storing processor executable instructions; wherein the processor is configured to: execute any of the first aspect One of the methods.
  • a computer-readable storage medium having computer program instructions stored thereon, the computer program instructions implementing the method of any one of the first aspects when executed by a processor .
  • An embodiment of the present disclosure proposes to use a bidirectional pyramid neural network to perform key point feature detection, in which not only the forward processing is used to obtain multi-scale features, but also the reverse processing is used to fuse more features, which can further improve the key Point detection accuracy.
  • FIG. 1 shows a flowchart of a key point detection method according to an embodiment of the present disclosure
  • step S100 in a key point detection method according to an embodiment of the present disclosure
  • FIG. 3 shows another flowchart of a key point detection method according to an embodiment of the present disclosure
  • step S200 shows a flowchart of step S200 in a key point detection method according to an embodiment of the present disclosure
  • step S300 shows a flowchart of step S300 in the key point detection method according to an embodiment of the present disclosure
  • step S400 is a flowchart of step S400 in the key point detection method according to an embodiment of the present disclosure
  • step S401 shows a flowchart of step S401 in the key point detection method according to an embodiment of the present disclosure
  • FIG. 8 shows another flowchart of a key point detection method according to an embodiment of the present disclosure
  • step S402 shows a flowchart of step S402 in the key point detection method according to an embodiment of the present disclosure
  • FIG. 10 shows a flowchart of training a first pyramid neural network in a keypoint detection method according to an embodiment of the present disclosure
  • FIG. 11 shows a flowchart of training a second pyramid neural network in a keypoint detection method according to an embodiment of the present disclosure
  • FIG. 12 shows a flowchart of a training feature extraction network model in a keypoint detection method according to an embodiment of the present disclosure
  • FIG. 13 shows a block diagram of a key point detection device according to an embodiment of the present disclosure
  • FIG. 14 shows a block diagram of an electronic device 800 according to an embodiment of the present disclosure
  • FIG. 15 shows a block diagram of an electronic device 1900 according to an embodiment of the present disclosure.
  • An embodiment of the present disclosure provides a key point detection method, which can be used to perform key point detection of a human image, which utilizes two pyramid network models to respectively perform forward processing and reverse processing of multi-scale features of key points ,
  • the fusion of more feature information can improve the accuracy of key point position detection.
  • FIG. 1 shows a flowchart of a key point detection method according to an embodiment of the present disclosure.
  • the key point detection method of the embodiments of the present disclosure may include:
  • the embodiments of the present disclosure use the fusion of multi-scale features of the input image to perform the above-mentioned key point detection.
  • the first feature maps in multiple scales of the input image can be obtained, the scales of the first feature maps are different, and there is a multiple relationship between the scales.
  • Embodiments of the present disclosure may use a multi-scale analysis algorithm to obtain multiple scale first feature maps of the input image, or may also obtain multiple scale first feature maps of the input image through a neural network model that can perform multi-scale analysis.
  • the disclosed embodiments are not specifically limited.
  • S200 Perform a forward processing on each of the first feature maps by using a first pyramid neural network to obtain a second feature map corresponding to each of the first feature maps, wherein the second feature map corresponds to each of them
  • the scale of the first feature map is the same.
  • the forward processing may include a first convolution process and a first linear interpolation process.
  • a second feature with the same scale as the corresponding first feature map may be obtained
  • each second feature map further integrates the features of the input image, and the obtained second feature map has the same number as the first feature map, and the second feature map has the same scale as the corresponding first feature map.
  • the first feature map obtained by the embodiment of the present disclosure may be C 1 , C 2 , C 3 and C 4
  • the corresponding second feature map obtained after the forward processing may be F 1 , F 2 , F 3 and F 4 .
  • S300 Perform a reverse processing on each second feature map by using a second pyramid neural network to obtain a third feature map corresponding to each of the second feature maps.
  • the reverse processing includes a second convolution process, where, The third feature map has the same scale as the second feature map in one-to-one correspondence.
  • the reverse processing includes second convolution processing and second linear interpolation processing.
  • a third feature map with the same scale as the corresponding second feature map can be obtained , And each third feature map further integrates the features of the input image relative to the second feature map, and the number of the obtained third feature map and the second feature map is the same, and the third feature map and the corresponding second feature map The scale is the same.
  • the second feature map obtained by the embodiment of the present disclosure may be F 1 , F 2 , F 3 and F 4
  • the corresponding third feature map obtained after the reverse processing may be R 1 , R 2 , R 3 and R 4 .
  • S400 Perform feature fusion processing on each of the third feature maps, and obtain the positions of key points in the input image by using the feature maps after feature fusion processing.
  • the features of each third feature map can be executed Fusion processing.
  • the embodiments of the present disclosure may use the corresponding convolution process to realize feature fusion of each third feature map, and when the scale of the third feature map is different, the scale conversion may be performed, and then the feature map stitching may be performed, and Extraction of key points.
  • the embodiments of the present disclosure can perform the detection of different key points of the input image, for example, when the input image is an image of a person, the key points may be left and right eyes, nose, right and left ears, left and right shoulders, right and left elbows, right and left wrists, right and left crotch , At least one of left and right knees, right and left ankles, or in other embodiments, the input image may also be other types of images, and other key points may be identified when performing key point detection. Therefore, the embodiment of the present disclosure may further perform key point detection and identification according to the feature fusion result of the third feature map.
  • the embodiments of the present disclosure can perform forward processing and further reverse processing based on the first feature map through bidirectional pyramid neural networks (first pyramid neural network and second pyramid neural network), respectively, which can effectively improve the input image
  • the degree of feature fusion further improves the detection accuracy of key points.
  • the embodiments of the present disclosure may first obtain an input image, which may be any image type, for example, a person image, a landscape image, an animal image, and so on. For different types of images, different key points can be identified. For example, the embodiment of the present disclosure takes a person image as an example for description.
  • the first feature map of the input image at multiple different scales can be obtained through step S100.
  • FIG. 2 shows a flowchart of step S100 in a key point detection method according to an embodiment of the present disclosure.
  • obtaining first feature maps of different scales for the input image may include:
  • the embodiment of the present disclosure may first normalize the size specifications of the input image, that is, the input image may be first adjusted to a first image of a preset specification, where the preset specification in the embodiment of the present disclosure may be 256pix * 192pix, and pix is a pixel value
  • the input image may be uniformly converted into images of other specifications, which is not specifically limited in the embodiments of the present disclosure.
  • S102 Input the first image to a residual neural network, and perform downsampling processing with different sampling frequencies on the first image to obtain first feature maps with different scales.
  • a sampling process of multiple sampling frequencies may be performed on the first image.
  • the first feature maps of different scales for the first image can be obtained through the residual neural network processing.
  • the first image can be sampled by using different sampling frequencies to obtain first feature maps of different scales.
  • the sampling frequency of the embodiment of the present disclosure may be 1/8, 1/16, 1/32, etc., but the embodiment of the present disclosure does not limit this.
  • the feature map in the embodiment of the present disclosure refers to the feature matrix of the image, for example, the feature matrix of the embodiment of the present disclosure may be a three-dimensional matrix, and the length and width of the feature map described in the embodiment of the present disclosure may be corresponding to The dimension of the feature matrix in the row and column directions.
  • FIG. 3 shows another flowchart of a key point detection method according to an embodiment of the present disclosure.
  • part (a) shows the process of step S100 in the embodiment of the present disclosure, and four first feature maps C 1 , C 2 , C 3 and C 4 can be obtained through step S100, wherein the first feature map C 1
  • the length and width may correspond to twice the length and width of the first feature map C 2, respectively
  • the length and width of the second feature map C 2 may correspond to twice the length and width of the third feature map C 3 , respectively
  • the length and width of the third feature map C 3 may correspond to twice the length and width of the fourth feature map C 4 , respectively.
  • the scale multiples between C 1 and C 2, between C 2 and C 3 , and between C 3 and C 4 may be the same, for example, k 1 takes the value 1.
  • k 1 may have different values, for example, the length and width of the first feature map C 1 may correspond to twice the length and width of the first feature map C 2 , respectively.
  • the length and width FIGS C 2 may correspond to the third feature, respectively length and width quadruple FIGS C 3, and a third length and width characteristics FIGS C 3 may correspond respectively to the fourth feature of the C 4 of FIG. Eight times the length and width, but this embodiment of the present disclosure does not limit this.
  • the forward processing of the first feature map may be performed through step S200 to obtain a plurality of second feature maps of different scales fused with the features of each first feature map .
  • FIG. 4 shows a flowchart of step S200 in a key point detection method according to an embodiment of the present disclosure.
  • using the first pyramid neural network to perform forward processing on each of the first feature maps to obtain a second feature map corresponding to each of the first feature maps includes:
  • S201 a first collation using the first convolution wherein FIG C 1 ... C n in FIG characteristic C n for the first convolution processing to obtain the first feature a second feature map corresponding to FIG C n F n, wherein , Where n represents the number of first feature maps, and n is an integer greater than 1, and the length and width of the first feature map C n correspond to the length and width of the second feature map F n , respectively.
  • the forward processing performed by the first pyramid neural network in the embodiment of the present disclosure may include first convolution processing and first linear interpolation processing, and may also include other processing procedures, which are not limited in the embodiment of the present disclosure.
  • the first feature map obtained in the embodiment of the present disclosure may be C 1 ... C n , that is, n first feature maps, and C n may be a feature map with the smallest length and width, That is, the first feature map with the smallest scale.
  • a first pyramid can be used wherein a first neural network convolution process for FIG C n, i.e., check the first convolution using a first characteristic diagram for convolution processing C n, to obtain a second characteristic graph F n.
  • the second feature map F n are the length and width are the same as the length and width of the first feature of C n in FIG.
  • the first convolution kernel may be a 3 * 3 convolution kernel, or may be another type of convolution kernel.
  • the second feature map F n performs linear interpolation processing to obtain a first and a second intermediate feature map F n corresponding to FIG feature F 'n, wherein the first intermediate feature map F' n first feature map scale The scale of C n-1 is the same.
  • the second feature map F n may be used to obtain the corresponding first intermediate feature map F ′ n , and the embodiment of the present disclosure may be obtained by performing linear interpolation processing on the second feature map F n
  • the first intermediate feature map F ′ n corresponding to the second feature map F n wherein the scale of the first intermediate feature map F ′ n is the same as the scale of the first feature map C n-1 , for example, at C n-1
  • the scale of the first intermediate feature map F ′ n is twice the length of the second feature map F n
  • the width of the first intermediate feature map F ′ n is the second feature Figure Fn is twice the width.
  • S203 Perform a convolution process on each of the first feature maps C 1 ... C n-1 except the first feature map C n by using the second convolution kernel to obtain each first feature other than the first feature map C n
  • Figures C 1 ... C n-1 correspond to the second intermediate feature map C ' 1 ... C' n-1
  • the scale of the second intermediate feature map corresponds to the first one-to-one correspondence
  • the scale of the feature map is the same.
  • the second convolution kernel can be used to perform a second convolution process on the first feature maps C 1 ... C n-1 , respectively, to obtain one -to- one with each of the first feature maps C 1 ... C n-1
  • the second convolution kernel may be a 1 * 1 convolution kernel, but this disclosure does not specifically limit it.
  • each second intermediate feature map obtained by the second convolution process is the same as the scale of the corresponding first feature map.
  • the present embodiment of the present disclosure may be characterized in a first descending FIG C 1 ... C n-1 to obtain each of the first feature FIG C 1 ... C n-1 wherein a second intermediate FIG C '1 .. .C ' n-1 . That is, it is possible to obtain the first characteristic corresponds to FIG. C n-1, the second intermediate FIG C 'n-1, then to obtain a first characteristic corresponding to FIG C n-2 second intermediate FIG C' n-2, in order to by analogy, until a first characteristic diagram corresponding to a second intermediate C 1 characterized in FIG C '1.
  • C n-1 is composed of the second intermediate feature map C ′ i and the first intermediate feature map F ′ I + 1 is obtained by superposition processing (addition processing), and the first intermediate feature map F ′ i is obtained by linear interpolation from the corresponding second feature map F i , and the second intermediate feature map C ′ i is The scale of the intermediate characteristic map F ′ i + 1 is the same, where i is an integer greater than or equal to 1 and less than n.
  • the second feature map F i other than the second feature map F n may still be obtained by using a reverse processing method. That is, the present embodiment of the present disclosure may first obtain the first intermediate characteristic graph F n-1, wherein the first feature may be utilized FIGS C n-1 corresponding to the second intermediate FIG C 'n-1 and the first intermediate feature map F' n performs superposition processing to obtain a second feature map F n-1 , wherein the length and width of the second intermediate feature map C ′ n-1 are the same as the length and width of the first intermediate feature map F ′ n , respectively, and the second feature The length and width of the graph F n-1 are the length and width of the second intermediate characteristic graphs C ′ n-1 and F ′ n .
  • a second characteristic graph F n-1 are the length and width of the second feature F in FIG twice the length and width of n (C n-1 C n scale of a scale of twice).
  • the second feature map F n-1 can be linearly interpolated to obtain the first intermediate feature map F ′ n-1 , so that the scale of F ′ n-1 is the same as the scale of C n-1 , which can then be used the first characteristic graph C n-2 corresponding to FIG second intermediate C 'n-2 and the first intermediate feature map F' n-1 to obtain the second superposing characteristic graph F n-2, wherein the second intermediate feature FIG.
  • the length and width of C ' n-2 are the same as those of the first intermediate feature map F' n-1
  • the length and width of the second feature map F n-2 are the second intermediate feature map C ' n- 2 and the length and width of F ′ n-1
  • the second feature for example, the length and width FIGS F n-2, respectively, twice a second characteristic diagram F n-1 in length and width.
  • the first intermediate feature map F ′ 2 can be finally obtained, and the length of the second feature map F 1 , F 1 can be obtained according to the superposition process of the first intermediate feature map F ′ 2 and the first feature map C ′ 1 and width with the length and width C 1 of the same.
  • step S200 may use a first pyramid neural network (Feature Pyramid Network-FPN) to obtain a multi-scale second feature map.
  • first C 4 may be calculated through a convolution core 3 * 3 to obtain a new feature F in FIG. 4 (a second characteristic diagram), the same length and width F 4 and C 4 size.
  • the up-sampling operation of double linear interpolation is performed on F 4 to obtain a feature map whose length and width are enlarged twice, that is, the first intermediate feature map F ′ 4 .
  • C 3 is calculated by a 1 * 1 second convolution kernel to obtain a second intermediate feature map C ′ 3 , C ′ 3 and F ′ 4 are the same size, and the two feature maps are added to obtain a new feature map F 3 ( FIG second feature), characterized in that the second length and a width F of FIG. 3 are the second feature F 4 twice FIG.
  • C 2 is calculated by a 1 * 1 second convolution kernel to obtain a second intermediate feature map C ′ 2 , C ′ 2 and F ′ 3 are the same size, and the two feature maps are added to obtain a new feature map F 2 ( FIG second feature), characterized in that the second length and a width F of FIG. 2, respectively, twice a second feature F 3 in FIG.
  • the up-sampling operation of bilinear interpolation is performed on F 2 to obtain a feature map whose length and width are enlarged twice, that is, the first intermediate feature map F ′ 2 .
  • C 1 is calculated by a 1 * 1 second convolution kernel to obtain a second intermediate feature map C ′ 1 , C ′ 1 and F ′ 2 are the same size, and the two feature maps are added to obtain a new feature map F 2 ( FIG second feature), characterized in that the second panel F, respectively length and width of a second F 2 characterized twice FIG.
  • FPN four second feature maps of different scales were also obtained, which were denoted as F 1 , F 2 , F 3 and F 4 .
  • the multiples of the length and width between F 1 and F 2 are the same as the multiples of the length and width between C 1 and C 2
  • the multiples of the length and width between F 2 and F 3 are the same as C 2 and C 3
  • the multiples of the length and width are the same
  • the multiples of the length and width between F 3 and F 4 are the same as the multiples of the length and width between C 3 and C 4 .
  • each second feature map performs reverse processing.
  • the reverse processing may include second convolution processing and second linear interpolation processing.
  • other processing may also be included, which is not specifically limited in the embodiment of the present disclosure.
  • FIG. 5 shows a flowchart of step S300 in the key point detection method according to an embodiment of the present disclosure.
  • using the second pyramid neural network to perform reverse processing on each second feature map to obtain a third feature map R i of different scales may include:
  • S301 third convolution using a second feature matching F in FIG. 1 ... F m F 1 of the convoluting process, wherein FIG obtain a second view corresponding to the third feature F 1 R 1, wherein the third feature of FIG.
  • the length and width of R 1 correspond to the length and width of the first feature map C 1 , respectively, where m represents the number of second feature maps, and m is an integer greater than 1, in which case m and the number of first feature maps n the same.
  • inverse processing can be performed first from the second feature map F 1 with the largest length and width.
  • the second feature map F 1 can be convoluted by a third convolution kernel to obtain the length are the same width F, and third intermediate wherein FIG 1 R 1.
  • the third convolution kernel may be a 3 * 3 convolution kernel or other types of convolution kernels. The technical field in the art may select a desired convolution kernel according to different requirements.
  • S302 Perform a convolution process on the second feature maps F 2 ... F m using a fourth convolution kernel to obtain corresponding third intermediate feature maps F ′′ 2 ... F ′′ m , where the third intermediate feature map
  • the scale of is the same as the scale of the corresponding second feature map.
  • the fourth convolution may be utilized to check each of the second feature of the second feature F in FIG other than FIG. 1
  • F 2 ... F m are the convolution process, wherein the third intermediate to give the corresponding Figure F ′′ 1 ... F ′′ m-1 .
  • the second feature maps F 2 ...
  • each third intermediate feature map F ′′ j may be the length and width of the corresponding second feature map F j .
  • the fourth convolution may be utilized to check each of the second feature of the second feature F in FIG other than FIG. 1
  • F 2 ... F m are the convolution process, wherein the third intermediate to give the corresponding Figure F ′′ 1 ... F ′′ m-1 .
  • the second feature maps F 2 ...
  • F m other than the second feature map F 1 can be convolved through a fourth convolution kernel, where F 2 can be first convolved to obtain the corresponding first Three intermediate feature maps F ′′ 2 , and then F 3 can be convolved to obtain a corresponding third intermediate feature map F ′′ 3 , and so on, to obtain a third intermediate feature map F ′′ n corresponding to the second feature map F m
  • the length and width of each third intermediate feature map F ′′ j may be half of the length and width of the corresponding second feature map F j .
  • the collation can also use the third volume V of FIG wherein R 1 convoluting process to obtain a third characteristic diagram corresponding to R 1 in FIG fourth intermediate wherein R '1.
  • the length and width of the fourth intermediate feature map R ′ 1 are the length and width of the second feature map F 2 .
  • step S302 of FIG. F in FIG fourth intermediate wherein R '1 i and obtained in step S303, to obtain the third characteristic feature of FIG third FIG other than R 1 R 2 ... R m.
  • each R in FIG third characteristic than the third feature of FIG 1 R 2 ... R m FIG characterized by a third intermediate F "j with the FIG fourth intermediate wherein R 'j-1 in superimposition processing obtained.
  • step S304 respectively, may be utilized wherein the third intermediate view corresponding to F "i and FIG fourth intermediate wherein R 'i-1 for each of the third superposition processing to obtain the third characteristic feature map R of FIG. 1 other than R j .
  • the third feature map R 2 can be obtained first by using the addition result of the third intermediate feature map F ′′ 2 and the fourth intermediate feature map R ′ 1 .
  • a fifth convolution kernel is used to convolve R 2 to obtain a fourth intermediate feature map R ′ 2
  • the sum of the results between the third intermediate feature map F ′′ 3 and the fourth intermediate feature map R ′ 2 is used to obtain the third Three feature maps R 3.
  • the remaining fourth intermediate feature maps R ′ 3 ... R ′ m and the third feature map R 4 ... R m can be further obtained.
  • each fourth intermediate feature map R ′ 1 obtained are the same as the length and width of the second feature map F 2 , respectively.
  • the length and width of the fourth intermediate feature map R ′ j are the same as the length and width of the fourth intermediate feature map F ′′ j + 1 respectively.
  • the length and width of the obtained third feature map R j are the second features
  • the length and width of the graph F i , and the lengths and widths of the further third feature maps R 1 ... Rn respectively correspond to those of the first feature maps C 1 ... C n .
  • the second feature pyramid network (Reverse Feature Pyramid Network-RFPN) is then used to further optimize the multi-scale features.
  • FIG via a second feature F a 3 * 3 convolution kernel (third convolution kernel), to give a new characteristic diagram R 1 (FIG third feature), R 1 length and width the same as the size of the F 1.
  • the feature map R 1 undergoes a convolution kernel with a convolution kernel of 3 * 3 (fifth convolution kernel) and a stride of 2 to calculate a new feature map, which is denoted as R ′ 1 and the length of R ′ 1 Both the width and the width can be half of R 1 .
  • the second feature graph F 2 is calculated by a 3 * 3 convolution kernel (fourth convolution kernel) to obtain a new feature graph, which is denoted as F ′′ 2.
  • R ′ 1 and F ′′ 2 are the same size, and R ′ 1 and F ′′ 2 are added to obtain a new feature map R 2.
  • RFPN four different scale feature maps are also obtained, which are denoted as R 1 , R 2 , R 3 and R 4.
  • R 1 and R 2 The multiples of the length and width between C 1 and C 2 are the same, and the multiples of the length and width between R 2 and R 3 are the same as the length and width between R 2 and R 3 The multiples are the same, and the multiples of the length and width between R 3 and R 4 are the same as the multiples of the length and width between C 3 and C 4 .
  • the third feature map R 1 ... Rn obtained by the reverse processing of the second fundraising network model can be obtained.
  • the forward and reverse processing can further improve the characteristics of image fusion, based on Each third feature map can accurately identify feature points.
  • step S300 the result may be fused according to the third characteristic feature of each of R i in FIG., The position of the key points of the input image.
  • step S400 performs feature fusion processing on each of the third feature maps and obtaining the position of each key point in the input image using the feature map after feature fusion processing (step S400) may include:
  • S401 Perform feature fusion processing on each third feature map to obtain a fourth feature map.
  • feature fusion may be performed on each third feature map, because the length and width of each third feature map in the embodiment of the present disclosure Different, so R 2 ... R n can be linearly interpolated, and finally the length and width of each third feature map R 2 ... R n are the same as the length and width of the third feature map R 1 . Then, the processed third feature maps can be combined to form a fourth feature map.
  • S402 Obtain the positions of key points in the input image based on the fourth feature map.
  • the fourth feature map may be subjected to dimensionality reduction processing, for example, the fourth feature map may be subjected to dimensionality reduction through convolution processing, and the feature point after the dimensionality reduction may be used to identify the position of the feature point of the input image .
  • step S401 shows a flowchart of step S401 in the key point detection method according to an embodiment of the present disclosure, wherein the feature fusion processing performed on each third feature map to obtain a fourth feature map (step S401) may include:
  • S4012 Use linear interpolation to adjust each third feature map to feature maps with the same scale.
  • the scales of the third feature maps R 1 ... R n obtained by the embodiments of the present disclosure are different, it is first necessary to adjust the third feature maps to the feature maps of the same scale.
  • the three feature maps perform different linear interpolation processes so that the scale of each feature map is the same, wherein the multiple of linear interpolation may be related to the scale multiple between each third feature map.
  • the feature maps can be stitched and combined to obtain the fourth feature map.
  • the feature maps after interpolation processing of the embodiments of the present disclosure have the same length and width, and the feature maps can be Connect in the height direction to obtain the fourth feature map.
  • each feature map after S4012 processing can be represented as A, B, C, and D, and the obtained fourth feature map can be:
  • step S401 in order to optimize small-scale features, the embodiment of the present disclosure may further optimize the third feature map with a smaller length and width, and may perform further convolution processing on the partial features.
  • FIG. 8 shows another flowchart of a key point detection method according to an embodiment of the present disclosure, where, before performing feature fusion processing on each third feature map to obtain a fourth feature map, S4011 may also be included:
  • S4011 Input the first set of third feature maps into different bottleneck block structures for convolution processing, respectively corresponding to the updated third feature maps, each of the bottleneck block structures includes a different number of volumes Product module; wherein, the third feature map includes a first set of third feature maps and a second set of third feature maps, both the first set of third feature maps and the second set of third feature maps include At least one third feature map.
  • the convolution processing can be further characterized in view of the small scale, wherein the third feature map may be R 1 ... R m is divided into two groups, a first group of the third feature
  • the scale of the map is smaller than the scale of the second set of third feature maps.
  • each third feature map in the first group of third feature maps can be input into different bottleneck block structures to obtain an updated third feature map.
  • the bottleneck block structure can include at least one volume
  • the number of convolution modules in different bottleneck block structures may be different.
  • the size of the feature map obtained after the convolution processing of the bottleneck block structure is the same as the size of the third feature map before input.
  • the first group of third feature maps may be determined according to a preset ratio value of the number of third feature maps.
  • the preset ratio can be 50%, that is, the third feature map with the smaller half of each third feature map can be input as the first set of third feature maps to different bottleneck block structures for feature optimization processing .
  • the preset ratio may also be other ratio values, which is not limited in this disclosure.
  • the first set of third feature maps input to the bottleneck block structure may also be determined according to the scale threshold.
  • the feature map smaller than the threshold value of the scale determines that it needs to be input into the bottleneck block structure for feature optimization processing.
  • the determination of the scale threshold may be determined according to the scale of each feature map, which is not specifically limited in the embodiments of the present disclosure.
  • bottleneck block structure is not specifically limited in the embodiments of the present disclosure, and the form of the convolution module can be selected according to requirements.
  • the optimized first set of third feature maps and the second set of third features may be scale normalized, that is, the feature maps are adjusted to feature maps of the same size.
  • corresponding linear interpolation processing is performed for the third feature map optimized for each S4011 and the second group of third feature maps, respectively, so as to obtain feature maps of the same size.
  • R 2 , R 3 and R 4 are followed by different numbers of bottleneck block structures, After R 2 is connected a bottleneck block, a new feature map is denoted as R ′′ 2 , after R 3 is connected with two bottleneck blocks, a new feature map is denoted as R ′′ 3 , followed by R 4 After three bottleneck blocks, the new feature map has to be rewritten, which is denoted as R ′′ 4.
  • step S4012 feature maps with the same scale can be connected, for example, the above four feature maps are concatenated to obtain a new feature map, which is the fourth feature map, for example, R 1 , R ′′ ′ 2 , R ′′ ′ 3
  • the four feature maps of R ′′ ′ 4 are all 256 dimensions, and the obtained fourth feature map can be 1024 dimensions.
  • the corresponding fourth feature map can be obtained through the configuration in the above different embodiments.
  • the key point position of the input image can be obtained according to the fourth feature map.
  • the fourth feature map may be directly subjected to dimensionality reduction processing, and the position of the key point of the input image may be determined using the dimensionality reduction processed feature map.
  • the feature map after dimensionality reduction may also be purified to further improve the accuracy of key points.
  • 9 shows a flowchart of step S402 in a key point detection method according to an embodiment of the present disclosure.
  • the obtaining the position of each key point in the input image based on the fourth feature map may include:
  • S4021 Perform dimensionality reduction processing on the fourth feature map using a fifth convolution kernel.
  • the manner of performing the dimensionality reduction processing may be convolution processing, that is, the preset feature convolution module is used to perform convolution processing on the fourth feature map to achieve the dimensionality reduction of the fourth feature map, for example, 256 Feature map.
  • S4022 Use the convolution block attention module to perform purification processing on the features in the fourth feature map after the dimensionality reduction process, to obtain a purified feature map.
  • the convolutional block attention module can be further used to purify the fourth feature map after the dimensionality reduction process.
  • the convolutional block attention module may be a convolutional block attention module in the prior art.
  • the convolutional block attention module of the embodiment of the present disclosure may include a channel attention unit and an importance attention unit.
  • the fourth feature map after dimensionality reduction processing can be first input to the channel attention unit, wherein firstly, the fourth feature map after dimensionality reduction processing can be subjected to global maximum pooling based on height and width, and Global average pooling, and then input the first result obtained by the global maximum pooling and the second result obtained by the global average pooling to the multi-layer perceptron (MLP), and the MLP processed The two results are summed to obtain a third result, and the third result is activated to obtain a channel attention feature map.
  • MLP multi-layer perceptron
  • the channel attention feature map is input to the importance attention unit.
  • the channel attention feature map can be input to the channel-based global maximum pooling (global maxpooling) and the global average Globalization (pooling) processing to obtain the fourth result and the fifth result respectively, and then connect the fourth result and the fifth result, and then perform dimensionality reduction on the connected result through convolution processing, and use the sigmoid function to reduce the
  • the dimension result is processed to obtain the importance attention feature map, and then the importance attention feature map is multiplied by the channel attention feature map to obtain the purified feature map.
  • the convolutional block attention module in the embodiment of the present disclosure.
  • other structures may also be used to purify the fourth feature map after dimensionality reduction.
  • S4023 Determine the position of key points of the input image using the purified feature map.
  • the feature map can be used to obtain the position information of key points, for example, the purified feature map can be input to a 3 * 3 convolution module to predict the position information of each key point in the input image .
  • the predicted key points may be the positions of 17 key points, for example, may include left and right eyes, nose, right and left ears, left and right shoulders, left and right elbows, left and right wrists, left and right crotch, left and right knee 3.
  • the position of the left and right ankles may also be obtained, which is not limited in the embodiments of the present disclosure.
  • FIG. 10 shows a flowchart of training a first pyramid neural network in a keypoint detection method according to an embodiment of the present disclosure.
  • the embodiment of the present disclosure may use the training image data set to train the first pyramid neural network, which includes:
  • S501 Perform the forward processing on the first feature map corresponding to each image in the training image data set using a first pyramid neural network to obtain a second feature map corresponding to each image in the training image data set.
  • the training image data set may be input to the first pyramid neural network for training.
  • the training image data set may include multiple images and the actual positions of key points corresponding to the images.
  • steps S100 and S200 extraction of multi-scale first feature map and forward processing
  • steps S100 and S200 extraction of multi-scale first feature map and forward processing
  • S502 Use each second feature map to determine the identified key points.
  • the obtained second feature map may be used to identify key points of the training image to obtain the first position of each key point of the training image.
  • S504 Use the first loss value to reversely adjust each convolution kernel in the first pyramid neural network until the training times reach the set first time threshold.
  • the first loss corresponding to the predicted first position can be obtained.
  • the parameters of the first pyramid neural network can be reversely adjusted according to the first loss obtained in each training, such as the parameters of the convolution kernel, until the number of trainings reaches the first number threshold, the first number threshold It can be set according to requirements, and is generally a value greater than 120.
  • the threshold of the first number of times in the embodiment of the present disclosure may be 140.
  • the first loss corresponding to the first position may be a loss value obtained by inputting the first difference between the first position and the real position into the first loss function, where the first loss function may be a logarithmic loss function.
  • the first position and the real position may be input to the first loss function to obtain the corresponding first loss.
  • the embodiments of the present disclosure do not limit this. Based on the above, the training process of the first pyramid neural network can be realized, and the parameters of the first pyramid neural network can be optimized.
  • FIG. 11 shows a flowchart of training a second pyramid neural network in a keypoint detection method according to an embodiment of the present disclosure.
  • the embodiment of the present disclosure may use the training image data set to train the second pyramid neural network, which includes:
  • S601 Use the second pyramid neural network to perform the reverse processing on the second feature map output by the first pyramid neural network and corresponding to each image in the training image data set, to obtain the first corresponding to each image in the training image data set.
  • the first pyramid neural network may be used first to obtain the second feature map of each image in the training data set, and then the second feature map corresponding to each image in the training image data set may be processed through the second pyramid neural network.
  • the third feature map corresponding to each image in the training image data set and then use the third feature map to predict the second position of the key point of the corresponding image.
  • S604 Use the second loss to reverse adjust the convolution kernel in the second pyramid neural network until the number of training times reaches the set second number threshold, or use the second loss to reverse adjust the first pyramid The convolution kernel in the network and the convolution kernel in the second pyramid neural network until the training times reach the set second times threshold.
  • the second loss corresponding to the predicted second position can be obtained after the second position of each key point is obtained.
  • the parameters of the second pyramid neural network can be reversely adjusted according to the second loss obtained in each training, such as the parameters of the convolution kernel, until the number of trainings reaches the second number threshold, the second number threshold can be based on
  • the requirement is set, generally a value greater than 120, for example, the threshold of the second number of times in the embodiment of the present disclosure may be 140.
  • the second loss corresponding to the second position may be a loss value obtained by inputting the second difference between the second position and the real position into the second loss function, where the second loss function may be a logarithmic loss function.
  • the second position and the real position may be input to the second loss function to obtain the corresponding second loss value.
  • the embodiments of the present disclosure do not limit this.
  • the first pyramid neural network while training the second pyramid neural network, can be further optimized and trained simultaneously. That is, in the embodiment of the present disclosure, in step S604, the obtained second loss can be used The value simultaneously reverses the parameters of the convolution kernel in the first pyramid neural network and the convolution kernel parameters in the second pyramid neural network sink. In order to achieve further optimization of the entire network model.
  • the training process of the second pyramid neural network can be realized, and the optimization of the first pyramid neural network can be realized.
  • step S400 may be implemented by a feature extraction network model, wherein the embodiment of the present disclosure may also perform an optimization process of the feature extraction network model, where FIG. 12 shows a first embodiment of the present disclosure.
  • Flowchart of a training feature extraction network model in a key point detection method, wherein training the feature extraction network model using a training image data set may include:
  • S701 Use the feature extraction network model to perform the feature fusion process on the third feature map corresponding to each image in the training image data set output by the second pyramid neural network, and use the feature map after feature fusion processing to identify the training The key points of each image in the image data set.
  • the third feature map obtained by the first pyramid neural network forward processing and the second pyramid neural network processing corresponding to the image training data set may be input to the feature extraction network model, and the feature extraction network The model performs feature fusion, purification and other processing to obtain the third position of the key point of each image in the training image data set.
  • S703 Use the third loss value to reverse adjust the parameters of the feature extraction network until the number of training times reaches the set third time threshold, or use the third loss function to reverse adjust the first pyramid neural network
  • the third loss value corresponding to the predicted third position can be obtained.
  • the parameters of the network model can be extracted based on the third loss reverse adjustment feature obtained in each training, such as the parameters of the convolution kernel, or the parameters of the above pooling process, until the number of training reaches the third number
  • the threshold, the third times threshold may be set according to requirements, and is generally a value greater than 120.
  • the third times threshold may be 140 in the embodiment of the present disclosure.
  • the third loss corresponding to the third position may be a loss value obtained by inputting the third difference between the third position and the real position into the first loss function, where the third loss function may be a logarithmic loss function.
  • the third position and the real position may be input to the third loss function to obtain the corresponding third loss value.
  • the embodiments of the present disclosure do not limit this.
  • the training process of the feature extraction network model can be realized, and the parameter optimization of the feature extraction network model can be realized.
  • the first pyramid neural network and the second pyramid neural network can be further optimized and trained simultaneously, that is, in the embodiment of the present disclosure, in step S703, the obtained At the same time, the third loss value reversely adjusts the parameters of the convolution kernel in the first pyramid neural network, the convolution kernel parameters in the second pyramid neural network sink, and the parameters of the feature extraction network model, thereby realizing the further network model. optimization.
  • the embodiment of the present disclosure proposes to use a bidirectional pyramid network model to perform key point feature detection, in which not only multi-scale features are obtained by forward processing, but also more features are fused by reverse processing. This can further improve the detection accuracy of key points.
  • the present disclosure also provides a key point detection device, an electronic device, a computer-readable storage medium, and a program, all of which can be used to implement any key point detection method provided by the present disclosure.
  • a key point detection device an electronic device, a computer-readable storage medium, and a program, all of which can be used to implement any key point detection method provided by the present disclosure.
  • FIG. 13 shows a block diagram of a key point detection device according to an embodiment of the present disclosure.
  • the key point detection device includes:
  • the multi-scale feature acquisition module 10 is configured to obtain a first feature map for multiple scales of the input image, and the scale of each first feature map is in a multiple relationship;
  • the forward processing module 20 is configured to use the first pyramid neural network for each Performing forward processing on the first feature map to obtain a second feature map corresponding to each of the first feature maps, wherein the second feature map has the same scale as the first feature map corresponding to each one
  • the reverse processing module 30 is configured to perform a reverse processing on each of the second feature maps using a second pyramid neural network to obtain a third feature map corresponding to each of the second feature maps, wherein, the first The three feature maps have the same scale as the one-to-one corresponding second feature map;
  • the key point detection module 40 is configured to perform feature fusion processing on each of the third feature maps, and use the feature fusion processed feature maps to obtain Describe the position of each key point in the input image.
  • the multi-scale feature acquisition module is configured to adjust the input image to a first image of a preset specification, and input the first image to a residual neural network.
  • the image is down-sampled at different sampling frequencies to obtain multiple first feature maps of different scales.
  • the forward processing includes first convolution processing and first linear interpolation processing
  • the reverse processing includes second convolution processing and second linear interpolation processing
  • the forward processing module is configured to perform a convolution process on the first feature map C n in the first feature map C 1 ... C n using the first convolution kernel to obtain FIG feature a second feature corresponding to C n F n in FIG, where n represents the number of a first characteristic diagram, and n is an integer greater than 1; and F n performs linear interpolation processing on the second characteristic diagram obtained with a second The first intermediate feature map F ′ n corresponding to the feature map F n , wherein the scale of the first intermediate feature map F ′ n is the same as the scale of the first feature map C n-1 ; and the second feature map is used to check the first feature map FIG respective first feature other than C n C 1 ...
  • C n- 1 performs convolution processing, respectively, to obtain C 1 ... C n-1-one correspondence of the second intermediate first feature characteristic diagram C of FIG. ' 1 ... C ' n-1 , wherein the scale of the second intermediate feature map is the same as the scale of the first feature map corresponding to it; and based on the second feature map F n and each of the first Two intermediate feature maps C ' 1 ... C' n-1 to obtain a second feature map F 1 ... F n-1 and a first intermediate feature map F ′ 1 ...
  • the second feature map F i is obtained by superimposing the second intermediate feature map C ′ i and the first intermediate feature map F ′ i + 1 , and the first intermediate feature map F ′ i is formed by the corresponding second feature
  • the graph F i is obtained by linear interpolation, and the second intermediate feature map C ′ i has the same scale as the first intermediate feature map F ′ i + 1 , where i is an integer greater than or equal to 1 and less than n.
  • the reverse processing module is configured to perform a convolution process on the second feature map F 1 in the second feature maps F 1 ... F m using a third convolution kernel to obtain 1 corresponding to the third characteristic feature of Figure II in FIG. F R 1, wherein m represents the number of the second characteristic diagram, and m is an integer greater than 1; and using a second feature matching fourth convolution F 2 ... F m FIG. Perform a convolution process to obtain the corresponding third intermediate feature maps F ′′ 2 ...
  • FIG third collation performed convolution processing to obtain the third characteristic corresponds to FIG fourth intermediate R 1 wherein FIG R '1; and FIG characterized by each of the third intermediate F "2 ... F" m and a fourth Intermediate feature map R ' 1 to obtain a third feature map R 2 ... R m and a fourth intermediate feature map R' 2 ...
  • the key point detection module is configured to perform feature fusion processing on each third feature map to obtain a fourth feature map, and obtain each key in the input image based on the fourth feature map The location of the point.
  • the key point detection module is configured to use linear interpolation to adjust each third feature map to a feature map with the same scale, and connect the feature maps with the same scale to obtain The fourth characteristic diagram is described.
  • the device further includes: an optimization module configured to input the first set of third feature maps to different bottleneck block structures for convolution processing to obtain updated third features, respectively Figures, each of the bottleneck block structures includes a different number of convolution modules, wherein the third feature map includes a first set of third feature maps and a second set of third feature maps, the first set of third Both the feature map and the second set of third feature maps include at least one third feature map.
  • an optimization module configured to input the first set of third feature maps to different bottleneck block structures for convolution processing to obtain updated third features, respectively Figures, each of the bottleneck block structures includes a different number of convolution modules, wherein the third feature map includes a first set of third feature maps and a second set of third feature maps, the first set of third Both the feature map and the second set of third feature maps include at least one third feature map.
  • the keypoint detection module is further configured to adjust each of the updated third feature map and the second set of third feature maps to the same scale using linear interpolation Feature map, and connect the feature maps with the same scale to obtain the fourth feature map.
  • the key point detection module is further configured to perform dimensionality reduction processing on the fourth feature map using a fifth convolution kernel, and determine the key of the input image using the fourth feature map after the dimensionality reduction processing The location of the point.
  • the keypoint detection module is further configured to perform a dimensionality reduction process on the fourth feature map using a fifth convolution kernel, and use a convolutional block attention module to perform the dimensionality reduction on the fourth feature
  • the features in the figure are purified to obtain a purified feature map, and the purified feature map is used to determine the position of the key point of the input image.
  • the forward processing module is further configured to train the first pyramid neural network using a training image data set, which includes: using the first pyramid neural network to correspond to each image in the training image data set The first feature map of the is subjected to the forward processing to obtain a second feature map corresponding to each image in the training image data set; the second feature map is used to determine the identified key points; the key points are obtained according to the first loss function The first loss; using the first loss to reversely adjust each convolution kernel in the first pyramid neural network until the training times reach the set first times threshold.
  • the reverse processing module is further configured to train the second pyramid neural network using a training image data set, which includes: using the second pyramid neural network to output the first pyramid neural network Perform the reverse processing on the second feature map corresponding to each image in the training image data set to obtain a third feature map corresponding to each image in the training image data set; use each third feature map to determine the identified key points;
  • the second loss function obtains the second loss of each identified key point; the second loss is used to reversely adjust the convolution kernel in the second pyramid neural network until the number of trainings reaches the set second number threshold; or, use The second loss reversely adjusts the convolution kernel in the first pyramid network and the convolution kernel in the second pyramid neural network until the number of training times reaches the set second number threshold.
  • the key point detection module is further configured to perform the feature fusion process on each of the third feature maps through a feature extraction network, and execute the Before performing feature fusion processing on the third feature map, the training image data set is used to train the feature extraction network, which includes: using the feature extraction network to output the second pyramid neural network with respect to each image corresponding to each image in the training image data set.
  • the three feature maps perform the feature fusion processing, and use the feature maps after feature fusion processing to identify the key points of each image in the training image data set; obtain the third loss of each key point according to the third loss function;
  • the three loss values reversely adjust the parameters of the feature extraction network until the training times reach the set third times threshold; or, use the third loss function to reversely adjust the convolution kernel in the first pyramid neural network
  • the functions provided by the apparatus provided by the embodiments of the present disclosure or the modules contained therein may be used to perform the methods described in the above method embodiments.
  • An embodiment of the present disclosure also proposes a computer-readable storage medium on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the above method is implemented.
  • the computer-readable storage medium may be a non-volatile computer-readable storage medium.
  • An embodiment of the present disclosure also provides an electronic device, including: a processor; a memory for storing processor executable instructions; wherein the processor is configured as the above method.
  • the electronic device may be provided as a terminal, server, or other form of device.
  • FIG. 14 shows a block diagram of an electronic device 800 according to an embodiment of the present disclosure.
  • the electronic device 800 may be a terminal such as a mobile phone, computer, digital broadcasting terminal, messaging device, game console, tablet device, medical device, fitness device, and personal digital assistant.
  • the electronic device 800 may include one or more of the following components: a processing component 802, a memory 804, a power supply component 806, a multimedia component 808, an audio component 810, an input / output (I / O) interface 812, and a sensor component 814 , ⁇ ⁇ ⁇ 816.
  • the processing component 802 generally controls the overall operations of the electronic device 800, such as operations associated with display, phone calls, data communications, camera operations, and recording operations.
  • the processing component 802 may include one or more processors 820 to execute instructions to complete all or part of the steps in the above method.
  • the processing component 802 may include one or more modules to facilitate interaction between the processing component 802 and other components.
  • the processing component 802 may include a multimedia module to facilitate interaction between the multimedia component 808 and the processing component 802.
  • the memory 804 is configured to store various types of data to support operation at the electronic device 800. Examples of these data include instructions for any application or method operating on the electronic device 800, contact data, phone book data, messages, pictures, videos, and so on.
  • the memory 804 may be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read only memory (EEPROM), erasable and removable Programmable read only memory (EPROM), programmable read only memory (PROM), read only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.
  • SRAM static random access memory
  • EEPROM electrically erasable programmable read only memory
  • EPROM erasable and removable Programmable read only memory
  • PROM programmable read only memory
  • ROM read only memory
  • magnetic memory flash memory
  • flash memory magnetic disk or optical disk.
  • the power supply component 806 provides power to various components of the electronic device 800.
  • the power component 806 may include a power management system, one or more power sources, and other components associated with generating, managing, and distributing power for the electronic device 800.
  • the multimedia component 808 includes a screen that provides an output interface between the electronic device 800 and the user.
  • the screen may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen may be implemented as a touch screen to receive input signals from the user.
  • the touch panel includes one or more touch sensors to sense touch, swipe, and gestures on the touch panel. The touch sensor may not only sense the boundary of the touch or sliding action, but also detect the duration and pressure related to the touch or sliding operation.
  • the multimedia component 808 includes a front camera and / or a rear camera. When the electronic device 800 is in an operation mode, such as a shooting mode or a video mode, the front camera and / or the rear camera may receive external multimedia data. Each front camera and rear camera can be a fixed optical lens system or have focal length and optical zoom capabilities.
  • the audio component 810 is configured to output and / or input audio signals.
  • the audio component 810 includes a microphone (MIC).
  • the microphone is configured to receive an external audio signal.
  • the received audio signal may be further stored in the memory 804 or transmitted via the communication component 816.
  • the audio component 810 further includes a speaker for outputting audio signals.
  • the I / O interface 812 provides an interface between the processing component 802 and a peripheral interface module.
  • the peripheral interface module may be a keyboard, a click wheel, or a button. These buttons may include, but are not limited to: home button, volume button, start button, and lock button.
  • the sensor component 814 includes one or more sensors for providing the electronic device 800 with status evaluation in various aspects.
  • the sensor component 814 can detect the on / off state of the electronic device 800, and the relative positioning of the components, for example, the component is the display and keypad of the electronic device 800, and the sensor component 814 can also detect the electronic device 800 or the electronic device 800.
  • the position of the component changes, the presence or absence of user contact with the electronic device 800, the orientation or acceleration / deceleration of the electronic device 800, and the temperature change of the electronic device 800.
  • the sensor assembly 814 may include a proximity sensor configured to detect the presence of nearby objects without any physical contact.
  • the sensor component 814 may also include a light sensor, such as a CMOS or CCD image sensor, for use in imaging applications.
  • the sensor component 814 may further include an acceleration sensor, a gyro sensor, a magnetic sensor, a pressure sensor, or a temperature sensor.
  • the communication component 816 is configured to facilitate wired or wireless communication between the electronic device 800 and other devices.
  • the electronic device 800 can access a wireless network based on a communication standard, such as WiFi, 2G, or 3G, or a combination thereof.
  • the communication component 816 receives a broadcast signal or broadcast related information from an external broadcast management system via a broadcast channel.
  • the communication component 816 also includes a near field communication (NFC) module to facilitate short-range communication.
  • the NFC module can be implemented based on radio frequency identification (RFID) technology, infrared data association (IrDA) technology, ultra-wideband (UWB) technology, Bluetooth (BT) technology, and other technologies.
  • RFID radio frequency identification
  • IrDA infrared data association
  • UWB ultra-wideband
  • Bluetooth Bluetooth
  • the electronic device 800 may be one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field Programming gate array (FPGA), controller, microcontroller, microprocessor or other electronic components are used to implement the above method.
  • ASICs application specific integrated circuits
  • DSPs digital signal processors
  • DSPDs digital signal processing devices
  • PLDs programmable logic devices
  • FPGA field Programming gate array
  • controller microcontroller, microprocessor or other electronic components are used to implement the above method.
  • a non-volatile computer-readable storage medium is also provided, for example, a memory 804 including computer program instructions, which can be executed by the processor 820 of the electronic device 800 to complete the above method.
  • FIG. 15 shows a block diagram of an electronic device 1900 according to an embodiment of the present disclosure.
  • the electronic device 1900 may be provided as a server.
  • the electronic device 1900 includes a processing component 1922, which further includes one or more processors, and memory resources represented by the memory 1932, for storing instructions executable by the processing component 1922, such as application programs.
  • the application programs stored in the memory 1932 may include one or more modules each corresponding to a set of instructions.
  • the processing component 1922 is configured to execute instructions to perform the above method.
  • the electronic device 1900 may also include a power component 1926 configured to perform power management of the electronic device 1900, a wired or wireless network interface 1950 configured to connect the electronic device 1900 to the network, and an input output (I / O) interface 1958 .
  • the electronic device 1900 can operate an operating system based on the memory 1932, such as Windows ServerTM, Mac OS XTM, UnixTM, LinuxTM, FreeBSDTM or the like.
  • a non-volatile computer-readable storage medium is also provided, for example, a memory 1932 including computer program instructions, which can be executed by the processing component 1922 of the electronic device 1900 to complete the above method.
  • the present disclosure may be a system, method, and / or computer program product.
  • the computer program product may include a computer-readable storage medium loaded with computer-readable program instructions for causing the processor to implement various aspects of the present disclosure.
  • the computer-readable storage medium may be a tangible device that can hold and store instructions used by the instruction execution device.
  • the computer-readable storage medium may be, but is not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing.
  • Computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read only memory (ROM), and erasable programmable read only memory (EPROM (Or flash memory), static random access memory (SRAM), portable compact disk read-only memory (CD-ROM), digital versatile disk (DVD), memory stick, floppy disk, mechanical coding device, such as a computer on which instructions are stored
  • RAM random access memory
  • ROM read only memory
  • EPROM erasable programmable read only memory
  • SRAM static random access memory
  • CD-ROM compact disk read-only memory
  • DVD digital versatile disk
  • memory stick floppy disk
  • mechanical coding device such as a computer on which instructions are stored
  • the convex structure in the hole card or the groove and any suitable combination of the above.
  • the computer-readable storage medium used herein is not to be interpreted as a transient signal itself, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (for example, optical pulses through fiber optic cables), or through wires The transmitted electrical signal.
  • the computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to various computing / processing devices, or to an external computer or external storage device through a network, such as the Internet, a local area network, a wide area network, and / or a wireless network.
  • the network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers.
  • the network adapter card or network interface in each computing / processing device receives computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in the computer-readable storage medium in each computing / processing device .
  • Computer program instructions for performing the operations of the present disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-related instructions, microcode, firmware instructions, state setting data, or in one or more programming languages Source code or object code written in any combination.
  • the programming languages include object-oriented programming languages such as Smalltalk, C ++, etc., and conventional procedural programming languages such as "C" language or similar programming languages.
  • Computer readable program instructions can be executed entirely on the user's computer, partly on the user's computer, as an independent software package, partly on the user's computer and partly on a remote computer, or completely on the remote computer or server carried out.
  • the remote computer may be connected to the user's computer through any kind of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (eg, using an Internet service provider to pass the Internet connection).
  • electronic circuits such as programmable logic circuits, field programmable gate arrays (FPGAs) or programmable logic arrays (PLA), can be personalized by utilizing the status information of computer-readable program instructions, which can be Computer-readable program instructions are executed to implement various aspects of the present disclosure.
  • These computer-readable program instructions can be provided to the processor of a general-purpose computer, special-purpose computer, or other programmable data processing device, thereby producing a machine that causes these instructions to be executed by the processor of a computer or other programmable data processing device A device that implements the functions / actions specified in one or more blocks in the flowchart and / or block diagram is generated.
  • the computer-readable program instructions may also be stored in a computer-readable storage medium. These instructions enable the computer, programmable data processing apparatus, and / or other devices to work in a specific manner. Therefore, the computer-readable medium storing the instructions includes An article of manufacture that includes instructions to implement various aspects of the functions / acts specified in one or more blocks in the flowchart and / or block diagram.
  • the computer-readable program instructions can also be loaded onto a computer, other programmable data processing apparatus, or other equipment, so that a series of operating steps are performed on the computer, other programmable data processing apparatus, or other equipment to produce a computer-implemented process , So that the instructions executed on the computer, other programmable data processing device, or other equipment implement the functions / acts specified in one or more blocks in the flowchart and / or block diagram.
  • each block in the flowchart or block diagram may represent a module, program segment, or part of an instruction, and the module, program segment, or part of an instruction contains one or more Executable instructions.
  • the functions marked in the blocks may also occur in an order different from that marked in the drawings. For example, two consecutive blocks can actually be executed substantially in parallel, and sometimes they can also be executed in reverse order, depending on the functions involved.
  • each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts can be implemented with dedicated hardware-based systems that perform specified functions or actions Or, it can be realized by a combination of dedicated hardware and computer instructions.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Data Mining & Analysis (AREA)
  • Computing Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Engineering & Computer Science (AREA)
  • Multimedia (AREA)
  • Medical Informatics (AREA)
  • Databases & Information Systems (AREA)
  • Biomedical Technology (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Molecular Biology (AREA)
  • Mathematical Physics (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Human Computer Interaction (AREA)
  • Image Analysis (AREA)

Abstract

本公开实施例涉及一种关键点检测方法及装置、电子设备和存储介质,所述方法包括:获得针对输入图像的多个尺度的第一特征图,各第一特征图的尺度成倍数关系;利用第一金字塔神经网络对各所述第一特征图进行正向处理得到与各个所述第一特征图一一对应的第二特征图,其中,所述第二特征图与其一一对应的所述第一特征图的尺度相同;利用第二金字塔神经网络对各个所述第二特征图进行反向处理得到与各个所述第二特征图一一对应的第三特征图,其中,所述第三特征图与其一一对应的所述第二特征图的尺度相同;对各所述第三特征图进行特征融合处理,并利用特征融合处理后的特征图获得所述输入图像中的各关键点的位置。本公开能够精确的提取关键点的位置。

Description

关键点检测方法及装置、电子设备和存储介质
相关申请的交叉引用
本申请基于申请号为201811367869.4、申请日为2018年11月16日的中国专利申请提出,并要求该中国专利申请的优先权,该中国专利申请的全部内容在此以引入方式并入本申请。
技术领域
本公开涉及计算机视觉技术领域,特别涉及一种关键点检测方法及装置、电子设备和存储介质。
背景技术
人体关键点检测是从人体图像上检测出关节或者五官等关键点的位置信息,从而通过这些关键点的位置信息来描述人体的姿态。
因为人体在图像中有大有小,现有的技术通常可以采用神经网络来获取图像的多尺度特征,用来最终预测人体关键点的位置。但是我们发现使用这种方式,还不能完全地挖掘和利用多尺度特征,关键点的检测精度较低。
发明内容
本公开实施例提供了一种有效的提高关键点检测精度的关键点检测方法及装置、电子设备和存储介质。
根据本公开实施例的第一方面,提供了一种关键点检测方法,其包括:
获得针对输入图像的多个尺度的第一特征图,各第一特征图的尺度成倍数关系;利用第一金字塔神经网络对各所述第一特征图进行正向处理得到与各个所述第一特征图一一对应的第二特征图,其中,所述第二特征图与其一一对应的所述第一特征图的尺度相同;利用第二金字塔神经网络对各个所述第二特征图进行反向处理得到与各个所述第二特征图一一对应的第三特征图,其中,所述第三特征图与其一一对应的所述第二特征图的尺度相同;对各所述第三特征图进行特征融合处理,并利用特征融合处理后的特征图获得所述输入图像中的各关键点的位置。
在一些可能的实施方式中,所述获得针对输入图像的多个尺度的第一特征图包括:将所述输入图像调整为预设规格的第一图像;将所述第一图像输入至残差神经网络,对第一图像执行不同采样频率的降采样处理得到多个不同尺度的第一特征图。
在一些可能的实施方式中,所述正向处理包括第一卷积处理和第一线性插值处理,所述反向处理包括第二卷积处理和第二线性插值处理。
在一些可能的实施方式中,所述利用第一金字塔神经网络对各所述第一特征图进行正向处理得到与各个所述第一特征图一一对应的第二特征图,包括:利用第一卷积核对第一特征图C 1...C n中的第一特征图C n进行卷积处理,获得与第一特征图C n对应的第二特征图F n,其中n表示第一特征图的数量,以及n为大于1的整数;对所述第二特征图F n执行线性插值处理获得与第二特征图F n对应的第一中间特征图F′ n,其中第一中间特征图F′ n的尺度与第一特征图C n-1的尺度相同;利用第二卷积核对第一特征图C n以外的各第一特征图C 1...C n-1进行卷积处理,得到分别与第一特征图C 1...C n-1一一对应的第二中间特征图C' 1...C' n-1,其中所述第二中间特征图的尺度与和其一一对应的第一特征图的尺度相同;基于所述第二特征图F n以及各所述第二中间特征图 C' 1...C' n-1,得到第二特征图F 1...F n-1以及第一中间特征图F′ 1...F′ n-1,其中所述第二特征图F i由所述第二中间特征图C′ i与所述第一中间特征图F′ i+1进行叠加处理得到,第一中间特征图F′ i由对应的第二特征图F i经线性插值得到,并且,所述第二中间特征图C′ i与第一中间特征图F′ i+1的尺度相同,其中,i为大于或者等于1且小于n的整数。
在一些可能的实施方式中,所述利用第二金字塔神经网络对各个所述第二特征图进行反向处理得到与各个所述第二特征图一一对应的第三特征图,包括:利用第三卷积核对第二特征图F 1...F m中的第二特征图F 1进行卷积处理,获得与第二特征图F 1对应的第三特征图R 1,其中m表示第二特征图的数量,以及m为大于1的整数;利用第四卷积核对第二特征图F 2...F m进行卷积处理,分别得到对应的第三中间特征图F″ 2...F″ m,其中,第三中间特征图的尺度与对应的第二特征图的尺度相同;
利用第五卷积核对第三特征图R 1进行卷积处理得到与第三特征图R 1对应的第四中间特征图R' 1;利用各第三中间特征图F″ 2...F″ m以及第四中间特征图R' 1,得到第三特征图R 2...R m以及第四中间特征图R' 2...R' m,其中,第三特征图R j由第三中间特征图F″ j与第四中间特征图R' j-1的叠加处理得到,第四中间特征图R' j-1由对应的第三特征图R j-1通过第五卷积核卷积处理获得,其中j为大于1且小于或者等于m。
在一些可能的实施方式中,所述对各所述第三特征图进行特征融合处理,并利用特征融合处理后的特征图获得所述输入图像中的各关键点的位置,包括:对各第三特征图进行特征融合处理,得到第四特征图:基于所述第四特征图获得所述输入图像中各关键点的位置。
在一些可能的实施方式中,所述对各第三特征图进行特征融合处理,得到第四特征图,包括:利用线性插值的方式,将各第三特征图调整为尺度相同的特征图;对所述尺度相同的特征图进行连接得到所述第四特征图。
在一些可能的实施方式中,在所述对各第三特征图进行特征融合处理,得到第四特征图之前,还包括:将第一组第三特征图分别输入至不同的瓶颈区块结构中进行卷积处理,分别得到更新后的第三特征图,各所述瓶颈区块结构中包括不同数量的卷积模块,其中,所述第三特征图包括第一组第三特征图和第二组第三特征图,所述第一组第三特征图和所述第二组第三特征图中均包括至少一个第三特征图。
在一些可能的实施方式中,所述对各第三特征图进行特征融合处理,得到第四特征图,包括:利用线性插值的方式,将各所述更新后的第三特征图以及所述第二组第三特征图,调整为尺度相同的特征图;对所述尺度相同的特征图进行连接得到所述第四特征图。
在一些可能的实施方式中,所述基于所述第四特征图获得所述输入图像中各关键点的位置,包括:利用第五卷积核对所述第四特征图进行降维处理;利用降维处理后的第四特征图确定输入图像的关键点的位置。
在一些可能的实施方式中,所述基于所述第四特征图获得所述输入图像中各关键点的位置,包括:利用第五卷积核对所述第四特征图进行降维处理;利用卷积块注意力模块对降维处理后的第四特征图中的特征进行提纯处理,得到提纯后的特征图;利用提纯后的特征图确定所述输入图像的关键点的位置。
在一些可能的实施方式中,所述方法还包括利用训练图像数据集训练所述第一金字塔神经网络,其包括:利用第一金字塔神经网络对所述训练图像数据集中各图像对应的第一特征图进行所述正向处理,得到所述训练图像数据集中各图像对应的第二特征图;利用各第二特征图确定识别的关键点;根据第一损失函数得到所述关键点的第一损失;利用所述第一损失反向调节所述第一金字塔神经网络中的各卷积核,直至训练次数达到设定的第一次数阈值。
在一些可能的实施方式中,所述方法还包括利用训练图像数据集训练所述第二金字塔神经网络,其包括:利用第二金字塔神经网络对所述第一金字塔神经网络输出的关于训练图像数据集中各图像对应的第二特征图进行所述反向处理,得到所述训练图像数据集中各图像对应的第三特征图;利用各第三特征图确定识别的关键点;根据第二损失函数得到识别的各关键点的第二损失;利用所述第二损失反向调节所述第二金字塔神经网络中卷积核,直至训练次数达到设定的第二次数阈值;或者,利用所述第二损失反向调节所述第一金字塔网络中的卷积核以及第二金字塔神经网络中的卷积核, 直至训练次数达到设定的第二次数阈值。
在一些可能的实施方式中,通过特征提取网络执行所述对各所述第三特征图进行特征融合处理,并且,在通过特征提取网络执行所述对各所述第三特征图进行特征融合处理之前,所述方法还包括:利用训练图像数据集训练所述特征提取网络,其包括:利用特征提取网络对所述第二金字塔神经网络输出的关于训练图像数据集中各图像对应的第三特征图进行所述特征融合处理,并利用特征融合处理后的特征图识别所述训练图像数据集中各图像的关键点;根据第三损失函数得到各关键点的第三损失;利用所述第三损失值反向调节所述特征提取网络的参数,直至训练次数达到设定的第三次数阈值;或者,利用所述第三损失函数反向调节所述第一金字塔神经网络中的卷积核参数、第二金字塔神经网络中的卷积核参数,以及所述特征提取网络的参数,直至训练次数达到设定的第三次数阈值。
根据本公开实施例的第二方面,提供了一种关键点检测装置,其包括:多尺度特征获取模块,配置为获得针对输入图像的多个尺度的第一特征图,各第一特征图的尺度成倍数关系;正向处理模块,配置为利用第一金字塔神经网络对各所述第一特征图进行正向处理得到与各个所述第一特征图一一对应的第二特征图,其中,所述第二特征图与其一一对应的所述第一特征图的尺度相同;反向处理模块,配置为利用第二金字塔神经网络对各个所述第二特征图进行反向处理得到与各个所述第二特征图一一对应的第三特征图,其中,所述第三特征图与其一一对应的所述第二特征图的尺度相同;关键点检测模块,配置为对各所述第三特征图进行特征融合处理,并利用特征融合处理后的特征图获得所述输入图像中的各关键点的位置。
在一些可能的实施方式中,所述多尺度特征获取模块,配置为将所述输入图像调整为预设规格的第一图像,并将所述第一图像输入至残差神经网络,对第一图像执行不同采样频率的降采样处理得到多个不同尺度的第一特征图。
在一些可能的实施方式中,所述正向处理包括第一卷积处理和第一线性插值处理,所述反向处理包括第二卷积处理和第二线性插值处理。
在一些可能的实施方式中,所述正向处理模块,配置为利用第一卷积核对第一特征图C 1...C n中的第一特征图C n进行卷积处理,获得与第一特征图C n对应的第二特征图F n,其中n表示第一特征图的数量,以及n为大于1的整数;以及对所述第二特征图F n执行线性插值处理获得与第二特征图F n对应的第一中间特征图F′ n,其中第一中间特征图F′ n的尺度与第一特征图C n-1的尺度相同;以及利用第二卷积核对第一特征图C n以外的各第一特征图C 1...C n-1进行卷积处理,得到分别与第一特征图C 1...C n-1一一对应的第二中间特征图C' 1...C' n-1,其中所述第二中间特征图的尺度与和其一一对应的第一特征图的尺度相同;并且基于所述第二特征图F n以及各所述第二中间特征图C' 1...C' n-1,得到第二特征图F 1...F n-1以及第一中间特征图F′ 1...F′ n-1,其中所述第二特征图F i由所述第二中间特征图C′ i与所述第一中间特征图F′ i+1进行叠加处理得到,第一中间特征图F′ i由对应的第二特征图F i经线性插值得到,并且,所述第二中间特征图C′ i与第一中间特征图F′ i+1的尺度相同,其中,i为大于或者等于1且小于n的整数。
在一些可能的实施方式中,所述反向处理模块,配置为利用第三卷积核对第二特征图F 1...F m中的第二特征图F 1进行卷积处理,获得与第二特征图F 1对应的第三特征图R 1,其中m表示第二特征图的数量,以及m为大于1的整数;以及利用第四卷积核对第二特征图F 2...F m进行卷积处理,分别得到对应的第三中间特征图F″ 2...F″ m,其中,第三中间特征图的尺度与对应的第二特征图的尺度相同;以及利用第五卷积核对第三特征图R 1进行卷积处理得到与第三特征图R 1对应的第四中间特征图R' 1;并且利用各第三中间特征图F″ 2...F″ m以及第四中间特征图R' 1,得到第三特征图R 2...R m以及第四中间特征图R' 2...R' m,其中,第三特征图R j由第三中间特征图F″ j与第四中间特征图R' j-1的叠加处理得到,第四中间特征图R' j-1由对应的第三特征图R j-1通过第五卷积核卷积处理获得,其中j为大于1且小于或者等于m。
在一些可能的实施方式中,所述关键点检测模块,配置为对各第三特征图进行特征融合处理, 得到第四特征图,并基于所述第四特征图获得所述输入图像中各关键点的位置。
在一些可能的实施方式中,所述关键点检测模块,配置为利用线性插值的方式,将各第三特征图调整为尺度相同的特征图,并对所述尺度相同的特征图进行连接得到所述第四特征图。
在一些可能的实施方式中,所述装置还包括:优化模块,配置为将第一组第三特征图分别输入至不同的瓶颈区块结构中进行卷积处理,分别得到更新后的第三特征图,各所述瓶颈区块结构中包括不同数量的卷积模块,其中,所述第三特征图包括第一组第三特征图和第二组第三特征图,所述第一组第三特征图和所述第二组第三特征图中均包括至少一个第三特征图。
在一些可能的实施方式中,所述关键点检测模块还配置为利用线性插值的方式,将各所述更新后的第三特征图以及所述第二组第三特征图,调整为尺度相同的特征图,并对所述尺度相同的特征图进行连接得到所述第四特征图。
在一些可能的实施方式中,所述关键点检测模块还配置为利用第五卷积核对所述第四特征图进行降维处理,并利用降维处理后的第四特征图确定输入图像的关键点的位置。
在一些可能的实施方式中,所述关键点检测模块还配置为利用第五卷积核对所述第四特征图进行降维处理,利用卷积块注意力模块对降维处理后的第四特征图中的特征进行提纯处理,得到提纯后的特征图,并利用提纯后的特征图确定所述输入图像的关键点的位置。
在一些可能的实施方式中,所述正向处理模块还配置为利用训练图像数据集训练所述第一金字塔神经网络,其包括:利用第一金字塔神经网络对所述训练图像数据集中各图像对应的第一特征图进行所述正向处理,得到所述训练图像数据集中各图像对应的第二特征图;利用各第二特征图确定识别的关键点;根据第一损失函数得到所述关键点的第一损失;利用所述第一损失反向调节所述第一金字塔神经网络中的各卷积核,直至训练次数达到设定的第一次数阈值。
在一些可能的实施方式中,所述反向处理模块还配置为利用训练图像数据集训练所述第二金字塔神经网络,其包括:利用第二金字塔神经网络对所述第一金字塔神经网络输出的关于训练图像数据集中各图像对应的第二特征图进行所述反向处理,得到所述训练图像数据集中各图像对应的第三特征图;利用各第三特征图确定识别的关键点;根据第二损失函数得到识别的各关键点的第二损失;利用所述第二损失反向调节所述第二金字塔神经网络中卷积核,直至训练次数达到设定的第二次数阈值;或者,利用所述第二损失反向调节所述第一金字塔网络中的卷积核以及第二金字塔神经网络中的卷积核,直至训练次数达到设定的第二次数阈值。
在一些可能的实施方式中,所述关键点检测模块还配置为通过特征提取网络执行所述对各所述第三特征图进行特征融合处理,并且在通过特征提取网络执行所述对各所述第三特征图进行特征融合处理之前,还利用训练图像数据集训练所述特征提取网络,其包括:利用特征提取网络对所述第二金字塔神经网络输出的关于训练图像数据集中各图像对应的第三特征图进行所述特征融合处理,并利用特征融合处理后的特征图识别所述训练图像数据集中各图像的关键点;根据第三损失函数得到各关键点的第三损失;利用所述第三损失值反向调节所述特征提取网络的参数,直至训练次数达到设定的第三次数阈值;或者,利用所述第三损失函数反向调节所述第一金字塔神经网络中的卷积核参数、第二金字塔神经网络中的卷积核参数,以及所述特征提取网络的参数,直至训练次数达到设定的第三次数阈值。
根据本公开实施例的第三方面,提供了一种电子设备,其包括:处理器;用于存储处理器可执行指令的存储器;其中,所述处理器被配置为:执行第一方面中任意一项所述的方法。
根据本公开实施例的第四方面,提供了一种计算机可读存储介质,其上存储有计算机程序指令,所述计算机程序指令被处理器执行时实现第一方面中任意一项所述的方法。
本公开实施例提出了一种利用双向金字塔神经网络来执行关键点特征检测,其中不仅利用正向处理的方式得到多尺度特征,同时还利用反向处理融合更多的特征,从而能够进一步提高关键点的检测精度。
应当理解的是,以上的一般描述和后文的细节描述仅是示例性和解释性的,而非限制本公开。
根据下面参考附图对示例性实施例的详细说明,本公开的其它特征及方面将变得清楚。
附图说明
此处的附图被并入说明书中并构成本说明书的一部分,这些附图示出了符合本公开的实施例,并与说明书一起用于说明本公开的技术方案。
图1示出根据本公开实施例的一种关键点检测方法的流程图;
图2示出根据本公开实施例的一种关键点检测方法中步骤S100的流程图;
图3示出本公开实施例的关键点检测方法的另一流程图;
图4示出根据本公开实施例的一种关键点检测方法中的步骤S200的流程图;
图5示出根据本公开实施例的关键点检测方法中步骤S300的流程图;
图6出根据本公开实施例的关键点检测方法中步骤S400的流程图;
图7示出根据本公开实施例的关键点检测方法中步骤S401的流程图;
图8示出根据本公开实施例的关键点检测方法的另一流程图;
图9示出根据本公开实施例的关键点检测方法中步骤S402的流程图;
图10示出根据本公开实施例的一种关键点检测方法中的训练第一金字塔神经网络的流程图;
图11示出根据本公开实施例的一种关键点检测方法中的训练第二金字塔神经网络的流程图;
图12示出根据本公开实施例的一种关键点检测方法中的训练特征提取网络模型的流程图;
图13示出根据本公开实施例的一种关键点检测装置的框图;
图14示出根据本公开实施例的一种电子设备800的框图;
图15示出根据本公开实施例的一种电子设备1900的框图。
具体实施方式
以下将参考附图详细说明本公开的各种示例性实施例、特征和方面。附图中相同的附图标记表示功能相同或相似的元件。尽管在附图中示出了实施例的各种方面,但是除非特别指出,不必按比例绘制附图。
在这里专用的词“示例性”意为“用作例子、实施例或说明性”。这里作为“示例性”所说明的任何实施例不必解释为优于或好于其它实施例。
本文中术语“和/或”,仅仅是一种描述关联对象的关联关系,表示可以存在三种关系,例如,A和/或B,可以表示:单独存在A,同时存在A和B,单独存在B这三种情况。另外,本文中术语“至少一种”表示多种中的任意一种或多种中的至少两种的任意组合,例如,包括A、B、C中的至少一种,可以表示包括从A、B和C构成的集合中选择的任意一个或多个元素。
另外,为了更好地说明本公开实施例,在下文的具体实施方式中给出了众多的具体细节。本领域技术人员应当理解,没有某些具体细节,本公开实施例同样可以实施。在一些实例中,对于本领域技术人员熟知的方法、手段、元件和电路未作详细描述,以便于凸显本公开实施例的主旨。
本公开实施例提供了一种关键点检测方法,该方法可以用于执行人体图像的关键点检测,其利用了两个金字塔网络模型分别执行关键点的多尺度特征的正向处理和反向处理,融合了更多的特征信息,能够提高关键点位置检测的精度。
图1示出根据本公开实施例的一种关键点检测方法的流程图。其中,本公开实施例的关键点检测方法可以包括:
S100:获得针对输入图像的多个尺度的第一特征图,各第一特征图的尺度成倍数关系。
本公开实施例采用输入图像的多尺度特征的融合的方式执行上述关键点的检。首先可以获取输入图像的多个尺度的第一特征图,各第一特征图的尺度不同,且各尺度之间存在倍数的关系。本公开实施例可以利用多尺度分析算法得到输入图像的多个尺度的第一特征图,或者也可以通过能够执行多尺度分析的神经网络模型获得输入图像的多个尺度的第一特征图,本公开实施例不作具体限定。
S200:利用第一金字塔神经网络对各所述第一特征图进行正向处理得到与各个所述第一特征图一一对应的第二特征图,其中,所述第二特征图与其一一对应的所述第一特征图的尺度相同。
在本实施例中,正向处理可以包括第一卷积处理以及第一线性插值处理,通过第一金字塔神经网络的正向处理过程,可以得到与相应的第一特征图尺度相同的第二特征图,各第二特征图的进一步融合了输入图像的各特征,并且得到的第二特征图与第一特征图的数量相同,且第二特征图与对应的第一特征图的尺度相同。例如,本公开实施例得到的第一特征图可以为C 1、C 2、C 3和C 4,对应的正向处理后得到的第二特征图可以为F 1、F 2、F 3和F 4。其中,在第一特征图C 1至C 4的尺度关系为C 1的尺度为C 2的尺度的2倍,C 2的尺度为C 3的尺度的二倍,以及C 3的尺度为C 4的二倍时,得到的第二特征图F 1至F 4中,F 1与C 1的尺度相同,F 2与C 2的尺度相同,F 3与C 3的尺度相同,以及F 4与C 4的尺度相同,并且第二特征图F 1的尺度为F 2的尺度的2倍,F 2的尺度为F 3的尺度的二倍,以及F 3的尺度为F 4的 二倍。上述仅为第一特征图经过正向处理得到第二特征图的示例性说明,不作为本公开的具体限定。
S300:利用第二金字塔神经网络对各第二特征图进行反向处理得到与各个所述第二特征图一一对应的第三特征图,所述反向处理包括第二卷积处理,其中,所述第三特征图与其一一对应的所述第二特征图的尺度相同。
在本实施例中,反向处理包括第二卷积处理以及第二线性插值处理,通过第二金字塔神经网络的反向处理过程,可以得到与相应的第二特征图尺度相同的第三特征图,且各第三特征图相对于第二特征图进一步融合了输入图像的特征,并且得到的第三特征图与第二特征图的数量相同,且第三特征图与对应的第二特征图的尺度相同。例如,本公开实施例得到的第二特征图可以为F 1、F 2、F 3和F 4,对应的反向处理后得到的第三特征图可以为R 1、R 2、R 3和R 4。其中,在第二特征图F 1、F 2、F 3和F 4的尺度关系为F 1的尺度为F 2的尺度的2倍,F 2的尺度为F 3的尺度的二倍,以及F 3的尺度为F 4的二倍时,得到的第三特征图R 1至R 4中,R 1与F 1的尺度相同,R 2与F 2的尺度相同,R 3与F 3的尺度相同,以及R 4与F 4的尺度相同,并且第三特征图R 1的尺度为R 2的尺度的2倍,R 2的尺度为R 3的尺度的二倍,以及R 3的尺度为R 4的二倍。上述仅为第二特征图经反向处理得到第三特征图的示例性说明,不作为本公开的具体限定。
S400:对各所述第三特征图进行特征融合处理,并利用特征融合处理后的特征图获得所述输入图像中的各关键点的位置。
本公开实施例中,在对各第一特征图经正向处理得到第二特征图,以及根据第二特征图的反向处理得到第三特征图后,即可以执行各第三特征图的特征融合处理。例如本公开实施例可以利用对应的卷积处理的方式实现各第三特征图的特征融合,以及在第三特征图的尺度不相同时还可以执行尺度的转变,而后执行特征图的拼接,以及关键点的提取。
本公开实施例可以执行对输入图像的不同关键点的检测,例如在输入图像为人物的图像时,关键点可以为左右眼睛、鼻子、左右耳朵、左右肩膀、左右手肘、左右手腕、左右胯部、左右膝盖、左右脚踝中的至少一种,或者在其他实施例中,输入图像也可以其他类型的图像,在执行关键点检测时,可以识别其他的关键点。因此,本公开实施例可以根据第三特征图的特征融合结果,进一步执行关键点的检测识别。
基于上述配置,本公开实施例可以通过双向金字塔神经网络(第一金字塔神经网络和第二金字塔神经网络)分别基于第一特征图执行正向处理以及进一步的反向处理,能够有效的提高输入图像的特征融合度,进一步提高关键点的检测精度。如上所示,本公开实施例可以首先获取输入图像,该输入图像可以为任意的图像类型,例如可以是人物图像、风景图像、动物图像等等。对于不同类型的图像,可以识别不同的关键点。例如,本公开实施例以人物图像为例进行说明。首先可以通过步骤S100获取输入图像在多个不同尺度下的第一特征图。图2示出根据本公开实施例的一种关键点检测方法中步骤S100的流程图。其中,获得针对输入图像的不同尺度的第一特征图(步骤S100)可以包括:
S101:将所述输入图像调整为预设规格的第一图像。
本公开实施例可以首先归一化输入图像的尺寸规格,即可以首先将输入图像调整为预设规格的第一图像,其中本公开实施例中预设规格可以为256pix*192pix,pix为像素值,在其他的实施例中,可以将输入图像统一转换为其他规格的图像,本公开实施例对此不进行具体限定。
S102:将所述第一图像输入至残差神经网络,对第一图像执行不同采样频率的降采样处理得到不同尺度的第一特征图。
在得到预设规格的第一图像之后,可以对该第一图像执行多个采样频率的采样处理。例如,本公开实施例可以通过将第一图像输入至残差神经网络,通过残差神经网络处理得到针对第一图像的不同尺度的第一特征图。其中,可以利用不同的采样频率对第一图像进行将采样处理从而得到不同尺度的第一特征图。本公开实施例的采样频率可以为1/8、1/16、1/32等,但本公开实施例对此不进行限定。另外,本公开实施例中的特征图是指图像的特征矩阵,例如本公开实施例的特征矩阵可以为三维矩阵,以及本公开实施例中所述的特征图的长度和宽度可以分别为对应的特征矩阵在行方向和列方向上的维度。
通过步骤S100处理后得到的输入图像的多个不同尺度的第一特征图。并且通过控制降采样的采样频率可以使得各第一特征图之间的尺度的关系为
Figure PCTCN2019083721-appb-000001
Figure PCTCN2019083721-appb-000002
其中,C i表示各第一特征图,L(C i)表示第一特征图C i的长度,W(C i)表示第一特征图C i的宽度,k 1为大于或者等于1的整数,i为变量,且i的范围为[2,n],n为第一特征图 的数量。即本公开实施例中的各第一特征图的长度和宽度之间的关系均为2的k 1次方倍。
图3示出本公开实施例的关键点检测方法的另一流程图。其中,(a)部分示出本公开实施例的步骤S100的过程,通过步骤S100可以获得四个第一特征图C 1、C 2、C 3和C 4,其中,第一特征图C 1的长度和宽度可以分别对应的为第一特征图C 2的长度和宽度的二倍,第二特征图C 2的长度和宽度可以分别对应的为第三特征图C 3的长度和宽度的二倍,以及第三特征图C 3的长度和宽度可以分别对应的为第四特征图C 4的长度和宽度的二倍。本公开实施例上述C 1和C 2之间、C 2和C 3之间,以及C 3和C 4之间的尺度倍数可以均相同,例如k 1取值为1。在其他的实施例中,k 1可以为不同的值,例如可以为,第一特征图C 1的长度和宽度可以分别对应的为第一特征图C 2的长度和宽度的二倍,第二特征图C 2的长度和宽度可以分别对应的为第三特征图C 3的长度和宽度的四倍,以及第三特征图C 3的长度和宽度可以分别对应的为第四特征图C 4的长度和宽度的八倍,但本公开实施例对此不进行限定。
在获得输入图像的不同尺度的第一特征图之后,可以对通过步骤S200执行第一特征图的正向处理过程,得到融合了各第一特征图的特征的多个不同尺度的第二特征图。
图4示出根据本公开实施例的一种关键点检测方法中的步骤S200的流程图。其中,所述利用第一金字塔神经网络对各所述第一特征图进行正向处理得到与各个所述第一特征图一一对应的第二特征图(步骤S200),包括:
S201:利用第一卷积核对第一特征图C 1...C n中的第一特征图C n进行卷积处理,获得与第一特征图C n对应的第二特征图F n,其中,其中n表示第一特征图的数量,以及n为大于1的整数,并且第一特征图C n的长度和宽度分别与第二特征图F n的长度和宽度对应相同。
本公开实施例中的第一金字塔神经网络执行的正向处理可以包括第一卷积处理以及第一线性插值处理,也可以包括其他的处理过程,本公开实施例对此不进行限定。
在一种可能的实施方式中,本公开实施例获得的第一特征图可以为C 1...C n,即n个第一特征图,且C n可以为长度和宽度最小的特征图,即尺度最小的第一特征图。其中,首先可以利用第一金字塔神经网络对第一特征图C n进行卷积处理,即利用第一卷积核对第一特征图C n进行卷积处理,得到第二特征图F n。该第二特征图F n的长度和宽度均与第一特征图C n的长度和宽度分别相同。其中,第一卷积核可以为3*3的卷积核,或者也可以是其他类型的卷积核。
S202:对所述第二特征图F n执行线性插值处理获得与第二特征图F n对应的第一中间特征图F′ n,其中第一中间特征图F′ n的尺度与第一特征图C n-1的尺度相同。
在得到第二特征图F n之后,可以利用该第二特征图F n获得与其对应的第一中间特征图F′ n,本公开实施例可以通过对第二特征图F n执行线性插值处理获得与第二特征图F n对应的第一中间特征图F′ n,其中,第一中间特征图F′ n的尺度与第一特征图C n-1的尺度相同,例如,在C n-1的尺度为C n的尺度的二倍时,第一中间特征图F′ n的长度为第二特征图F n的长度的二倍,以及第一中间特征图F′ n的宽度为第二特征图F n的宽度的二倍。
S203:利用第二卷积核对第一特征图C n以外的各第一特征图C 1...C n-1进行卷积处理,得到分别与第一特征图C n以外的各第一特征图C 1...C n-1一一对应的第二中间特征图C' 1...C' n-1,其中所述第二中间特征图的尺度与和其一一对应的第一特征图的尺度相同。
同时,本公开实施例还可以获得第一特征图C n以外的各第一特征图C 1...C n-1对应的第二中间特征图C' 1...C' n-1,其中,可以利用第二卷积核分别对第一特征图C 1...C n-1进行第二卷积处理,分别得到与各第一特征图C 1...C n-1一一对应的第二中间特征图C' 1...C' n-1,其中第二卷积核可以为1*1的卷积核,但本公开对此不作具体限定。通过第二卷积处理得到的各第二中间特征图的尺度与对应的第一特征图的尺度分别相同。其中,本公开实施例可以按照第一特 征图C 1...C n-1的倒序,获得各第一特征图C 1...C n-1的第二中间特征图C' 1...C' n-1。即,可以先获得第一特征图C n-1对应的第二中间图C' n-1,而后获得第一特征图C n-2的对应的第二中间图C' n-2,以此类推,直至获得第一特征图C 1对应的第二中间特征图C' 1
S204:基于所述第二特征图F n以及各所述第二中间特征图C' 1...C' n-1,得到第二特征图F 1...F n-1以及第一中间特征图F′ 1...F′ n-1,,第二特征图F 1...F n-1(可记为第二特征图F i)为与第一特征图C n以外的各第一特征图C 1...C n-1对应的第二特征图;第一中间特征图F′ 1...F′ n-1为与各第二特征图F i对应的第一中间特征图;其中与第一特征图C 1...C n-1中的第一特征图C i对应的第二特征图F i由第二中间特征图C′ i与第一中间特征图F′ i+1进行叠加处理(加和处理)得到,并且第一中间特征图F′ i由对应的第二特征图F i经线性插值得到,并且,所述第二中间特征图C′ i与第以中间特征图F′ i+1的尺度相同,其中,i为大于或者等于1且小于n的整数。
另外,在获得各第二中间特征图的同时,或者获得各第二中间特征图之后还可以对应的获得第一中间特征图F′ n以外的其他第一中间特征图F′ 1...F′ n-1,本公开实施例中,与第一特征图C 1...C n-1中的第一特征图C i对应的第二特征图F i=C′ i+F′ i+1,其中,第二中间特征图C′ i的尺度(长度和宽度)分别与第一中间特征图F′ i+1的尺度(长度和宽度)相等,并且第二中间特征图C′ i的长度和宽度与第一特征图C i的长度和宽度相同,因此得到的第二特征图F i的长度和宽度分别为第一特征图C i的长度和宽度。其中,i为大于或者等于1且小于n的整数。
具体的,本公开实施例依然可以采用倒序的处理方式获得第二特征图F n以外的各第二特征图F i。即,本公开实施例可以首先获得第一中间特征图F n-1,其中,可以利用第一特征图C n-1对应的第二中间图C' n-1与第一中间特征图F′ n进行叠加处理得到第二特征图F n-1,其中,第二中间特征图C' n-1的长度和宽度分别与第一中间特征图F′ n的长度和宽度相同,以及第二特征图F n-1的长度和宽度为第二中间特征图C' n-1和F′ n的长度和宽度。此时第二特征图F n-1的长度和宽度分别为第二特征图F n的长度和宽度的二倍(C n-1的尺度为C n的尺度的二倍)。进一步地,可以对第二特征图F n-1进行线性插值处理,得到第一中间特征图F′ n-1,使得F′ n-1的尺度与C n-1的尺度相同,继而可以利用第一特征图C n-2对应的第二中间图C' n-2与第一中间特征图F′ n-1进行叠加处理得到第二特征图F n-2,其中,第二中间特征图C' n-2的长度和宽度分别与第一中间特征图F′ n-1的长度和宽度相同,以及第二特征图F n-2的长度和宽度为第二中间特征图C' n-2和F′ n-1的长度和宽度。例如第二特征图F n-2的长度和宽度分别为第二特征图F n-1的长度和宽度的二倍。以此类推,可以最终获得第一中间特征图F′ 2,以及根据该第一中间特征图F′ 2与第一特征图C' 1的叠加处理得到第二特征图F 1,F 1的长度和宽度分别为与C 1的长度和宽度的相同。从而得到各第二特征图,并满足
Figure PCTCN2019083721-appb-000003
Figure PCTCN2019083721-appb-000004
并且L(F n)=L(C n),W(F n)=W(C n)。
例如,以上述四个第一特征图C 1、C 2、C 3和C 4为例进行说明。如图3所示,步骤S200可以使用第一金字塔神经网络(Feature Pyramid Network--FPN)来获得多尺度的第二特征图。其中,首先可以将C 4经过一个3*3的第一卷积核计算得到一个新的特征图F 4(第二特征图),F 4的长度和宽度的大小与C 4相同。对F 4进行双线形插值的上采样(upsample)操作,得到一个长和宽都放大两倍的特征图,即第一中间特征图F′ 4。C 3经过一个1*1的第二卷积核计算得到一个第二中间特征图C' 3,C' 3与F′ 4大小相同,两个特征图相加,得到新的特征图F 3(第二特征图),使得第二特征图F 3的长度和宽度分别为第二特征图F 4二倍。对F 3进行双线形插值的上采样(upsample)操作,得到一个长和 宽都放大两倍的特征图,即第一中间特征图F′ 3。C 2经过一个1*1的第二卷积核计算得到一个第二中间特征图C' 2,C' 2与F′ 3大小相同,两个特征图相加,得到新的特征图F 2(第二特征图),使得第二特征图F 2的长度和宽度分别为第二特征图F 3二倍。对F 2进行双线形插值的上采样(upsample)操作,得到一个长和宽都放大两倍的特征图,即第一中间特征图F′ 2。C 1经过一个1*1的第二卷积核计算得到一个第二中间特征图C' 1,C' 1与F′ 2大小相同,两个特征图相加,得到新的特征图F 2(第二特征图),使得第二特征图F 1的长度和宽度分别为第二特征图F 2二倍。经过FPN之后,同样得到了四个不同尺度的第二特征图,分别记为F 1、F 2、F 3和F 4。并且F 1和F 2之间的长度和宽度的倍数与C 1和C 2之间的长度和宽度的倍数相同,以及F 2和F 3之间的长度和宽度的倍数与C 2和C 3之间的长度和宽度的倍数相同,F 3和F 4之间的长度和宽度的倍数与C 3和C 4之间的长度和宽度的倍数相同。
通过上述金字塔网络模型的正向处理之后,可以使得各第二特征图中融合更多的特征,为了进一步提高特征的提取精度,本公开实施例在步骤S200之后,还利用第二金字塔神经网络对各第二特征图执行反向处理。其中,反向处理可以包括第二卷积处理以及第二线性插值处理,同样,也可以包括其他处理,本公开实施例对此不进行具体限定。
图5示出根据本公开实施例的关键点检测方法中步骤S300的流程图。其中,所述利用第二金字塔神经网络对各第二特征图进行反向处理得到不同尺度的第三特征图R i(步骤S300),可以包括:
S301:利用第三卷积核对F 1...F m中的第二特征图F 1进行卷积处理,获得与第二特征图F 1对应的第三特征图R 1,其中第三特征图R 1的长度和宽度分别与第一特征图C 1的长度和宽度对应相同,其中m表示第二特征图的数量,以及m为大于1的整数,此时m与第一特征图的数量n相同。
在反向处理的过程中,可以首先从长度和宽度最大的第二特征图F 1进行反向处理,例如,可以通过第三卷积核对该第二特征图F 1进行卷积处理,得到长度和宽度都与F 1相同的第三中间特征图R 1。其中,第三卷积核可以为3*3的卷积核,也可以是其他类型的卷积核,本领域技术领域可以根据不同的需求选择所需的卷积核。
S302:利用第四卷积核对第二特征图F 2...F m进行卷积处理,分别得到对应的第三中间特征图F″ 2...F″ m,其中,第三中间特征图的尺度与对应的第二特征图的尺度相同。
在得到第三特征图R 1之后,可以利用第四卷积核对第二特征图F 1以外的各第二特征图F 2...F m分别执行卷积处理,得到对应的第三中间特征图F″ 1...F″ m-1。步骤S302中,可以将第二特征图F 1以外的第二特征图F 2...F m通过第四卷积核做卷积处理,其中可以首先对F 2进行卷积处理得到对应的第三中间特征图F″ 2,继而可以对F 3进行卷积处理得到对应的第三中间特征图F″ 3,以此类推,得到第二特征图F m对应的第三中间特征图F″ n。其中,本公开实施例中,各第三中间特征图F″ j的长度和宽度可以为对应的第二特征图F j的长度和宽度。
S303:利用第五卷积核对第三特征图R 1进行卷积处理得到与第三特征图R 1对应的第四中间特征图R' 1
在得到第三特征图R 1之后,可以利用第四卷积核对第二特征图F 1以外的各第二特征图F 2...F m分别执行卷积处理,得到对应的第三中间特征图F″ 1...F″ m-1。步骤S302中,可以将第二特征图F 1以外的第二特征图F 2...F m通过第四卷积核做卷积处理,其中可以首先对F 2进行卷积处理得到对应的第三中间特征图F″ 2,继而可以对F 3进行卷积处理得到对应的第三中间特征图F″ 3,以此类推,得到第二特征图F m对应的第三中间特征图F″ n。其中,本公开实施例中,各第三中间特征图F″ j的长度和宽度可以为对应的第二特征图F j的长度和宽度的一半。
S304:利用各第三中间特征图F″ 2...F″ m以及第四中间特征图R' 1,得到第三特征图R 2...R m,其中, 第三特征图R j由第三中间特征图F″ j与第四中间特征图R' j-1的叠加处理得到,以及第四中间特征图R' j-1由对应的第三特征图R j-1通过第五卷积核卷积处理获得,其中j为大于1且小于或者等于m。
在执行步骤S301之后,或者执行S302之后,还可以利用第五卷积核对第三特征图R 1进行卷积处理得到第三特征图R 1对应的第四中间特征图R' 1。其中,第四中间特征图R' 1的长度和宽度为第二特征图F 2的长度和宽度。
另外,还可以利用步骤S302得到的第三中间特征图F″ i以及步骤S303得到的第四中间特征图R' 1,得到第三特征图R 1以外的第三特征图R 2...R m。其中,第三特征图R 1之外的各第三特征图R 2...R m由第三中间特征图F″ j与第四中间特征图R' j-1的叠加处理得到。
具体的,步骤S304中,可以分别利用对应的第三中间特征图F″ i与第四中间特征图R' i-1进行叠加处理得到第三特征图R 1之外的各第三特征图R j。其中,可以首先利用第三中间特征图F″ 2与第四中间特征图R' 1的加和结果获得第三特征图R 2。而后,利用第五卷积核对R 2进行卷积处理得到第四中间特征图R' 2,通过第三中间特征图F″ 3与第四中间特征图R' 2之间的加和结果获得第三特征图R 3。以此类推,可以进一步得到其余第四中间特征图R' 3...R' m,以及第三特征图R 4…R m
另外,本公开实施例中,获得的各第四中间特征图R' 1的长度和宽度分别与第二特征图F 2的长度和宽度相同。以及第四中间特征图R' j的长度和宽度分别与第四中间特征图F″ j+1的长度和宽度相同。从而,得到的第三特征图R j的长度和宽度分别为第二特征图F i的长度和宽度,进一步的各第三特征图R 1…Rn的长度和宽度分别对应的与第一特征图C 1…C n的长度和宽度相等。
下面举例说明反向处理的过程。如图3所示,接着利用第二特征金字塔网络(Reverse Feature Pyramid Network--RFPN)来进一步优化多尺度特征。第二特征图F 1经过一个3*3的卷积核(第三卷积核),得到一个新的特征图R 1(第三特征图),R 1长和宽的大小与F 1相同。特征图R 1经过一个卷积核为3*3(第五卷积核),步长(stride)为2的卷积计算得到一个新的特征图,记为R′ 1,R′ 1的长和宽均可以是R 1的一半。第二特征图F 2经过一个3*3的卷积核(第四卷积核)计算得到一个新的特征图,记为F″ 2。R′ 1与F″ 2的大小相同,将R′ 1与F″ 2相加得到新的特征图R 2。对R 2和F 3重复R 1和F 2的操作,得到新的特征图R 3。对R 3和F 4重复R 1和F 2的操作,得到新的特征图R 4。经过RFPN之后,同样得到了四个不同尺度的特征图,分别记为R 1、R 2、R 3和R 4。同样的,R 1和R 2之间的长度和宽度的倍数与C 1和C 2之间的长度和宽度的倍数相同,以及R 2和R 3之间的长度和宽度的倍数与R 2和R 3之间的长度和宽度的倍数相同,R 3和R 4之间的长度和宽度的倍数与C 3和C 4之间的长度和宽度的倍数相同。
基于上述配置,即可以得到经第二集资他网络模型进行反向处理得到的第三特征图R 1…Rn,经过正向和反向处理两个处理过程可以进一步提高图像的融合的特征,基于各第三特征图可以精确的识别特征点。
在步骤S300之后,则可以根据各第三特征图R i的特征融合结果,获得输入图像的各关键点的位置。其中,图6示出根据本公开实施例的关键点检测方法中步骤S400的流程图。其中,所述对各所述第三特征图进行特征融合处理,并利用特征融合处理后的特征图获得所述输入图像中的各关键点的位置(步骤S400),可以包括:
S401:对各第三特征图进行特征融合处理,得到第四特征图。
本公开实施例中,在获得各尺度的第三特征图R 1...R n之后,可以对各第三特征图进行特征融合,由于本公开实施例中各第三特征图的长度和宽度不同,因此可以将分别R 2…R n进行线性插值处理,最终使得各第三特征图R 2…R n的长度和宽度与第三特征图R 1的长度和宽度相同。继而可以将处理后的第三特征图进行组合形成第四特征图。
S402:基于所述第四特征图获得所述输入图像中各关键点的位置。
在获得第四特征图之后,可以对第四特征图进行降维处理,例如可以通过卷积 处理对第四特征图进行降维,并利用降维后的特征图识别输入图像的特征点的位置。
图7示出根据本公开实施例的关键点检测方法中步骤S401的流程图,其中,所述对各第三特征图进行特征融合处理,得到第四特征图(步骤S401)可以包括:
S4012:利用线性插值的方式,将各第三特征图调整为尺度相同的特征图。
由于本公开实施例获得的各第三特征图R 1...R n的尺度不同,因此首先需要将各第三特征图调整为尺度相同的特征图,其中,本公开实施例可以对各第三特征图执行不同的线性插值处理使得各特征图的尺度相同,其中线性插值的倍数可以与各第三特征图之间的尺度倍数相关。
S4013:对线性插值处理后的各特征图进行连接得到所述第四特征图。
在得到尺度相同的各特征图后,可以将各特征图进行拼接组合得到第四特征图,例如本公开实施例的各插值处理后的特征图的长度和宽度均相同,可以将各特征图在高度方向上进行连接得到第四特征图,如,经过S4012处理后的各特征图可以表示为A、B、C和D,得到的第四特征图可以为
Figure PCTCN2019083721-appb-000005
另外,步骤S401之前,本公开实施例为了对小尺度的特征进行优化,可以将长度和宽度较小的第三特征图进一步的优化,可以对该部分特征进行进一步的卷积处理。图8示出根据本公开实施例的关键点检测方法的另一流程图,其中,在所述对各第三特征图进行特征融合处理,得到第四特征图之前,还可以包括S4011:
S4011:将第一组第三特征图分别输入至不同的瓶颈区块结构中进行卷积处理,分别对应的得到更新后的第三特征图,各所述瓶颈区块结构中包括不同数量的卷积模块;其中,所述第三特征图包括第一组第三特征图和第二组第三特征图,所述第一组第三特征图和所述第二组第三特征图中均包括至少一个第三特征图。
如上所述,为了优化小尺度特征图内的特征,可以对小尺度的特征图进一步卷积处理,其中,可以将第三特征图R 1…R m分成两组,其中第一组第三特征图的尺度小于第二组第三特征图的尺度。对应的,可以将第一组第三特征图内的各第三特征图分别输入至不同的瓶颈区块结构内,得到更新后的第三特征图,该瓶颈区块结构内可以包括至少一个卷积模块,不同的瓶颈区块结构中的卷积模块的数量可以不同,其中,经过瓶颈区块结构卷积处理后得到的特征图的大小与输入之前的第三特征图的大小相同。
其中,可以按照第三特征图的数量的预设比例值确定该第一组第三特征图。例如,预设比例可以为50%,即可以将各第三特征图中尺度较小的一半的第三特征图作为第一组第三特征图输入至不同的瓶颈区块结构中进行特征优化处理。该预设比例可以也可以为其他的比例值,本公开对此不进行限定。或者,在另一些可能的实施例中,也可以按照尺度阈值确定该输入至瓶颈区块结构中的第一组第三特征图。小于该尺度阈值的特征图即确定需要输入至瓶颈区块结构中进行特征优化处理。对于尺度阈值的确定可以根据各特征图的尺度进行确定,本公开实施例对此不进行具体限定。
另外,对于瓶颈区块结构的选择,本公开实施例不作具体限定,其中卷积模块的形式可以根据需求进行选择。
S4012:利用线性插值的方式,将更新后的第三特征图以及第二组第三特征图,调整为尺度相同的特征图。
在执行步骤S4011之后,可以将优化后的第一组第三特征图以及第二组第三特征进行尺度归一化,即将各特征图调整为尺寸相同的特征图。本公开实施例通过为各S4011优化后的第三特征图以及第二组第三特征图分别执行对应的线性插值处理,从而得到大小相同的特征图。
本公开实施例中,如图3所示的(d)部分,为了对小尺度的特征进行优化在R 2、R 3和R 4后接了不同个数的瓶颈区块(bottleneck block)结构,在R 2后接一个bottleneck block后得倒新的特征图,记为R″ 2,在R 3后接两个bottleneck block后得倒新的特征图,记为R″ 3,在R 4后接三个bottleneck block后得倒新的特征图,记为R″ 4。为了进行融合,我们需 要将四个特征图R 1、R″ 2、R″ 3、R″ 4的大小统一,所以对R″ 2进行双线形插值的上采样(upsample)操作放大2倍,得到特征图R″′ 2,对R″ 3进行双线形插值的上采样(upsample)操作放大4倍,得到特征图R″′ 3,对R″ 4进行双线形插值的上采样(upsample)操作放大8倍,得到特征图R″′ 4。此时,R 1、R″′ 2、R″′ 3、R″′ 4尺度相同。
S4013:对各尺度相同的特征图进行连接得到所述第四特征图。
步骤S4012之后,可以将尺度相同的特征图进行连接,例如将上述四个特征图连接(concat)得到新的特征图即为第四特征图,例如R 1、R″′ 2、R″′ 3、R″′ 4四个特征图都是256维,得到的第四特征图即可以为1024维。
通过上述不同实施例中的配置可以得到相应的第四特征图,在获得第四特征图之后,即可以根据第四特征图得到输入图像的关键点位置。其中,可以直接对第四特征图进行降维处理,利用降维处理后的特征图确定输入图像的关键点的位置。在另一些实施例中,还可以对降维后的特征图进行提纯处理,进一步提高关键点的精度。图9示出根据本公开实施例的关键点检测方法中步骤S402的流程图,所述基于所述第四特征图获得所述输入图像中各关键点的位置,可以包括:
S4021:利用第五卷积核对所述第四特征图进行降维处理。
本公开实施例中,执行降维处理的方式可以为卷积处理,即利用预设的卷积模块对第四特征图进行卷积处理,以实现第四特征图的降维,得到例如256维的特征图。
S4022:利用卷积块注意力模块对降维处理后的第四特征图中的特征进行提纯处理,得到提纯后的特征图。
而后,可以进一步利用卷积块注意力模块对降维处理后的第四特征图进行提纯处理。其中卷积块注意力模块可以为现有技术中的卷积块注意力模块。例如本公开实施例的卷积块注意力模块可以包括通道注意力单元以及重要度注意力单元。其中,可以首先将降维处理后的第四特征图输入至通道注意力单元,其中首先可以对降维处理后的第四特征图进行基于高度和宽度的全局最大池化(global max pooling)以及全局平均池化(global average pooling),而后分别将经全局最大池化得到的第一结果以及经全局平均池化得到的第二结果输入至多层感知器(MLP),并对经MLP处理后的两个结果进行加和处理得到第三结果,对将第三结果经过激活处理得到通道注意力特征图。
在得到通道注意力特征图之后,将该通道注意力特征图输入至重要度注意力单元,首先可以对该通道注意力特征图输入至基于通道的全局最大池化(global max pooling)以及全局平均池化(global average pooling)处理,分别得到第四结果和第五结果,再将第四结果和第五结果进行连接,而后对连接后的结果通过卷积处理进行降维,利用sigmoid函数对降维结果进行处理得到重要度注意力特征图,而后将重要度注意力特征图与通道注意力特征图相乘积,得到提纯后的特征图。上述仅为本公开实施例对于卷积块注意力模块的示例性说明,在其他实施例中,也可以采用其他的结构对降维后的第四特征图进行提纯处理。
S4023:利用提纯后的特征图确定输入图像的关键点的位置。
在获得提纯后特征图之后,可以利用该特征图获取关键点的位置信息,例如可以将该提纯后的特征图输入至3*3的卷积模块,来预测输入图像中各关键点的位置信息。其中,在输入图像为面部图像时,预测的关键点可以为17个关键点的位置,比如可以包括对于左右眼睛、鼻子、左右耳朵、左右肩膀、左右手肘、左右手腕、左右胯部、左右膝盖、左右脚踝的位置。在其他的实施例中,也可以获取其他关键点的位置,本公开实施例对此不进行限定。
基于上述配置,即可以通过第一金字塔神经网络的正向处理以及第二金字塔神经网络的反向处理更充分的融合特征,从而提高关键点的检测精度。
在本公开实施例中,还可以执行对于第一金字塔神经网络以及第二金字塔神经网络的训练,从而使得正向处理和反向处理满足工作精度。其中,图10示出根据本公开实施例的一种关键点检测方法中的训练第一金字塔神经网络的流程图。其中,本公开实施例可以利用训练图像数据集训练所述第一金字塔神经网络,其包括:
S501:利用第一金字塔神经网络对所述训练图像数据集中各图像对应的第一特征图进行所述正向处理,得到所述训练图像数据集中各图像对应的第二特征图。
本公开实施例中,可以将训练图像数据集输入至第一金字塔神经网络进行训练。其中,训练图像数据集中可以包括多个图像以及与图像对应的关键点的真实位置。利用第一金字塔网络可以执行如上所述步骤S100和S200(多尺度第一特征图的提取以及正向处理),得到各图像的第二特征图。
S502:利用各第二特征图确定识别的关键点。
在步骤S201之后,可以利用得到的第二特征图识别训练图像的关键点,获得训练图像的各关键点的第一位置。
S503:根据第一损失函数得到所述关键点的第一损失。
S504:利用所述第一损失值反向调节所述第一金字塔神经网络中的各卷积核,直至训练次数达到设定的第一次数阈值。
对应的,在得到各关键点的第一位置之后,可以得到该预测得到的第一位置对应的第一损失。在训练的过程中,可以根据每次训练得到的第一损失反向调节第一金字塔神经网络的参数,例如卷积核的参数,直到训练次数达到第一次数阈值,该第一次数阈值可以根据需求进行设定,一般为大于120的数值,例如本公开实施例中第一次数阈值可以为140。
其中,第一位置对应的第一损失可以为将第一位置与真实位置之间的第一差值输入至第一损失函数获得的损失值,其中第一损失函数可以为对数损失函数。或者也可以是将第一位置和真实位置输入至第一损失函数,获得对应的第一损失。本公开实施例对此不进行限定。基于上述即可以实现第一金字塔神经网络的训练过程,实现第一金字塔神经网络参数的优化。
另外,对应的,图11示出根据本公开实施例的一种关键点检测方法中的训练第二金字塔神经网络的流程图。其中,本公开实施例可以利用训练图像数据集训练所述第二金字塔神经网络,其包括:
S601:利用第二金字塔神经网络对所述第一金字塔神经网络输出的关于训练图像数据集中各图像对应的第二特征图进行所述反向处理,得到所述训练图像数据集中各图像对应的第三特征图。
S602:利用各第三特征图识别关键点。
本公开实施例中,可以首先利用第一金字塔神经网络获得训练数据集中各图像的第二特征图,而后通过第二金字塔神经网络对所述训练图像数据集中各图像对应的第二特征图进行上述的反向处理,得到所述训练图像数据集中各图像对应的第三特征图,而后利用第三特征图预测对应的图像的关键点的第二位置。
S603:根据第二损失函数得到识别的关键点的第二损失。
S604:利用所述第二损失反向调节所述第二金字塔神经网络中卷积核,直至训练次数达到设定的第二次数阈值,或者利用所述第二损失反向调节所述第一金字塔网络中的卷积核以及第二金字塔神经网络中的卷积核,直至训练次数达到设定的第二次数阈值。
对应的,在得到各关键点的第二位置之后可以得到该预测得到的第二位置对应的第二损失。在训练的过程中,可以根据每次训练得到的第二损失反向调节第二金字塔神经网络的参数,例如卷积核的参数,直到训练次数达到第二次数阈值,该第二次数阈值可以根据需求进行设定,一般为大于120的数值,例如本公开实施例中第二次数阈值可以为140。
其中,第二位置对应的第二损失可以为将第二位置与真实位置之间的第二差值输入至第二损失函数获得的损失值,其中第二损失函数可以为对数损失函数。或者也可以是将第二位置和真实位置输入至第二损失函数,获得对应的第二损失值。本公开实施例对此不进行限定。
在本公开的另一些实施例中,在训练第二金字塔神经网络的同时,还可以同时进一步优化训练第一金字塔神经网络,即本公开实施例中,步骤S604时,可以利用获得的第二损失值同时反向调节第一金字塔神经网络中的卷积核的参数以及第二金字塔神经网络汇中的卷积核参数。从而实现整个网络模型的进一步优化。
基于上述即可以实现第二金字塔神经网络的训练过程,实现第一金字塔神经网络的优化。
另外,在本公开实施例中,步骤S400可以通过特征提取网络模型来实现,其中,本公开实施例还可以执行特征提取网络模型的优化过程,其中,图12示出根据本公开实施例的一种关键点检测方法中的训练特征提取网络模型的流程图,其中,利用训练图像数据集训练所述特征提取网络模型,可以包括:
S701:利用特征提取网络模型对所述第二金字塔神经网络输出的关于训练图像数据集中各图像对应的第三特征图进行所述特征融合处理,并利用特征融合处理后的特征图识别所述训练图像数据集中各图像的关键点。
本公开实施例中,可以将与图像训练数据集对应的经第一金字塔神经网络正向处理以及经第二金字塔神经网络处理得到的第三特征图输入至特征提取网络模型,并通过特征提取网络模型执行特征融合,以及提纯等处理得到训练图像数据集中的各图像的关键点的第三位置。
S702:根据第三损失函数得到各关键点的第三损失。
S703:利用所述第三损失值反向调节所述特征提取网络的参数,直至训练次数达到设定的第三 次数阈值,或者利用所述第三损失函数反向调节所述第一金字塔神经网络中的卷积核参数、第二金字塔神经网络中的卷积核参数,以及所述特征提取网络的参数,直至训练次数达到设定的第三次数阈值。
对应的在得到各关键点的第三位置之后可以得到该预测得到的第三位置对应的第三损失值。在训练的过程中,可以根据每次训练得到的第三损失反向调节特征提取网络模型的参数,例如卷积核的参数,或者上述池化等过程的各参数,直到训练次数达到第三次数阈值,该第三次数阈值可以根据需求进行设定,一般为大于120的数值,例如本公开实施例中第三次数阈值可以为140。
其中,第三位置对应的第三损失可以为将第三位置与真实位置之间的第三差值输入至第一损失函数获得的损失值,其中第三损失函数可以为对数损失函数。或者也可以是将第三位置和真实位置输入至第三损失函数,获得对应的第三损失值。本公开实施例对此不进行限定。
基于上述即可以实现特征提取网络模型的训练过程,实现特征提取网络模型参数的优化。
在本公开的另一些实施例中,在训练特征提取网络的同时,还可以同时进一步优化训练第一金字塔神经网络和第二金字塔神经网络,即本公开实施例中,步骤S703时,可以利用获得的第三损失值同时反向调节第一金字塔神经网络中的卷积核的参数、第二金字塔神经网络汇中的卷积核参数,以及特征提取网络模型的参数,从而实现整个网络模型的进一步优化。
综上所述,本公开实施例提出了一种利用双向金字塔网络模型来执行关键点特征检测,其中不仅利用正向处理的方式得到多尺度特征,同时还利用反向处理融合更多的特征,从而能够进一步提高关键点的检测精度。
本领域技术人员可以理解,在具体实施方式的上述方法中,各步骤的撰写顺序并不意味着严格的执行顺序而对实施过程构成任何限定,各步骤的具体执行顺序应当以其功能和可能的内在逻辑确定。
可以理解,本公开提及的上述各个方法实施例,在不违背原理逻辑的情况下,均可以彼此相互结合形成结合后的实施例,限于篇幅,本公开不再赘述。
此外,本公开还提供了关键点检测装置、电子设备、计算机可读存储介质、程序,上述均可用来实现本公开提供的任一种关键点检测方法,相应技术方案和描述和参见方法部分的相应记载,不再赘述。
图13示出根据本公开实施例的关键点检测装置的框图,如图13所示,所述关键点检测装置包括:
多尺度特征获取模块10,配置为获得针对输入图像的多个尺度的第一特征图,各第一特征图的尺度成倍数关系;正向处理模块20,配置为利用第一金字塔神经网络对各所述第一特征图进行正向处理得到与各个所述第一特征图一一对应的第二特征图,其中,所述第二特征图与其一一对应的所述第一特征图的尺度相同;反向处理模块30,配置为利用第二金字塔神经网络对各个所述第二特征图进行反向处理得到与各个所述第二特征图一一对应的第三特征图,其中,所述第三特征图与其一一对应的所述第二特征图的尺度相同;关键点检测模块40,配置为对各所述第三特征图进行特征融合处理,并利用特征融合处理后的特征图获得所述输入图像中的各关键点的位置。
在一些可能的实施方式中,所述多尺度特征获取模块,配置为将所述输入图像调整为预设规格的第一图像,并将所述第一图像输入至残差神经网络,对第一图像执行不同采样频率的降采样处理得到多个不同尺度的第一特征图。
在一些可能的实施方式中,所述正向处理包括第一卷积处理和第一线性插值处理,所述反向处理包括第二卷积处理和第二线性插值处理。
在一些可能的实施方式中,所述正向处理模块,配置为利用第一卷积核对第一特征图C 1...C n中的第一特征图C n进行卷积处理,获得与第一特征图C n对应的第二特征图F n,其中n表示第一特征图的数量,以及n为大于1的整数;以及对所述第二特征图F n执行线性插值处理获得与第二特征图F n对应的第一中间特征图F′ n,其中第一中间特征图F′ n的尺度与第一特征图C n-1的尺度相同;以及利用第二卷积核对第一特征图C n以外的各第一特征图C 1...C n-1进行卷积处理,得到分别与第一特征图C 1...C n-1一一对应的第二中间特征图C' 1...C' n-1,其中所述第二中间特征图的尺度与和其一一对应的第一特征图的尺度相同;并且基于所述第二特征图F n以及各所述第二中间特征图C' 1...C' n-1,得到第二特征图F 1...F n-1以及第一中间特征图F′ 1...F′ n-1,其中所述第二特征图F i由 所述第二中间特征图C′ i与所述第一中间特征图F′ i+1进行叠加处理得到,第一中间特征图F′ i由对应的第二特征图F i经线性插值得到,并且,所述第二中间特征图C′ i与第一中间特征图F′ i+1的尺度相同,其中,i为大于或者等于1且小于n的整数。
在一些可能的实施方式中,所述反向处理模块,配置为利用第三卷积核对第二特征图F 1...F m中的第二特征图F 1进行卷积处理,获得与第二特征图F 1对应的第三特征图R 1,其中m表示第二特征图的数量,以及m为大于1的整数;以及利用第四卷积核对第二特征图F 2...F m进行卷积处理,分别得到对应的第三中间特征图F″ 2...F″ m,其中,第三中间特征图的尺度与对应的第二特征图的尺度相同;以及利用第五卷积核对第三特征图R 1进行卷积处理得到与第三特征图R 1对应的第四中间特征图R' 1;并且利用各第三中间特征图F″ 2...F″ m以及第四中间特征图R' 1,得到第三特征图R 2...R m以及第四中间特征图R' 2...R' m,其中,第三特征图R j由第三中间特征图F″ j与第四中间特征图R' j-1的叠加处理得到,第四中间特征图R' j-1由对应的第三特征图R j-1通过第五卷积核卷积处理获得,其中j为大于1且小于或者等于m。
在一些可能的实施方式中,所述关键点检测模块,配置为对各第三特征图进行特征融合处理,得到第四特征图,并基于所述第四特征图获得所述输入图像中各关键点的位置。
在一些可能的实施方式中,所述关键点检测模块,配置为利用线性插值的方式,将各第三特征图调整为尺度相同的特征图,并对所述尺度相同的特征图进行连接得到所述第四特征图。
在一些可能的实施方式中,所述装置还包括:优化模块,配置为将第一组第三特征图分别输入至不同的瓶颈区块结构中进行卷积处理,分别得到更新后的第三特征图,各所述瓶颈区块结构中包括不同数量的卷积模块,其中,所述第三特征图包括第一组第三特征图和第二组第三特征图,所述第一组第三特征图和所述第二组第三特征图中均包括至少一个第三特征图。
在一些可能的实施方式中,所述关键点检测模块还配置为利用线性插值的方式,将各所述更新后的第三特征图以及所述第二组第三特征图,调整为尺度相同的特征图,并对所述尺度相同的特征图进行连接得到所述第四特征图。
在一些可能的实施方式中,所述关键点检测模块还配置为利用第五卷积核对所述第四特征图进行降维处理,并利用降维处理后的第四特征图确定输入图像的关键点的位置。
在一些可能的实施方式中,所述关键点检测模块还配置为利用第五卷积核对所述第四特征图进行降维处理,利用卷积块注意力模块对降维处理后的第四特征图中的特征进行提纯处理,得到提纯后的特征图,并利用提纯后的特征图确定所述输入图像的关键点的位置。
在一些可能的实施方式中,所述正向处理模块还配置为利用训练图像数据集训练所述第一金字塔神经网络,其包括:利用第一金字塔神经网络对所述训练图像数据集中各图像对应的第一特征图进行所述正向处理,得到所述训练图像数据集中各图像对应的第二特征图;利用各第二特征图确定识别的关键点;根据第一损失函数得到所述关键点的第一损失;利用所述第一损失反向调节所述第一金字塔神经网络中的各卷积核,直至训练次数达到设定的第一次数阈值。
在一些可能的实施方式中,所述反向处理模块还配置为利用训练图像数据集训练所述第二金字塔神经网络,其包括:利用第二金字塔神经网络对所述第一金字塔神经网络输出的关于训练图像数据集中各图像对应的第二特征图进行所述反向处理,得到所述训练图像数据集中各图像对应的第三特征图;利用各第三特征图确定识别的关键点;根据第二损失函数得到识别的各关键点的第二损失;利用所述第二损失反向调节所述第二金字塔神经网络中卷积核,直至训练次数达到设定的第二次数阈值;或者,利用所述第二损失反向调节所述第一金字塔网络中的卷积核以及第二金字塔神经网络中的卷积核,直至训练次数达到设定的第二次数阈值。
在一些可能的实施方式中,所述关键点检测模块还配置为通过特征提取网络执行所述对各所述第三特征图进行特征融合处理,并且在通过特征提取网络执行所述对各所述第三特征图进行特征融合处理之前,还利用训练图像数据集训练所述特征提取网络,其包括:利用特征提取网络对所述第二金字塔神经网络输出的关于训练图像数据集中各图像对应的第三特征图进行所述特征融合处理,并利用特征融合处理后的特征图识别所述训练图像数据集中各图像的关键点;根据第三损失函数得到各关键点的第三损失;利用所述第三损失值反向调节所述特征提取网络的参数,直至训练次数达到设定的第三次数阈值;或者,利用所述第三损失函数反向调节所述第一金字塔神经网络中的卷积 核参数、第二金字塔神经网络中的卷积核参数,以及所述特征提取网络的参数,直至训练次数达到设定的第三次数阈值。
在一些实施例中,本公开实施例提供的装置具有的功能或包含的模块可以用于执行上文方法实施例描述的方法,其具体实现可以参照上文方法实施例的描述,为了简洁,这里不再赘述
本公开实施例还提出一种计算机可读存储介质,其上存储有计算机程序指令,所述计算机程序指令被处理器执行时实现上述方法。计算机可读存储介质可以是非易失性计算机可读存储介质。
本公开实施例还提出一种电子设备,包括:处理器;用于存储处理器可执行指令的存储器;其中,所述处理器被配置为上述方法。
电子设备可以被提供为终端、服务器或其它形态的设备。
图14示出根据本公开实施例的一种电子设备800的框图。例如,电子设备800可以是移动电话,计算机,数字广播终端,消息收发设备,游戏控制台,平板设备,医疗设备,健身设备,个人数字助理等终端。
参照图14,电子设备800可以包括以下一个或多个组件:处理组件802,存储器804,电源组件806,多媒体组件808,音频组件810,输入/输出(I/O)的接口812,传感器组件814,以及通信组件816。
处理组件802通常控制电子设备800的整体操作,诸如与显示,电话呼叫,数据通信,相机操作和记录操作相关联的操作。处理组件802可以包括一个或多个处理器820来执行指令,以完成上述的方法的全部或部分步骤。此外,处理组件802可以包括一个或多个模块,便于处理组件802和其他组件之间的交互。例如,处理组件802可以包括多媒体模块,以方便多媒体组件808和处理组件802之间的交互。
存储器804被配置为存储各种类型的数据以支持在电子设备800的操作。这些数据的示例包括用于在电子设备800上操作的任何应用程序或方法的指令,联系人数据,电话簿数据,消息,图片,视频等。存储器804可以由任何类型的易失性或非易失性存储设备或者它们的组合实现,如静态随机存取存储器(SRAM),电可擦除可编程只读存储器(EEPROM),可擦除可编程只读存储器(EPROM),可编程只读存储器(PROM),只读存储器(ROM),磁存储器,快闪存储器,磁盘或光盘。
电源组件806为电子设备800的各种组件提供电力。电源组件806可以包括电源管理系统,一个或多个电源,及其他与为电子设备800生成、管理和分配电力相关联的组件。
多媒体组件808包括在所述电子设备800和用户之间的提供一个输出接口的屏幕。在一些实施例中,屏幕可以包括液晶显示器(LCD)和触摸面板(TP)。如果屏幕包括触摸面板,屏幕可以被实现为触摸屏,以接收来自用户的输入信号。触摸面板包括一个或多个触摸传感器以感测触摸、滑动和触摸面板上的手势。所述触摸传感器可以不仅感测触摸或滑动动作的边界,而且还检测与所述触摸或滑动操作相关的持续时间和压力。在一些实施例中,多媒体组件808包括一个前置摄像头和/或后置摄像头。当电子设备800处于操作模式,如拍摄模式或视频模式时,前置摄像头和/或后置摄像头可以接收外部的多媒体数据。每个前置摄像头和后置摄像头可以是一个固定的光学透镜系统或具有焦距和光学变焦能力。
音频组件810被配置为输出和/或输入音频信号。例如,音频组件810包括一个麦克风(MIC),当电子设备800处于操作模式,如呼叫模式、记录模式和语音识别模式时,麦克风被配置为接收外部音频信号。所接收的音频信号可以被进一步存储在存储器804或经由通信组件816发送。在一些实施例中,音频组件810还包括一个扬声器,用于输出音频信号。
I/O接口812为处理组件802和外围接口模块之间提供接口,上述外围接口模块可以是键盘,点击轮,按钮等。这些按钮可包括但不限于:主页按钮、音量按钮、启动按钮和锁定按钮。
传感器组件814包括一个或多个传感器,用于为电子设备800提供各个方面的状态评估。例如,传感器组件814可以检测到电子设备800的打开/关闭状态,组件的相对定位,例如所述组件为电子设备800的显示器和小键盘,传感器组件814还可以检测电子设备800或电子设备800一个组件的位置改变,用户与电子设备800接触的存在或不存在,电子设备800方位或加速/减速和电子设备800的温度变化。传感器组件814可以包括接近传感器,被配置用来在没有任何的物理接触时检测附近物体的存在。传感器组件814还可以包括光传感器,如CMOS或CCD图像传感器,用于在成像应用中使用。在一些实施例中,该传感器组件814还可以包括加速度传感器,陀螺仪传感器,磁传感器,压力传感器或温度传感器。
通信组件816被配置为便于电子设备800和其他设备之间有线或无线方式的通信。电子设备800可以接入基于通信标准的无线网络,如WiFi,2G或3G,或它们的组合。在一个示例性实施例中,通信组件816经由广播信道接收来自外部广播管理系统的广播信号或广播相关信息。在一个示例性实施 例中,所述通信组件816还包括近场通信(NFC)模块,以促进短程通信。例如,在NFC模块可基于射频识别(RFID)技术,红外数据协会(IrDA)技术,超宽带(UWB)技术,蓝牙(BT)技术和其他技术来实现。
在示例性实施例中,电子设备800可以被一个或多个应用专用集成电路(ASIC)、数字信号处理器(DSP)、数字信号处理设备(DSPD)、可编程逻辑器件(PLD)、现场可编程门阵列(FPGA)、控制器、微控制器、微处理器或其他电子元件实现,用于执行上述方法。
在示例性实施例中,还提供了一种非易失性计算机可读存储介质,例如包括计算机程序指令的存储器804,上述计算机程序指令可由电子设备800的处理器820执行以完成上述方法。
图15示出根据本公开实施例的一种电子设备1900的框图。例如,电子设备1900可以被提供为一服务器。参照图15,电子设备1900包括处理组件1922,其进一步包括一个或多个处理器,以及由存储器1932所代表的存储器资源,用于存储可由处理组件1922的执行的指令,例如应用程序。存储器1932中存储的应用程序可以包括一个或一个以上的每一个对应于一组指令的模块。此外,处理组件1922被配置为执行指令,以执行上述方法。
电子设备1900还可以包括一个电源组件1926被配置为执行电子设备1900的电源管理,一个有线或无线网络接口1950被配置为将电子设备1900连接到网络,和一个输入输出(I/O)接口1958。电子设备1900可以操作基于存储在存储器1932的操作系统,例如Windows ServerTM,Mac OS XTM,UnixTM,LinuxTM,FreeBSDTM或类似。
在示例性实施例中,还提供了一种非易失性计算机可读存储介质,例如包括计算机程序指令的存储器1932,上述计算机程序指令可由电子设备1900的处理组件1922执行以完成上述方法。
本公开可以是系统、方法和/或计算机程序产品。计算机程序产品可以包括计算机可读存储介质,其上载有用于使处理器实现本公开的各个方面的计算机可读程序指令。
计算机可读存储介质可以是可以保持和存储由指令执行设备使用的指令的有形设备。计算机可读存储介质例如可以是――但不限于――电存储设备、磁存储设备、光存储设备、电磁存储设备、半导体存储设备或者上述的任意合适的组合。计算机可读存储介质的更具体的例子(非穷举的列表)包括:便携式计算机盘、硬盘、随机存取存储器(RAM)、只读存储器(ROM)、可擦式可编程只读存储器(EPROM或闪存)、静态随机存取存储器(SRAM)、便携式压缩盘只读存储器(CD-ROM)、数字多功能盘(DVD)、记忆棒、软盘、机械编码设备、例如其上存储有指令的打孔卡或凹槽内凸起结构、以及上述的任意合适的组合。这里所使用的计算机可读存储介质不被解释为瞬时信号本身,诸如无线电波或者其他自由传播的电磁波、通过波导或其他传输媒介传播的电磁波(例如,通过光纤电缆的光脉冲)、或者通过电线传输的电信号。
这里所描述的计算机可读程序指令可以从计算机可读存储介质下载到各个计算/处理设备,或者通过网络、例如因特网、局域网、广域网和/或无线网下载到外部计算机或外部存储设备。网络可以包括铜传输电缆、光纤传输、无线传输、路由器、防火墙、交换机、网关计算机和/或边缘服务器。每个计算/处理设备中的网络适配卡或者网络接口从网络接收计算机可读程序指令,并转发该计算机可读程序指令,以供存储在各个计算/处理设备中的计算机可读存储介质中。
用于执行本公开操作的计算机程序指令可以是汇编指令、指令集架构(ISA)指令、机器指令、机器相关指令、微代码、固件指令、状态设置数据、或者以一种或多种编程语言的任意组合编写的源代码或目标代码,所述编程语言包括面向对象的编程语言—诸如Smalltalk、C++等,以及常规的过程式编程语言—诸如“C”语言或类似的编程语言。计算机可读程序指令可以完全地在用户计算机上执行、部分地在用户计算机上执行、作为一个独立的软件包执行、部分在用户计算机上部分在远程计算机上执行、或者完全在远程计算机或服务器上执行。在涉及远程计算机的情形中,远程计算机可以通过任意种类的网络—包括局域网(LAN)或广域网(WAN)—连接到用户计算机,或者,可以连接到外部计算机(例如利用因特网服务提供商来通过因特网连接)。在一些实施例中,通过利用计算机可读程序指令的状态信息来个性化定制电子电路,例如可编程逻辑电路、现场可编程门阵列(FPGA)或可编程逻辑阵列(PLA),该电子电路可以执行计算机可读程序指令,从而实现本公开的各个方面。
这里参照根据本公开实施例的方法、装置(系统)和计算机程序产品的流程图和/或框图描述了本公开的各个方面。应当理解,流程图和/或框图的每个方框以及流程图和/或框图中各方框的组合,都可以由计算机可读程序指令实现。
这些计算机可读程序指令可以提供给通用计算机、专用计算机或其它可编程数据处理装置的处理器,从而生产出一种机器,使得这些指令在通过计算机或其它可编程数据处理装置的处理器执行时,产生了实现流程图和/或框图中的一个或多个方框中规定的功能/动作的装置。也可以把这些计算 机可读程序指令存储在计算机可读存储介质中,这些指令使得计算机、可编程数据处理装置和/或其他设备以特定方式工作,从而,存储有指令的计算机可读介质则包括一个制造品,其包括实现流程图和/或框图中的一个或多个方框中规定的功能/动作的各个方面的指令。
也可以把计算机可读程序指令加载到计算机、其它可编程数据处理装置、或其它设备上,使得在计算机、其它可编程数据处理装置或其它设备上执行一系列操作步骤,以产生计算机实现的过程,从而使得在计算机、其它可编程数据处理装置、或其它设备上执行的指令实现流程图和/或框图中的一个或多个方框中规定的功能/动作。
附图中的流程图和框图显示了根据本公开的多个实施例的系统、方法和计算机程序产品的可能实现的体系架构、功能和操作。在这点上,流程图或框图中的每个方框可以代表一个模块、程序段或指令的一部分,所述模块、程序段或指令的一部分包含一个或多个用于实现规定的逻辑功能的可执行指令。在有些作为替换的实现中,方框中所标注的功能也可以以不同于附图中所标注的顺序发生。例如,两个连续的方框实际上可以基本并行地执行,它们有时也可以按相反的顺序执行,这依所涉及的功能而定。也要注意的是,框图和/或流程图中的每个方框、以及框图和/或流程图中的方框的组合,可以用执行规定的功能或动作的专用的基于硬件的系统来实现,或者可以用专用硬件与计算机指令的组合来实现。
以上已经描述了本公开的各实施例,上述说明是示例性的,并非穷尽性的,并且也不限于所披露的各实施例。在不偏离所说明的各实施例的范围和精神的情况下,对于本技术领域的普通技术人员来说许多修改和变更都是显而易见的。本文中所用术语的选择,旨在最好地解释各实施例的原理、实际应用或对市场中的技术改进,或者使本技术领域的其它普通技术人员能理解本文披露的各实施例。

Claims (30)

  1. 一种关键点检测方法,包括:
    获得针对输入图像的多个尺度的第一特征图,各第一特征图的尺度成倍数关系;
    利用第一金字塔神经网络对各所述第一特征图进行正向处理得到与各个所述第一特征图一一对应的第二特征图,其中,所述第二特征图与其一一对应的所述第一特征图的尺度相同;
    利用第二金字塔神经网络对各个所述第二特征图进行反向处理得到与各个所述第二特征图一一对应的第三特征图,其中,所述第三特征图与其一一对应的所述第二特征图的尺度相同;
    对各所述第三特征图进行特征融合处理,并利用特征融合处理后的特征图获得所述输入图像中的各关键点的位置。
  2. 根据权利要求1所述的方法,其中,所述获得针对输入图像的多个尺度的第一特征图包括:
    将所述输入图像调整为预设规格的第一图像;
    将所述第一图像输入至残差神经网络,对第一图像执行不同采样频率的降采样处理得到多个不同尺度的第一特征图。
  3. 根据权利要求1所述的方法,其中,所述正向处理包括第一卷积处理和第一线性插值处理,所述反向处理包括第二卷积处理和第二线性插值处理。
  4. 根据权利要求1-3中任意一项所述的方法,其中,所述利用第一金字塔神经网络对各所述第一特征图进行正向处理得到与各个所述第一特征图一一对应的第二特征图,包括:
    利用第一卷积核对第一特征图C 1...C n中的第一特征图C n进行卷积处理,获得与第一特征图C n对应的第二特征图F n,其中n表示第一特征图的数量,以及n为大于1的整数;
    对所述第二特征图F n执行线性插值处理获得与第二特征图F n对应的第一中间特征图F′ n,其中第一中间特征图F′ n的尺度与第一特征图C n-1的尺度相同;
    利用第二卷积核对第一特征图C n以外的各第一特征图C 1...C n-1进行卷积处理,得到分别与第一特征图C 1...C n-1一一对应的第二中间特征图C′ 1...C′ n-1,其中所述第二中间特征图的尺度与和其一一对应的第一特征图的尺度相同;
    基于所述第二特征图F n以及各所述第二中间特征图C′ 1...C′ n-1,得到第二特征图F 1...F n-1以及第一中间特征图F′ 1...F′ n-1,其中所述第二特征图F i由所述第二中间特征图C′ i与所述第一中间特征图F′ i+1进行叠加处理得到,第一中间特征图F′ i由对应的第二特征图F i经线性插值得到,并且,所述第二中间特征图C′ i与第一中间特征图F′ i+1的尺度相同,其中,i为大于或者等于1且小于n的整数。
  5. 根据权利要求1-4中任意一项所述的方法,其中,所述利用第二金字塔神经网络对各个所述第二特征图进行反向处理得到与各个所述第二特征图一一对应的第三特征图,包括:
    利用第三卷积核对第二特征图F 1...F m中的第二特征图F 1进行卷积处理,获得与第二特征图F 1对应的第三特征图R 1,其中m表示第二特征图的数量,以及m为大于1的整数;
    利用第四卷积核对第二特征图F 2...F m进行卷积处理,分别得到对应的第三中间特征图F″ 2...F″ m,其中,第三中间特征图的尺度与对应的第二特征图的尺度相同;
    利用第五卷积核对第三特征图R 1进行卷积处理得到与第三特征图R 1对应的第四中间特征图R′ 1
    利用各第三中间特征图F″ 2...F″ m以及第四中间特征图R′ 1,得到第三特征图R 2...R m以及第四中间特征图R′ 2...R′ m,其中,第三特征图R j由第三中间特征图F″ j与第四中间特征图R′ j-1的叠加处理得到,第四中间特征图R′ j-1由对应的第三特征图R j-1通过第五卷积核卷积处理获得,其中j为大于1且小于或者等于m。
  6. 根据权利要求1-5中任意一项所述的方法,其中,所述对各所述第三特征图进行特征融合处理,并利用特征融合处理后的特征图获得所述输入图像中的各关键点的位置,包括:
    对各第三特征图进行特征融合处理,得到第四特征图:
    基于所述第四特征图获得所述输入图像中各关键点的位置。
  7. 根据权利要求6所述的方法,其中,所述对各第三特征图进行特征融合处理,得到第四特征图,包括:利用线性插值的方式,将各第三特征图调整为尺度相同的特征图;
    对所述尺度相同的特征图进行连接得到所述第四特征图。
  8. 根据权利要求6或7所述的方法,其中,在所述对各第三特征图进行特征融合处理,得到第四特征图之前,还包括:将第一组第三特征图分别输入至不同的瓶颈区块结构中进行卷积处理,分别得到更新后的第三特征图,各所述瓶颈区块结构中包括不同数量的卷积模块,其中,所述第三特征图包括第一组第三特征图和第二组第三特征图,所述第一组第三特征图和所述第二组第三特征图中均包括至少一个第三特征图。
  9. 根据权利要求8所述的方法,其中,所述对各第三特征图进行特征融合处理,得到第四特征图,包括:利用线性插值的方式,将各所述更新后的第三特征图以及所述第二组第三特征图,调整为尺度相同的特征图;
    对所述尺度相同的特征图进行连接得到所述第四特征图。
  10. 根据权利要求6-9中任意一项所述的方法,其中,所述基于所述第四特征图获得所述输入图像中各关键点的位置,包括:利用第五卷积核对所述第四特征图进行降维处理;
    利用降维处理后的第四特征图确定输入图像的关键点的位置。
  11. 根据权利要求6-9中任意一项所述的方法,其中,所述基于所述第四特征图获得所述输入图像中各关键点的位置,包括:利用第五卷积核对所述第四特征图进行降维处理;
    利用卷积块注意力模块对降维处理后的第四特征图中的特征进行提纯处理,得到提纯后的特征图;
    利用提纯后的特征图确定所述输入图像的关键点的位置。
  12. 根据权利要求1-11中任意一项所述的方法,其中,所述方法还包括利用训练图像数据集训练所述第一金字塔神经网络,其包括:利用第一金字塔神经网络对所述训练图像数据集中各图像对应的第一特征图进行所述正向处理,得到所述训练图像数据集中各图像对应的第二特征图;
    利用各第二特征图确定识别的关键点;
    根据第一损失函数得到所述关键点的第一损失;
    利用所述第一损失反向调节所述第一金字塔神经网络中的各卷积核,直至训练次数达到设定的第一次数阈值。
  13. 根据权利要求1-12中任意一项所述的方法,其中,所述方法还包括利用训练图像数据集训练所述第二金字塔神经网络,其包括:
    利用第二金字塔神经网络对所述第一金字塔神经网络输出的关于训练图像数据集中各图像对应的第二特征图进行所述反向处理,得到所述训练图像数据集中各图像对应的第三特征图;
    利用各第三特征图确定识别的关键点;
    根据第二损失函数得到识别的各关键点的第二损失;
    利用所述第二损失反向调节所述第二金字塔神经网络中卷积核,直至训练次数达到设定的第二次数阈值;或者,
    利用所述第二损失反向调节所述第一金字塔网络中的卷积核以及第二金字塔神经网络中的卷积核,直至训练次数达到设定的第二次数阈值。
  14. 根据权利要求1-13中任意一项所述的方法,其中,通过特征提取网络执行所述对各所述第三特征图进行特征融合处理,并且,
    在通过特征提取网络执行所述对各所述第三特征图进行特征融合处理之前,所述方法还包括:利用训练图像数据集训练所述特征提取网络,其包括:
    利用特征提取网络对所述第二金字塔神经网络输出的关于训练图像数据集中各图像对应的第三特征图进行所述特征融合处理,并利用特征融合处理后的特征图识别所述训练图像数据集中各图像的关键点;
    根据第三损失函数得到各关键点的第三损失;
    利用所述第三损失值反向调节所述特征提取网络的参数,直至训练次数达到设定的第三次数阈值;或者,利用所述第三损失函数反向调节所述第一金字塔神经网络中的卷积核参数、第二金字塔 神经网络中的卷积核参数,以及所述特征提取网络的参数,直至训练次数达到设定的第三次数阈值。
  15. 一种关键点检测装置,包括:
    多尺度特征获取模块,配置为获得针对输入图像的多个尺度的第一特征图,各第一特征图的尺度成倍数关系;
    正向处理模块,配置为利用第一金字塔神经网络对各所述第一特征图进行正向处理得到与各个所述第一特征图一一对应的第二特征图,其中,所述第二特征图与其一一对应的所述第一特征图的尺度相同;
    反向处理模块,配置为利用第二金字塔神经网络对各个所述第二特征图进行反向处理得到与各个所述第二特征图一一对应的第三特征图,其中,所述第三特征图与其一一对应的所述第二特征图的尺度相同;
    关键点检测模块,配置为对各所述第三特征图进行特征融合处理,并利用特征融合处理后的特征图获得所述输入图像中的各关键点的位置。
  16. 根据权利要求15所述的装置,其中,所述多尺度特征获取模块,配置为将所述输入图像调整为预设规格的第一图像,并将所述第一图像输入至残差神经网络,对第一图像执行不同采样频率的降采样处理得到多个不同尺度的第一特征图。
  17. 根据权利要求15所述的装置,其中,所述正向处理包括第一卷积处理和第一线性插值处理,所述反向处理包括第二卷积处理和第二线性插值处理。
  18. 根据权利要求15-17中任意一项所述的装置,其中,所述正向处理模块,配置为利用第一卷积核对第一特征图C 1...C n中的第一特征图C n进行卷积处理,获得与第一特征图C n对应的第二特征图F n,其中n表示第一特征图的数量,以及n为大于1的整数;以及
    对所述第二特征图F n执行线性插值处理获得与第二特征图F n对应的第一中间特征图F′ n,其中第一中间特征图F′ n的尺度与第一特征图C n-1的尺度相同;以及
    利用第二卷积核对第一特征图C n以外的各第一特征图C 1...C n-1进行卷积处理,得到分别与第一特征图C 1...C n-1一一对应的第二中间特征图C′ 1...C′ n-1,其中所述第二中间特征图的尺度与和其一一对应的第一特征图的尺度相同;并且
    基于所述第二特征图F n以及各所述第二中间特征图C′ 1...C′ n-1,得到第二特征图F 1...F n-1以及第一中间特征图F′ 1...F′ n-1,其中所述第二特征图F i由所述第二中间特征图C′ i与所述第一中间特征图F′ i+1进行叠加处理得到,第一中间特征图F′ i由对应的第二特征图F i经线性插值得到,并且,所述第二中间特征图C′ i与第一中间特征图F′ i+1的尺度相同,其中,i为大于或者等于1且小于n的整数。
  19. 根据权利要求15-18中任意一项所述的装置,其中,所述反向处理模块,配置为利用第三卷积核对第二特征图F 1...F m中的第二特征图F 1进行卷积处理,获得与第二特征图F 1对应的第三特征图R 1,其中m表示第二特征图的数量,以及m为大于1的整数;以及利用第四卷积核对第二特征图F 2...F m进行卷积处理,分别得到对应的第三中间特征图F″ 2...F″ m,其中,第三中间特征图的尺度与对应的第二特征图的尺度相同;以及利用第五卷积核对第三特征图R 1进行卷积处理得到与第三特征图R 1对应的第四中间特征图R′ 1;并且利用各第三中间特征图F″ 2...F″ m以及第四中间特征图R′ 1,得到第三特征图R 2...R m以及第四中间特征图R′ 2...R′ m,其中,第三特征图R j由第三中间特征图F″ j与第四中间特征图R′ j-1的叠加处理得到,第四中间特征图R′ j-1由对应的第三特征图R j-1通过第五卷积核卷积处理获得,其中j为大于1且小于或者等于m。
  20. 根据权利要求15-19中任意一项所述的装置,其中,所述关键点检测模块,配置为对各第三特征图进行特征融合处理,得到第四特征图,并基于所述第四特征图获得所述输入图像中各关键点的位置。
  21. 根据权利要求20所述的装置,其中,所述关键点检测模块,配置为利用线性插值的方式,将各第三特征图调整为尺度相同的特征图,并对所述尺度相同的特征图进行连接得到所述第四特征 图。
  22. 根据权利要求20或21所述的装置,其中,所述装置还包括:
    优化模块,配置为将第一组第三特征图分别输入至不同的瓶颈区块结构中进行卷积处理,分别得到更新后的第三特征图,各所述瓶颈区块结构中包括不同数量的卷积模块,其中,所述第三特征图包括第一组第三特征图和第二组第三特征图,所述第一组第三特征图和所述第二组第三特征图中均包括至少一个第三特征图。
  23. 根据权利要求22所述的装置,其中,所述关键点检测模块还配置为利用线性插值的方式,将各所述更新后的第三特征图以及所述第二组第三特征图,调整为尺度相同的特征图,并对所述尺度相同的特征图进行连接得到所述第四特征图。
  24. 根据权利要求20-23中任意一项所述的装置,其中,所述关键点检测模块还配置为利用第五卷积核对所述第四特征图进行降维处理,并利用降维处理后的第四特征图确定输入图像的关键点的位置。
  25. 根据权利要求20-23中任意一项所述的装置,其中,所述关键点检测模块还配置为利用第五卷积核对所述第四特征图进行降维处理,利用卷积块注意力模块对降维处理后的第四特征图中的特征进行提纯处理,得到提纯后的特征图,并利用提纯后的特征图确定所述输入图像的关键点的位置。
  26. 根据权利要求15-25中任意一项所述的装置,其中,所述正向处理模块还配置为利用训练图像数据集训练所述第一金字塔神经网络,其包括:利用第一金字塔神经网络对所述训练图像数据集中各图像对应的第一特征图进行所述正向处理,得到所述训练图像数据集中各图像对应的第二特征图;利用各第二特征图确定识别的关键点;根据第一损失函数得到所述关键点的第一损失;利用所述第一损失反向调节所述第一金字塔神经网络中的各卷积核,直至训练次数达到设定的第一次数阈值。
  27. 根据权利要求15-26中任意一项所述的装置,其中,所述反向处理模块还配置为利用训练图像数据集训练所述第二金字塔神经网络,其包括:利用第二金字塔神经网络对所述第一金字塔神经网络输出的关于训练图像数据集中各图像对应的第二特征图进行所述反向处理,得到所述训练图像数据集中各图像对应的第三特征图;利用各第三特征图确定识别的关键点;根据第二损失函数得到识别的各关键点的第二损失;利用所述第二损失反向调节所述第二金字塔神经网络中卷积核,直至训练次数达到设定的第二次数阈值;或者,利用所述第二损失反向调节所述第一金字塔网络中的卷积核以及第二金字塔神经网络中的卷积核,直至训练次数达到设定的第二次数阈值。
  28. 根据权利要求15-27中任意一项所述的装置,其中,所述关键点检测模块还配置为通过特征提取网络执行所述对各所述第三特征图进行特征融合处理,并且在通过特征提取网络执行所述对各所述第三特征图进行特征融合处理之前,还利用训练图像数据集训练所述特征提取网络,其包括:利用特征提取网络对所述第二金字塔神经网络输出的关于训练图像数据集中各图像对应的第三特征图进行所述特征融合处理,并利用特征融合处理后的特征图识别所述训练图像数据集中各图像的关键点;根据第三损失函数得到各关键点的第三损失;利用所述第三损失值反向调节所述特征提取网络的参数,直至训练次数达到设定的第三次数阈值;或者,利用所述第三损失函数反向调节所述第一金字塔神经网络中的卷积核参数、第二金字塔神经网络中的卷积核参数,以及所述特征提取网络的参数,直至训练次数达到设定的第三次数阈值。
  29. 一种电子设备,包括:
    处理器;
    用于存储处理器可执行指令的存储器;
    其中,所述处理器被配置为:执行权利要求1至14中任意一项所述的方法。
  30. 一种计算机可读存储介质,其上存储有计算机程序指令,其中,所述计算机程序指令被处理器执行时实现权利要求1至14中任意一项所述的方法。
PCT/CN2019/083721 2018-11-16 2019-04-22 关键点检测方法及装置、电子设备和存储介质 WO2020098225A1 (zh)

Priority Applications (4)

Application Number Priority Date Filing Date Title
KR1020207012580A KR102394354B1 (ko) 2018-11-16 2019-04-22 키 포인트 검출 방법 및 장치, 전자 기기 및 저장 매체
JP2020518758A JP6944051B2 (ja) 2018-11-16 2019-04-22 キーポイント検出方法及び装置、電子機器並びに記憶媒体
SG11202003818YA SG11202003818YA (en) 2018-11-16 2019-04-22 Key point detection method and apparatus, electronic device, and storage medium
US16/855,630 US20200250462A1 (en) 2018-11-16 2020-04-22 Key point detection method and apparatus, and storage medium

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201811367869.4 2018-11-16
CN201811367869.4A CN109614876B (zh) 2018-11-16 2018-11-16 关键点检测方法及装置、电子设备和存储介质

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US16/855,630 Continuation US20200250462A1 (en) 2018-11-16 2020-04-22 Key point detection method and apparatus, and storage medium

Publications (1)

Publication Number Publication Date
WO2020098225A1 true WO2020098225A1 (zh) 2020-05-22

Family

ID=66003175

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/083721 WO2020098225A1 (zh) 2018-11-16 2019-04-22 关键点检测方法及装置、电子设备和存储介质

Country Status (7)

Country Link
US (1) US20200250462A1 (zh)
JP (1) JP6944051B2 (zh)
KR (1) KR102394354B1 (zh)
CN (7) CN113591755B (zh)
SG (1) SG11202003818YA (zh)
TW (1) TWI720598B (zh)
WO (1) WO2020098225A1 (zh)

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111695519A (zh) * 2020-06-12 2020-09-22 北京百度网讯科技有限公司 关键点定位方法、装置、设备以及存储介质
CN111709945A (zh) * 2020-07-17 2020-09-25 成都三零凯天通信实业有限公司 一种基于深度局部特征的视频拷贝检测方法
CN111784642A (zh) * 2020-06-10 2020-10-16 中铁四局集团有限公司 一种图像处理方法、目标识别模型训练方法和目标识别方法
CN112131925A (zh) * 2020-07-22 2020-12-25 浙江元亨通信技术股份有限公司 一种多通道特征空间金字塔的构造方法
CN112836710A (zh) * 2021-02-23 2021-05-25 浙大宁波理工学院 一种基于特征金字塔网络的房间布局估计获取方法与系统
CN116738296A (zh) * 2023-08-14 2023-09-12 大有期货有限公司 机房状况综合智能监控系统

Families Citing this family (32)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR102227583B1 (ko) * 2018-08-03 2021-03-15 한국과학기술원 딥 러닝 기반의 카메라 캘리브레이션 방법 및 장치
CN113591755B (zh) * 2018-11-16 2024-04-16 北京市商汤科技开发有限公司 关键点检测方法及装置、电子设备和存储介质
JP7103240B2 (ja) * 2019-01-10 2022-07-20 日本電信電話株式会社 物体検出認識装置、方法、及びプログラム
CN110378253B (zh) * 2019-07-01 2021-03-26 浙江大学 一种基于轻量化神经网络的实时关键点检测方法
CN110378976B (zh) * 2019-07-18 2020-11-13 北京市商汤科技开发有限公司 图像处理方法及装置、电子设备和存储介质
CN110705563B (zh) * 2019-09-07 2020-12-29 创新奇智(重庆)科技有限公司 一种基于深度学习的工业零件关键点检测方法
CN110647834B (zh) * 2019-09-18 2021-06-25 北京市商汤科技开发有限公司 人脸和人手关联检测方法及装置、电子设备和存储介质
KR20210062477A (ko) * 2019-11-21 2021-05-31 삼성전자주식회사 전자 장치 및 그 제어 방법
US20220092735A1 (en) * 2019-11-21 2022-03-24 Samsung Electronics Co., Ltd. Electronic apparatus and controlling method thereof
US11080833B2 (en) * 2019-11-22 2021-08-03 Adobe Inc. Image manipulation using deep learning techniques in a patch matching operation
WO2021146890A1 (en) * 2020-01-21 2021-07-29 Beijing Didi Infinity Technology And Development Co., Ltd. Systems and methods for object detection in image using detection model
CN111414823B (zh) * 2020-03-12 2023-09-12 Oppo广东移动通信有限公司 人体特征点的检测方法、装置、电子设备以及存储介质
CN111382714B (zh) * 2020-03-13 2023-02-17 Oppo广东移动通信有限公司 图像检测方法、装置、终端及存储介质
CN111401335B (zh) * 2020-04-29 2023-06-30 Oppo广东移动通信有限公司 一种关键点检测方法及装置、存储介质
CN111709428B (zh) * 2020-05-29 2023-09-15 北京百度网讯科技有限公司 图像中关键点位置的识别方法、装置、电子设备及介质
US11847823B2 (en) 2020-06-18 2023-12-19 Apple Inc. Object and keypoint detection system with low spatial jitter, low latency and low power usage
CN112132011B (zh) * 2020-09-22 2024-04-26 深圳市捷顺科技实业股份有限公司 一种面部识别方法、装置、设备及存储介质
CN112149558A (zh) * 2020-09-22 2020-12-29 驭势科技(南京)有限公司 一种用于关键点检测的图像处理方法、网络和电子设备
CN112232361B (zh) * 2020-10-13 2021-09-21 国网电子商务有限公司 图像处理的方法及装置、电子设备及计算机可读存储介质
CN112364699B (zh) * 2020-10-14 2024-08-02 珠海欧比特宇航科技股份有限公司 基于加权损失融合网络的遥感图像分割方法、装置及介质
CN112257728B (zh) * 2020-11-12 2021-08-17 腾讯科技(深圳)有限公司 图像处理方法、装置、计算机设备以及存储介质
CN112329888B (zh) * 2020-11-26 2023-11-14 Oppo广东移动通信有限公司 图像处理方法、装置、电子设备以及存储介质
CN112434713A (zh) * 2020-12-02 2021-03-02 携程计算机技术(上海)有限公司 图像特征提取方法、装置、电子设备、存储介质
CN112581450B (zh) * 2020-12-21 2024-04-16 北京工业大学 基于膨胀卷积金字塔与多尺度金字塔的花粉检测方法
CN112800834B (zh) * 2020-12-25 2022-08-12 温州晶彩光电有限公司 一种基于跪拜行为识别来定位炫彩射灯的方法及系统
JP2023527615A (ja) * 2021-04-28 2023-06-30 ベイジン バイドゥ ネットコム サイエンス テクノロジー カンパニー リミテッド 目標対象検出モデルのトレーニング方法、目標対象検出方法、機器、電子機器、記憶媒体及びコンピュータプログラム
CN113902903B (zh) * 2021-09-30 2024-08-02 北京工业大学 一种基于下采样的双注意力多尺度融合方法
KR102647320B1 (ko) * 2021-11-23 2024-03-12 숭실대학교산학협력단 객체 추적 장치 및 방법
CN114022657B (zh) * 2022-01-06 2022-05-24 高视科技(苏州)有限公司 一种屏幕缺陷分类方法、电子设备及存储介质
CN114724175B (zh) * 2022-03-04 2024-03-29 亿达信息技术有限公司 行人图像的检测网络、检测方法、训练方法、电子设备和介质
WO2024011281A1 (en) * 2022-07-11 2024-01-18 James Cook University A method and a system for automated prediction of characteristics of aquaculture animals
KR20240083242A (ko) * 2022-12-02 2024-06-12 주식회사 Lg 경영개발원 기계 학습 기반 이상 검출 장치 및 방법

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106339680A (zh) * 2016-08-25 2017-01-18 北京小米移动软件有限公司 人脸关键点定位方法及装置
US9552510B2 (en) * 2015-03-18 2017-01-24 Adobe Systems Incorporated Facial expression capture for character animation
CN108280455A (zh) * 2018-01-19 2018-07-13 北京市商汤科技开发有限公司 人体关键点检测方法和装置、电子设备、程序和介质
CN109614876A (zh) * 2018-11-16 2019-04-12 北京市商汤科技开发有限公司 关键点检测方法及装置、电子设备和存储介质

Family Cites Families (35)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP0486635A1 (en) * 1990-05-22 1992-05-27 International Business Machines Corporation Scalable flow virtual learning neurocomputer
CN101510257B (zh) * 2009-03-31 2011-08-10 华为技术有限公司 一种人脸相似度匹配方法及装置
CN101980290B (zh) * 2010-10-29 2012-06-20 西安电子科技大学 抗噪声环境多聚焦图像融合方法
CN102622730A (zh) * 2012-03-09 2012-08-01 武汉理工大学 基于非降采样Laplacian金字塔和BEMD的遥感图像融合处理方法
CN103049895B (zh) * 2012-12-17 2016-01-20 华南理工大学 基于平移不变剪切波变换的多模态医学图像融合方法
CN103279957B (zh) * 2013-05-31 2015-11-25 北京师范大学 一种基于多尺度特征融合的遥感图像感兴趣区域提取方法
CN103793692A (zh) * 2014-01-29 2014-05-14 五邑大学 低分辨率多光谱掌纹、掌静脉实时身份识别方法与系统
JP6474210B2 (ja) * 2014-07-31 2019-02-27 インターナショナル・ビジネス・マシーンズ・コーポレーションInternational Business Machines Corporation 大規模画像データベースの高速検索手法
WO2016054779A1 (en) * 2014-10-09 2016-04-14 Microsoft Technology Licensing, Llc Spatial pyramid pooling networks for image processing
CN104346607B (zh) * 2014-11-06 2017-12-22 上海电机学院 基于卷积神经网络的人脸识别方法
CN104793620B (zh) * 2015-04-17 2019-06-18 中国矿业大学 基于视觉特征捆绑和强化学习理论的避障机器人
CN104866868B (zh) * 2015-05-22 2018-09-07 杭州朗和科技有限公司 基于深度神经网络的金属币识别方法和装置
US10007863B1 (en) * 2015-06-05 2018-06-26 Gracenote, Inc. Logo recognition in images and videos
CN105184779B (zh) * 2015-08-26 2018-04-06 电子科技大学 一种基于快速特征金字塔的车辆多尺度跟踪方法
CN105912990B (zh) * 2016-04-05 2019-10-08 深圳先进技术研究院 人脸检测的方法及装置
GB2549554A (en) * 2016-04-21 2017-10-25 Ramot At Tel-Aviv Univ Ltd Method and system for detecting an object in an image
US10032067B2 (en) * 2016-05-28 2018-07-24 Samsung Electronics Co., Ltd. System and method for a unified architecture multi-task deep learning machine for object recognition
US10993697B2 (en) * 2016-06-20 2021-05-04 Butterfly Network, Inc. Automated image acquisition for assisting a user to operate an ultrasound device
US10365617B2 (en) * 2016-12-12 2019-07-30 Dmo Systems Limited Auto defect screening using adaptive machine learning in semiconductor device manufacturing flow
CN110475505B (zh) * 2017-01-27 2022-04-05 阿特瑞斯公司 利用全卷积网络的自动分割
CN108229490B (zh) * 2017-02-23 2021-01-05 北京市商汤科技开发有限公司 关键点检测方法、神经网络训练方法、装置和电子设备
CN106934397B (zh) * 2017-03-13 2020-09-01 北京市商汤科技开发有限公司 图像处理方法、装置及电子设备
WO2018169639A1 (en) * 2017-03-17 2018-09-20 Nec Laboratories America, Inc Recognition in unlabeled videos with domain adversarial learning and knowledge distillation
CN108664981B (zh) * 2017-03-30 2021-10-26 北京航空航天大学 显著图像提取方法及装置
CN107194318B (zh) * 2017-04-24 2020-06-12 北京航空航天大学 目标检测辅助的场景识别方法
CN108229281B (zh) * 2017-04-25 2020-07-17 北京市商汤科技开发有限公司 神经网络的生成方法和人脸检测方法、装置及电子设备
CN108229497B (zh) * 2017-07-28 2021-01-05 北京市商汤科技开发有限公司 图像处理方法、装置、存储介质、计算机程序和电子设备
CN107909041A (zh) * 2017-11-21 2018-04-13 清华大学 一种基于时空金字塔网络的视频识别方法
CN108021923B (zh) * 2017-12-07 2020-10-23 上海为森车载传感技术有限公司 一种用于深度神经网络的图像特征提取方法
CN108182384B (zh) * 2017-12-07 2020-09-29 浙江大华技术股份有限公司 一种人脸特征点定位方法及装置
CN108229445A (zh) * 2018-02-09 2018-06-29 深圳市唯特视科技有限公司 一种基于级联金字塔网络的多人姿态估计方法
CN108664885B (zh) * 2018-03-19 2021-08-31 杭州电子科技大学 基于多尺度级联HourGlass网络的人体关键点检测方法
CN108520251A (zh) * 2018-04-20 2018-09-11 北京市商汤科技开发有限公司 关键点检测方法及装置、电子设备和存储介质
CN108596087B (zh) * 2018-04-23 2020-09-15 合肥湛达智能科技有限公司 一种基于双网络结果的驾驶疲劳程度检测回归模型
CN108764133B (zh) * 2018-05-25 2020-10-20 北京旷视科技有限公司 图像识别方法、装置及系统

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9552510B2 (en) * 2015-03-18 2017-01-24 Adobe Systems Incorporated Facial expression capture for character animation
CN106339680A (zh) * 2016-08-25 2017-01-18 北京小米移动软件有限公司 人脸关键点定位方法及装置
CN108280455A (zh) * 2018-01-19 2018-07-13 北京市商汤科技开发有限公司 人体关键点检测方法和装置、电子设备、程序和介质
CN109614876A (zh) * 2018-11-16 2019-04-12 北京市商汤科技开发有限公司 关键点检测方法及装置、电子设备和存储介质

Cited By (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111784642A (zh) * 2020-06-10 2020-10-16 中铁四局集团有限公司 一种图像处理方法、目标识别模型训练方法和目标识别方法
CN111695519B (zh) * 2020-06-12 2023-08-08 北京百度网讯科技有限公司 关键点定位方法、装置、设备以及存储介质
EP3869402A1 (en) * 2020-06-12 2021-08-25 Beijing Baidu Netcom Science And Technology Co., Ltd. Method and apparatus for positioning key point, device, storage medium and computer program product
JP2021197157A (ja) * 2020-06-12 2021-12-27 ベイジン バイドゥ ネットコム サイエンス テクノロジー カンパニー リミテッド キーポイントの特定方法及び装置、機器、記憶媒体
JP7194215B2 (ja) 2020-06-12 2022-12-21 ベイジン バイドゥ ネットコム サイエンス テクノロジー カンパニー リミテッド キーポイントの特定方法及び装置、機器、記憶媒体
US11610389B2 (en) 2020-06-12 2023-03-21 Beijing Baidu Netcom Science And Technology Co., Ltd. Method and apparatus for positioning key point, device, and storage medium
CN111695519A (zh) * 2020-06-12 2020-09-22 北京百度网讯科技有限公司 关键点定位方法、装置、设备以及存储介质
CN111709945A (zh) * 2020-07-17 2020-09-25 成都三零凯天通信实业有限公司 一种基于深度局部特征的视频拷贝检测方法
CN112131925A (zh) * 2020-07-22 2020-12-25 浙江元亨通信技术股份有限公司 一种多通道特征空间金字塔的构造方法
CN112131925B (zh) * 2020-07-22 2024-06-07 随锐科技集团股份有限公司 一种多通道特征空间金字塔的构造方法
CN112836710A (zh) * 2021-02-23 2021-05-25 浙大宁波理工学院 一种基于特征金字塔网络的房间布局估计获取方法与系统
CN116738296A (zh) * 2023-08-14 2023-09-12 大有期货有限公司 机房状况综合智能监控系统
CN116738296B (zh) * 2023-08-14 2024-04-02 大有期货有限公司 机房状况综合智能监控系统

Also Published As

Publication number Publication date
CN113591750A (zh) 2021-11-02
KR20200065033A (ko) 2020-06-08
CN113569798B (zh) 2024-05-24
CN113591754A (zh) 2021-11-02
CN113591755B (zh) 2024-04-16
CN113569796B (zh) 2024-06-11
CN113569798A (zh) 2021-10-29
CN113591755A (zh) 2021-11-02
SG11202003818YA (en) 2020-06-29
CN109614876A (zh) 2019-04-12
CN113591750B (zh) 2024-07-19
CN113569797B (zh) 2024-05-21
CN113569796A (zh) 2021-10-29
US20200250462A1 (en) 2020-08-06
JP6944051B2 (ja) 2021-10-06
TW202020806A (zh) 2020-06-01
TWI720598B (zh) 2021-03-01
CN113591754B (zh) 2022-08-02
KR102394354B1 (ko) 2022-05-04
CN113569797A (zh) 2021-10-29
CN109614876B (zh) 2021-07-27
JP2021508388A (ja) 2021-03-04

Similar Documents

Publication Publication Date Title
WO2020098225A1 (zh) 关键点检测方法及装置、电子设备和存储介质
KR102406354B1 (ko) 비디오 수복 방법 및 장치, 전자 기기 및 기억 매체
JP6916970B2 (ja) ビデオ処理方法及び装置、電子機器並びに記憶媒体
TWI740309B (zh) 圖像處理方法及裝置、電子設備和電腦可讀儲存介質
WO2021051650A1 (zh) 人脸和人手关联检测方法及装置、电子设备和存储介质
WO2020155711A1 (zh) 图像生成方法及装置、电子设备和存储介质
WO2020107813A1 (zh) 图像的描述语句定位方法及装置、电子设备和存储介质
TW202029125A (zh) 圖像處理方法及裝置、電子設備和儲存介質
TWI718631B (zh) 人臉圖像的處理方法及裝置、電子設備和儲存介質
TW202109449A (zh) 影像處理方法、電子設備,和電腦可讀儲存介質
KR102334279B1 (ko) 얼굴 특징점 위치결정 방법 및 장치
TWI719777B (zh) 圖像重建方法、圖像重建裝置、電子設備和電腦可讀儲存媒體
CN110188865B (zh) 信息处理方法及装置、电子设备和存储介质
WO2023142645A1 (zh) 图像处理方法、装置、电子设备、存储介质和计算机程序产品
KR102324001B1 (ko) 위치자세 검출 방법 및 장치, 전자 기기 및 저장 매체
CN110929616B (zh) 一种人手识别方法、装置、电子设备和存储介质
WO2024124913A1 (zh) 实体信息确定方法、装置和设备
CN111046780A (zh) 神经网络训练及图像识别方法、装置、设备和存储介质
CN114821799B (zh) 基于时空图卷积网络的动作识别方法、装置和设备
CN108227927B (zh) 基于vr的产品展示方法、装置及电子设备
CN111753596B (zh) 神经网络的训练方法及装置、电子设备和存储介质
CN114489333A (zh) 图像的处理方法、装置、电子设备及存储介质

Legal Events

Date Code Title Description
ENP Entry into the national phase

Ref document number: 2020518758

Country of ref document: JP

Kind code of ref document: A

ENP Entry into the national phase

Ref document number: 20207012580

Country of ref document: KR

Kind code of ref document: A

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

Ref document number: 19884469

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

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 18/08/2021)

122 Ep: pct application non-entry in european phase

Ref document number: 19884469

Country of ref document: EP

Kind code of ref document: A1