CN102622589A - Multispectral face detection method based on graphics processing unit (GPU) - Google Patents
Multispectral face detection method based on graphics processing unit (GPU) Download PDFInfo
- Publication number
- CN102622589A CN102622589A CN2012100644917A CN201210064491A CN102622589A CN 102622589 A CN102622589 A CN 102622589A CN 2012100644917 A CN2012100644917 A CN 2012100644917A CN 201210064491 A CN201210064491 A CN 201210064491A CN 102622589 A CN102622589 A CN 102622589A
- Authority
- CN
- China
- Prior art keywords
- image
- face
- visible light
- detection
- infrared
- Prior art date
- Legal status (The legal status 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 status listed.)
- Pending
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 117
- 238000012545 processing Methods 0.000 title abstract description 3
- 238000000034 method Methods 0.000 claims description 46
- 230000036544 posture Effects 0.000 claims description 45
- HPTJABJPZMULFH-UHFFFAOYSA-N 12-[(Cyclohexylcarbamoyl)amino]dodecanoic acid Chemical compound OC(=O)CCCCCCCCCCCNC(=O)NC1CCCCC1 HPTJABJPZMULFH-UHFFFAOYSA-N 0.000 claims description 9
- 230000004927 fusion Effects 0.000 claims description 5
- 238000012706 support-vector machine Methods 0.000 abstract description 4
- 238000005516 engineering process Methods 0.000 abstract description 3
- 238000005286 illumination Methods 0.000 description 14
- 230000006870 function Effects 0.000 description 13
- 210000000887 face Anatomy 0.000 description 12
- 238000012549 training Methods 0.000 description 11
- 239000013598 vector Substances 0.000 description 11
- 230000008859 change Effects 0.000 description 5
- 238000010586 diagram Methods 0.000 description 5
- 238000013459 approach Methods 0.000 description 4
- 230000008901 benefit Effects 0.000 description 4
- 238000011160 research Methods 0.000 description 4
- 238000004364 calculation method Methods 0.000 description 3
- 210000001331 nose Anatomy 0.000 description 3
- 230000008569 process Effects 0.000 description 3
- 230000000007 visual effect Effects 0.000 description 3
- 238000013528 artificial neural network Methods 0.000 description 2
- 238000006243 chemical reaction Methods 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 230000018109 developmental process Effects 0.000 description 2
- 210000004709 eyebrow Anatomy 0.000 description 2
- 230000001815 facial effect Effects 0.000 description 2
- 210000004209 hair Anatomy 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- 230000004807 localization Effects 0.000 description 2
- 239000011159 matrix material Substances 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 230000011218 segmentation Effects 0.000 description 2
- 230000009466 transformation Effects 0.000 description 2
- 206010017577 Gait disturbance Diseases 0.000 description 1
- 238000007476 Maximum Likelihood Methods 0.000 description 1
- 241000405217 Viola <butterfly> Species 0.000 description 1
- 239000000654 additive Substances 0.000 description 1
- 230000000996 additive effect Effects 0.000 description 1
- 238000003340 combinatorial analysis Methods 0.000 description 1
- 238000000354 decomposition reaction Methods 0.000 description 1
- 230000003247 decreasing effect Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000009977 dual effect Effects 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 210000001061 forehead Anatomy 0.000 description 1
- 238000003064 k means clustering Methods 0.000 description 1
- 238000013178 mathematical model Methods 0.000 description 1
- 239000000203 mixture Substances 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 230000000717 retained effect Effects 0.000 description 1
- 238000000926 separation method Methods 0.000 description 1
Images
Landscapes
- Image Processing (AREA)
Abstract
The invention discloses a multispectral face detection method based on a graphics processing unit (GPU). The GPU based on compute unified device architecture (CUDA) is used for calculating an infrared light video and a visible light video which are recorded synchronously, so that features of a face in an infrared light image and a visible light image are detected respectively; and an infrared light detection result and a visible light detection result are combined synchronously, and a combined result is used as a face feature of a human and output. According to the multispectral face detection method, the face detection results based on the infrared light image and the visible light image are combined. The detection method is not influenced by light, and a detected face image is an accurate visible light image; and a face in an image can be detected under a severe environment. During detection, a support vector machine (SVM) classifier for classifying attitudes of faces is constructed; and due to the classification of the attitude classifier, face detection based on an adaboos detection algorithm is performed on sub types of infrared images. According to the technology, faces with various attitudes in the infrared images can be detected; and the limitation that only front faces in the infrared images can be detected is broken through.
Description
Technical Field
The invention relates to the technical field of human target query and search, in particular to a multispectral face detection method based on a GPU.
Background
At present, methods for unilaterally detecting human faces of visible light images and infrared images can be classified into four categories, specifically, 1) methods based on prior knowledge are introduced as follows. These a priori knowledge based methods encode the knowledge that makes up a typical face. Typically, the a priori knowledge contains the interrelationships between these facial features. The method is mainly used for face positioning. One difficulty with this approach is how to translate the face knowledge into well-defined criteria. If the criteria are too detailed, some faces will be missed because they do not pass all criteria. If the criteria are too coarse, many positive errors will likely be made. Furthermore, this method is difficult to extend to detecting faces of different poses, as it is difficult to enumerate all possible cases. On the other hand, the heuristic method has a good effect on detecting the positive human face in a specific scene. 2) A feature invariant method. The goal of these algorithms is to find some structural features that are present, which remain unchanged in the case of changes in pose, viewpoint, lighting conditions. These characteristics are then used to locate the face. These methods are mainly used for face localization. Researchers have been trying to find invariant quantities of faces, as opposed to knowledge-based methods, for detecting faces. Global features like skin color, size, shape etc. are used to find candidate faces, which are then verified with local features, such as eyebrows, nose and hair. One typical approach is to initially detect skin-like regions and then group and associate the face-like pixels using a combinatorial analysis or clustering algorithm. If the associated region has an elliptical or oval shape, then the region becomes a candidate face. 3) Template matching method. The method firstly stores several standard templates of a face to describe the whole face or part of the face. Detection is then performed by calculating the degree of correlation between the input image and the template that has been stored. These methods can be used for both face detection and face localization. Early Sakai et al created a face model using sub-templates of the eyes, nose, mouth and face contours. Miao et al propose a hierarchical template matching method for face detection. The input image is rotated according to a certain step length to form image layering, and edges are extracted by using Laplace transform. The face template contains the edges of six face structural elements: two eyebrows, two eyes, one nose, one mouth. And finally, determining whether the human face exists or not by using a heuristic method. Sinha describes the spatial characteristics of the face pattern using a set of spatial image invariants. When the variable changes the brightness of different parts of the face as the lighting changes, the relative brightness of these parts remains substantially unchanged. The pairwise brightness ratios of similar regions are determined, and the general trend of these ratios is retained, such as one region being brighter or darker than another, which provides us with a good invariance to the increase. Thus, the observed luminance law is encoded as a coarse face-space scale template containing appropriately selected sub-regions corresponding to the dominant facial features, such as eyes, cheeks, and forehead. The brightness limit between the face features is obtained through a set of proper bright-dark relation pairs between the sub-regions. 4) Appearance-based methods. Unlike template matching, the template is learned from a set of training images that should include representative variations in the appearance of a human face. These methods are mainly used for face detection. The appearance-based approach can be understood by using a probabilistic framework. An image vector or a feature vector acquired from an image is taken as a random variable x, and the value of x is determined by a conditional density function. The bayesian classification method and the maximum likelihood method can then be used to decide whether the candidate image location is a face or not. Another way to implement an appearance-based approach is to find a discriminant function between a face and a non-face. Conventionally, image patterns are mapped into a low-dimensional space and then discriminant functions are formed for classification, or a nonlinear decision surface is formed using a multi-layer neural network.
The development of the face detection technology in speed can be divided into several stages, and the improvement of detection precision and the face detection of various visual angles are mainly used as main contents in the initial research stage, so that the attention on speed is relatively low. Some representative line researches have a k-means clustering method, which establishes a plurality of face templates in a feature space by a clustering method and learns the distance between a training sample and each template by using a neural network. The human face detection method based on the skin color characteristics considers that the skin color of the human face has consistency and can be described by a uniform model. When skin color is used for face detection, different modeling methods can be adopted, and the methods mainly comprise a Gaussian model, a Gaussian mixture model, non-parameter estimation and the like. The nonparametric kernel function probability density estimation method can also be used for establishing a skin color model, and on the basis, the mean shift method can realize the detection and tracking of the face. The algorithm improves the detection speed of the human face and has certain robustness on shielding and illumination. But this method has difficulties in dealing with complex backgrounds and multiple faces. In order to solve the illumination problem, some scholars propose a method of compensating for different illumination and then detecting a skin color area. One turning point for improving the face detection speed is that the Adaboost and Cascade algorithms proposed by p. Viola realize a real-time face detection system, so that the face detection is practical in a real sense. Learning algorithms based on AdaBoost. It can select a small fraction of key features from a large set of features, resulting in an extremely efficient classifier. Systems based on Boosting and Cascade algorithms have great advantages in speed. On the basis of a series of proposed Haar-like characteristics, weak classifiers are learned through a Boosting algorithm and then combined into a strong classifier. However, generally, a strong classifier is not enough to complete the task satisfactorily, and a series of such strong classifiers are cascaded. However, to further improve the detection accuracy, more strong classifiers need to be cascaded, which reduces the detection speed. .
Disclosure of Invention
In order to make up for the defect that the face detection under multispectral has blank and meet the requirements of image enhancement and detection speed under complex environments, the invention provides a multispectral face detection method based on a GPU.
The method adopts the technical scheme that a GPU based on a CUDA framework is adopted to operate synchronously recorded infrared light videos and visible light videos, the characteristics of human faces under an infrared light image and a visible light image are respectively detected, the infrared light detection result and the visible light detection result are synchronously fused, and the fused result is used as the human face characteristic to be output
The human face feature detection step in the visible light video comprises the following steps:
(1) extracting feature points of images in the video by adopting an LBP (local Binary pattern) operator mode;
(2) dividing the feature points extracted in the step (1) into three subclasses of positive posture, left posture and right posture by an SVM classifier according to different image postures represented by the feature points;
(3) and (3) classifying the feature points of the three subclasses in the step (2) by adopting an SVM mode again, and distinguishing the human face features from all the feature points by utilizing the difference between the human skin color chromaticity and the non-skin color chromaticity.
The human face feature detection step in the infrared light video comprises the following steps:
firstly, extracting feature points of images in a video by adopting an LBP (local Binary pattern) operator mode;
dividing the feature points extracted in the step I into three subclasses of positive posture, left posture and right posture by an SVM classifier according to the difference of image postures represented by the feature points;
and thirdly, operating the divided three subclasses in the step two by using an adaboost algorithm, and distinguishing and identifying the human face features.
The division rule of the human face gesture in the visible light or infrared light image is as follows: an image vertical direction axis is taken as an angle of 0 degree, clockwise is taken as a positive direction, a space between-30 degrees and 30 degrees is divided into a positive posture, a space between-90 degrees and-30 degrees is divided into a left posture, and a space between 30 degrees and 90 degrees is divided into a right posture.
When the detection and identification result of the infrared light content and the detection and identification result of the visible light content are fused, the detection results of the infrared image and the visible light image which are synchronously recorded at the same time are fused by referring to the recording time of the two videos, and the fusion formula is as follows:
wherein,
and
respectively human face characteristics detected by human faces in the visible light image and the infrared image,for features of human face in visible and infrared images
Corresponding to the average luminance of the image area,
is the threshold value when judging that the average brightness of the image block is low，
Judging the threshold value when the average brightness of the image block is higher; in the above formula, if the detection results in the visible light image and the infrared image are the same, the detection result of the visible light image is taken; if the human face is not detected in the visible light image and the human face is detected in the infrared image, judging the brightness of the corresponding area in the visible light image, and if the human face detection fails due to overlarge or undersize brightness, taking the detection result of the infrared image as the final detection result.
The multispectral-based face detection method disclosed by the invention fuses face detection results of visible light images and infrared images. The detection method is not influenced by illumination, and the detected face image is an accurate visible light image. The face detection of the image in the severe environment can be realized. In the detection, an SVM classifier for classifying postures is established for the face postures, classification is carried out through the posture classifier, and face detection based on the adaboost algorithm is carried out on subclasses of each posture. The technology can detect the human faces in various postures in the infrared image, and breaks through the limitation that only the human face at the front side can be detected in the previous infrared image.
Drawings
FIG. 1 is a schematic flow chart of one embodiment of the method of the present invention;
FIG. 2 is a diagram illustrating an embodiment of face feature detection for infrared video in accordance with the present invention;
fig. 3A is a schematic diagram of an LBP operator with P =4, R = 1;
fig. 3B is a schematic diagram of an LBP operator with P =8, R = 1;
fig. 3C is a schematic diagram of an LBP operator with P =16, R = 1;
fig. 4 is a flowchart illustrating an LBP texture feature calculation process.
Detailed Description
The invention is further described below with reference to the accompanying drawings.
Referring to fig. 1, as shown in the flowchart of an embodiment of the method, the method adopts a GPU based on a CUDA framework to operate on synchronously recorded infrared light videos and visible light videos, respectively detects features of a human face under an infrared light image and a visible light image, synchronously fuses an infrared light detection result and a visible light detection result, and outputs the fused result as a human face feature.
In the method, the step of detecting the face features in the visible light video comprises the following steps:
(1) extracting feature points of images in the video by adopting an LBP (local Binary pattern) operator mode;
(2) dividing the feature points extracted in the step (1) into three subclasses of positive posture, left posture and right posture by an SVM classifier according to different image postures represented by the feature points;
(3) and (3) classifying the feature points of the three subclasses in the step (2) by adopting an SVM mode again, and distinguishing the human face features from all the feature points by utilizing the difference between the human skin color chromaticity and the non-skin color chromaticity.
Referring to fig. 2, the detection step of the face feature in the infrared light video in the above method includes:
firstly, extracting feature points of images in a video by adopting an LBP (local Binary pattern) operator mode;
dividing the feature points extracted in the step I into three subclasses of positive posture, left posture and right posture by an SVM classifier according to the difference of image postures represented by the feature points;
and thirdly, operating the divided three subclasses in the step two by using an adaboost algorithm, and distinguishing and identifying the human face features.
The division rule of the human face gesture in the visible light or infrared light image features is as follows: an image vertical direction axis is taken as an angle of 0 degree, clockwise is taken as a positive direction, a space between-30 degrees and 30 degrees is divided into a positive posture, a space between-90 degrees and-30 degrees is divided into a left posture, and a space between 30 degrees and 90 degrees is divided into a right posture.
In the method, when the detection and identification result of the infrared light content and the detection and identification result of the visible light content are fused, the detection results of the infrared image and the visible light image which are synchronously recorded at the same time are fused by referring to the recording time of the two videos, and the fusion formula is as follows:
wherein,
and
respectively human face characteristics detected by human faces in the visible light image and the infrared image,
for features of human face in visible and infrared images
Corresponding to the average luminance of the image area,is the threshold value when the average luminance of the image block is judged to be low,
is to judge the average brightness of the image blockA threshold at a higher time; in the above formula, if the detection results in the visible light image and the infrared image are the same, the detection result of the visible light image is taken; if the human face is not detected in the visible light image and the human face is detected in the infrared image, judging the brightness of the corresponding area in the visible light image, and if the human face detection fails due to overlarge or undersize brightness, taking the detection result of the infrared image as the final detection result.
The method is a face detection method based on the combination of the visible light image and the infrared image, which is realized on the CUDA platform, integrates the advantages that the infrared image is not influenced by the illumination environment and the accuracy of the visible light image is high, and can ensure that the face can be detected quickly and accurately.
The whole process of the invention comprises three parts of infrared image face detection, visible light image face detection and detection result fusion.
Firstly, extracting a frame of image in an infrared video and a visible light video respectively, and extracting feature points of the images respectively by adopting an LBP operator.
LBP utilizes the joint distribution of each pixel in the image and P pixel points on the annular neighborhood with the radius of R
To describe the texture features of the image. Wherein
The gray value representing the center point of the local neighborhood,
the LBP operators are different for different (P, R) combinations corresponding to the gray values of P bisectors on a circle with radius R, see fig. 3A, 3B, 3C, 3 different LBP operators.
In order to realize the invariance of the texture features to the gray level, P equal divisions on the annular neighborhood are usedGray value of pointGray value minus center point
Combined distributed T conversion to
In the formula,
the gray distribution of the whole image is described, the local texture feature distribution of the image is not influenced, and therefore, the texture feature of the image can be described by the differential joint distribution, namely
When the illumination of the image is additively changed, the relative size of the gray value of the central pixel and the gray value of the pixels in the annular neighborhood of the central pixel is generally not changed, namely
Independent of additive variations in illumination, the texture of the image can be described by using a sign function of the difference between the central pixel and the neighboring pixels instead of a specific value, i.e. by using a sign function of the difference between the central pixel and the neighboring pixels
In the above formula: s is a symbolic function
The results obtained from the joint distribution T are ordered according to the specific order of the pixels on the ring-shaped neighborhood to form an 0/1 sequence, in this embodiment, the calculation is started in the counterclockwise direction by using the right neighborhood pixel of the central pixel as the starting pixel, and each item is given
Imparting a binomial factor
The Local spatial texture structure of a pixel can be represented as a unique decimal number called LBP, R-number, which is why the texture operator is called Local Binary Pattern (Local Binary Pattern), which can be calculated by the following formula
A specific LBP texture feature calculation process is described with reference to fig. 4, (in the figure, P =8, R = 1).
And (3) thresholding the template on the left side of the image 4, comparing each neighborhood pixel point with a central pixel (131), setting the pixel values to be greater than 0 to 1 and less than 0 to obtain an 0/1 table of the middle position, constructing a 0/1 sequence (10100101) by taking the lower right corner as the start in a counterclockwise sequence, finally calculating a corresponding decimal number (165), wherein the LBP texture characteristic value of the pixel point is 165, and solving the LBP characteristic value of each pixel in the image to obtain the LBP texture characteristic map of the image. Because the LBP texture features at the edge of the image are less affected by the neighborhood, the original pixel gray values are reserved for the pixel points at the edge of the image.
And after the LBP operator is adopted to extract all the image characteristics, an SVM classifier is adopted to classify the characteristic points in the infrared image and the visible light image respectively.
In the invention, the SVM classifier solves the problem of seeking an optimal interface in an original space. The mathematical model of the problem is as follows:
whereinIn order to be at intervals of time,
is the number of training samples and is,
is a vector of training samples that is,
is a vector of the weights that is,
is the threshold value of the threshold value,
in order to mark the sample, the sample is marked,
，
represents the first
And (4) class.
Constructing a Lagrangian function to obtain
Are respectively paired
Differentiating and substituting into Lagrange function to obtain
=
To find
Is optimized to obtain
Obtaining a dual Lagrange function
The original problem is converted into the following optimization problem
According to the theory of optimization,
for the KKT additional condition, only a few samples, i.e., support vectors, which are the most informative data in the data set, have non-zero lagrange multipliers. At this point, the SVM classifier classifies the feature points in the image.
After the first classification by the SVM, the infrared image is detected by adopting an adaboost algorithm.
The infrared image can avoid the influence of illumination on the detection algorithm, and the invention provides a multi-view face detection algorithm based on a continuous adaboost algorithm on the infrared image. A flow diagram of one embodiment of infrared image detection. Firstly, carrying out visual angle estimation on an image, and dividing the posture of the face into 3 subclasses by adopting a statistical learning method for the estimation of the posture of the face. The y-axis in the image is a 0-degree reference, and the image is divided into three subclasses of positive attitude, left attitude and right attitude, and the front angle is [ -30 ]。,30。]Right angle [30 ]。,90。]And left angle [ -90 ]。,-30。]. Extracting local binary pattern (local binary pattern) from the face samples of the three subclassesLBP) feature training support vector machine SVM. And carrying out posture classification on the input human face image through the SVM. The face is divided into a plurality of viewpoint subclasses according to the three-dimensional posture, a look-up phenotype weak classifier form with continuous confidence degree output is designed by using LBP characteristics for each subclass, a weak classifier space is constructed, and a waterfall type face detector based on the view is learned by adopting a continuous Adaboost algorithm. During face detection, after classification is completed, the adaboost face detectors of all the subclasses are called for detection.
The core of the AdaBoost algorithm is that a plurality of key weak classifiers are automatically screened from a weak classifier space by adjusting sample distribution and weak classifier weights, and are integrated into a strong classifier in a certain mode. The Adaboost learning algorithm is as follows:
sample set
，
。
Initialization: for each one
，。
Obtaining basic classification rules
x-y
Updating:
here, the
Is a normalized constant such that
。
And (5) after iteration is finished, finally forming a cascade classifier as follows:
in the above-mentioned algorithm,
for the weights of the samples in the iterative method,the device is a weak-type device,
is an integrated strong classifier.
And calculating the image by an AdaBoost algorithm to detect the characteristics of the human face.
And detecting the visible light image based on a skin color model and an SVM classifier.
A large number of experiments prove that the change range of human skin color and chroma is obviously different from non-skin color (such as hair and surrounding objects). There are many common color spaces that can express skin color, such asEtc. we choose to use
The color space detects skin tones.
The space has the characteristic of separating the chrominance from the luminance, i.e. the space is to
(brightness) of the light emitted from the light source,
(blue chromaticity) and
(red chroma) separation. And in
The clustering characteristic of the skin color in the color space is better, the influence of brightness change is smaller, and the skin color distribution area can be better limited.
And
the coordinate correspondence relationship of mutual conversion is as follows:
selecting a large number of skin color samples for statistics, wherein the statistical distribution is selected to meet the requirement
A skin tone segmentation threshold. On a two-dimensional chromaticity plane, the area of skin color is more concentrated, and skin color pixels are subject to mean
Variance (variance)A gaussian distribution of (a). A gaussian skin tone model can thus be built in YCbCr space. According to the Gaussian distribution of skin color in the chromaticity space, for each pixel in the color image, after the pixel is converted from the RGB color space to the YCbCr space, the probability that the point belongs to the skin area can be calculated. The formula is as follows:
wherein
And calculating the skin color likelihood of each pixel point by using the formula, multiplying each point by 255 to obtain the gray value of the point, and setting a proper threshold value to segment a skin area in the gray image to obtain a binary image. The divided binary image needs to be processed by a mathematical morphology method, and the application of the mathematical morphology can simplify the image data, keep the basic shape characteristics and remove irrelevant structures. And carrying out face detection based on SVM classification on the basis of the image obtained by the skin color model segmentation.
The specific method comprises the following steps: the training image is normalized to a standard 64 x 64 image, and the sample image is expanded into a one-dimensional vector to obtain a sample mean and covariance matrix:
decomposing the covariance matrix by singular value decomposition, and arranging the calculated eigenvalues in a monotonically decreasing order
Feature vector corresponding thereto
,
Is stretched into a subspace, is called
Is a characteristic face space. Any given face image x, may be projected into the subspace,
. This set of coefficients, which indicates the position of the image in the subspace, will be used as a new feature vector for training to the SVM. Projection coefficient of image
The quantity is called the main feature of the face pattern Xi, and the space formed by the main features is called the feature space. And inputting the obtained coefficients serving as feature vectors into the SVM for training. And selecting the radial basis function as a kernel function when training the support vector machine.
，
. For any training component
Face modelThe formula is defined as 1, and the non-face mode is defined as-1. The training function is
In the formula
Support vector is used;
is the corresponding weight;
is composed of
The corresponding class label;
is the number of support vectors.
And after the infrared image and the visible light image are detected for the image characteristics, fusing the detection results. When the detection and identification result of the infrared light content and the detection and identification result of the visible light content are fused, the detection results of the infrared image and the visible light image which are synchronously recorded at the same time are fused by referring to the recording time of the two videos, and the fusion formula is as follows:
wherein,
andrespectively human face characteristics detected by human faces in the visible light image and the infrared image,for features of human face in visible and infrared images
Corresponding to the average luminance of the image area,
is the threshold value when the average luminance of the image block is judged to be low,
judging the threshold value when the average brightness of the image block is higher; in the above formula, if the detection results in the visible light image and the infrared image are the same, the detection result of the visible light image is taken; if the human face is not detected in the visible light image and the human face is detected in the infrared image, judging the brightness of the corresponding area in the visible light image, and if the human face detection fails due to overlarge or undersize brightness, taking the detection result of the infrared image as the final detection result.
The hardware platform for detection and operation is a GPU with a CUDA framework, has wide application range, can be applied to the fields of face recognition, new generation human-computer interfaces, safe access, visual monitoring, content-based retrieval and the like, and is generally regarded by researchers in recent years.
The face detection needs to be put to practical application, and the precision and the speed are two key problems which need to be solved urgently. Through the development of more than ten years after the 90 s of the 20 th century, the precision of face detection is greatly improved, but the speed is always the stumbling block which hinders the face detection to be practical. Therefore, researchers have made hard efforts. NVIDIA later introduced a parallel computing architecture that enabled GPUs to solve complex computational problems. It contains the CUDA Instruction Set Architecture (ISA) and the parallel computing engine inside the GPU. The system realizes a face detection algorithm on the framework of the CUDA. On a parallel processing architecture, the image is divided into grids in parallel, and the divided image data is sent to the GPU in parallel. The computer simultaneously judges the divided image data.
The multispectral face detection method based on the GPU fuses the face detection results of the visible light image and the infrared image. The method combines the characteristics of high accuracy and clear image of the visible light image and the advantage that the infrared image is not influenced by illumination, and accelerates the detection algorithm on the GPU based on the CUDA. In terms of algorithm, the problem of human face posture change is solved to a certain extent through a multi-angle human face detection algorithm. The problem of illumination change is solved by a multispectral face detection method.
The invention provides and realizes the multi-pose face detection method in the infrared image for the first time. Illumination transformation is a difficult problem to solve in face detection research. Infrared images are of interest because they are not affected by concerns. The front face image is less affected by illumination change, and the multi-pose face, especially the side face, is easily affected by illumination. The research of the literature finds that the face detection in the infrared image does not provide a multi-pose method. The method is based on the face detection method with multiple postures in the infrared image, the postures of the face are classified through the SVM, and the face detection is carried out by the adaboost algorithm for each class.
The invention combines the human face detection method of visible light image and infrared image, and fuses the detection result in the result layer. The method not only avoids the influence of illumination transformation on the human face detection algorithm, but also keeps the advantage of high accuracy of visible light images, and achieves the purpose of improving the human face detection accuracy.
The multispectral face detection method is implemented on the GPU based on the CUDA framework based on the GPU serving as an operation platform, and the purpose of high-speed detection of face images is achieved.
In summary, the above description is only a preferred embodiment of the present invention, and is not intended to limit the scope of the present invention. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims (5)
1. A multispectral face detection method based on a GPU is characterized in that the method adopts the GPU based on a CUDA framework to operate synchronously recorded infrared light videos and visible light videos, the characteristics of a face under an infrared light image and a visible light image are respectively detected, the infrared light detection result and the visible light detection result are synchronously fused, and the fused result is used as the face characteristic of a person to be output.
2. The GPU-based multispectral face detection method according to claim 1, wherein the GPU detects the face features in the visible light video by:
(1) extracting feature points of images in the video by adopting an LBP (local Binary pattern) operator mode;
(2) dividing the feature points extracted in the step (1) into three subclasses of positive posture, left posture and right posture by an SVM classifier according to different image postures represented by the feature points;
(3) and (3) classifying the feature points of the three subclasses in the step (2) by adopting an SVM mode again, and distinguishing the human face features from all the feature points by utilizing the difference between the human skin color chromaticity and the non-skin color chromaticity.
3. The GPU-based multispectral face detection method according to claim 1, wherein the GPU detects the face features in the ir video by:
firstly, extracting feature points of images in a video by adopting an LBP (local Binary pattern) operator mode;
dividing the feature points extracted in the step I into three subclasses of positive posture, left posture and right posture by an SVM classifier according to the difference of image postures represented by the feature points;
and thirdly, operating the divided three subclasses in the step two by using an adaboost algorithm, and distinguishing and identifying the human face features.
4. A GPU-based multispectral face detection method according to claim 2 or 3, wherein the classification rule for the face pose in the visible or infrared light image is: an image vertical direction axis is taken as an angle of 0 degree, clockwise is taken as a positive direction, a space between-30 degrees and 30 degrees is divided into a positive posture, a space between-90 degrees and-30 degrees is divided into a left posture, and a space between 30 degrees and 90 degrees is divided into a right posture.
5. The GPU-based multispectral face detection method according to claim 1, wherein when the detection and identification result of the infrared light content and the detection and identification result of the visible light content are fused, the detection and identification result of the infrared image and the detection result of the visible light image that are synchronously recorded at the same time are fused with reference to the recording time of the two videos, and the fusion formula is as follows:
wherein,
andrespectively human face characteristics detected by human faces in the visible light image and the infrared image,for features of human face in visible and infrared images
Corresponding to the average luminance of the image area,is the threshold value when the average luminance of the image block is judged to be low,
judging the threshold value when the average brightness of the image block is higher; in the above formula, if the detection results in the visible light image and the infrared image are the same, the detection result of the visible light image is taken; if the human face is not detected in the visible light image and the human face is detected in the infrared image, judging the brightness of the corresponding area in the visible light image, and if the human face detection fails due to overlarge or undersize brightness, taking the detection result of the infrared image as the final detection result.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2012100644917A CN102622589A (en) | 2012-03-13 | 2012-03-13 | Multispectral face detection method based on graphics processing unit (GPU) |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2012100644917A CN102622589A (en) | 2012-03-13 | 2012-03-13 | Multispectral face detection method based on graphics processing unit (GPU) |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN102622589A true CN102622589A (en) | 2012-08-01 |
Family
ID=46562499
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN2012100644917A Pending CN102622589A (en) | 2012-03-13 | 2012-03-13 | Multispectral face detection method based on graphics processing unit (GPU) |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN102622589A (en) |
Cited By (27)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103208103A (en) * | 2013-04-15 | 2013-07-17 | 中国科学院苏州纳米技术与纳米仿生研究所 | Graphic processing unit (GPU)-based low-luminance image enhancement method |
| CN103528990A (en) * | 2013-10-31 | 2014-01-22 | 天津工业大学 | Method for establishing multiple models of near infrared spectrums |
| CN103646235A (en) * | 2013-11-25 | 2014-03-19 | 华为技术有限公司 | Method, device and server for target detection in multi-channel video detection system |
| CN103685268A (en) * | 2013-12-10 | 2014-03-26 | 华东理工大学 | Network intrusion detection method based on GPU and SVM |
| CN103745461A (en) * | 2013-03-17 | 2014-04-23 | 北京理工大学 | Printing image defect detection method based on area combination feature |
| CN103914679A (en) * | 2013-01-05 | 2014-07-09 | 联想(北京)有限公司 | Image identification method, device and electronic device |
| CN104573640A (en) * | 2013-10-23 | 2015-04-29 | 想象技术有限公司 | Face detection |
| WO2015070764A1 (en) * | 2013-11-13 | 2015-05-21 | 智慧城市系统服务（中国）有限公司 | Face positioning method and device |
| CN105138956A (en) * | 2015-07-22 | 2015-12-09 | 小米科技有限责任公司 | Face detection method and device |
| CN105224911A (en) * | 2015-08-27 | 2016-01-06 | 湖北文理学院 | A kind of various visual angles pedestrian detection method and system in real time |
| CN105426848A (en) * | 2014-11-03 | 2016-03-23 | 倪蔚民 | Imaging method for improving success rate of biological recognition |
| CN106503687A (en) * | 2016-11-09 | 2017-03-15 | 合肥工业大学 | The monitor video system for identifying figures of fusion face multi-angle feature and its method |
| CN106707293A (en) * | 2016-12-01 | 2017-05-24 | 百度在线网络技术（北京）有限公司 | Obstacle recognition method and device for vehicles |
| CN106922192A (en) * | 2014-12-10 | 2017-07-04 | 英特尔公司 | Using the type of face detection method and device of look-up table |
| CN107153807A (en) * | 2016-03-03 | 2017-09-12 | 重庆信科设计有限公司 | A kind of non-greedy face identification method of two-dimensional principal component analysis |
| CN107194351A (en) * | 2017-05-22 | 2017-09-22 | 天津科技大学 | Face recognition features' extraction algorithm based on weber Local Symmetric graph structure |
| CN108062521A (en) * | 2017-12-12 | 2018-05-22 | 深圳大学 | Method for detecting human face, device, terminal and medium based on convolutional neural networks |
| CN108090477A (en) * | 2018-01-23 | 2018-05-29 | 北京易智能科技有限公司 | A kind of face identification method and device based on Multi-spectral image fusion |
| CN108268839A (en) * | 2018-01-05 | 2018-07-10 | 北京万相融通科技股份有限公司 | A kind of live body verification method and its system |
| CN108416326A (en) * | 2018-03-27 | 2018-08-17 | 百度在线网络技术（北京）有限公司 | Face identification method and device |
| CN108415037A (en) * | 2018-05-14 | 2018-08-17 | 珠海市微半导体有限公司 | A kind of infrared and visual signature obstacle avoidance apparatus and control method |
| CN109753848A (en) * | 2017-11-03 | 2019-05-14 | 杭州海康威视数字技术股份有限公司 | Execute the methods, devices and systems of face identifying processing |
| CN110110606A (en) * | 2019-04-12 | 2019-08-09 | 天津大学 | The fusion method of visible light neural network based and infrared face image |
| CN110175562A (en) * | 2019-05-24 | 2019-08-27 | 深圳职业技术学院 | A kind of gpu acceleration pedestrian detection algorithm based on parallel computation |
| CN112258592A (en) * | 2020-09-17 | 2021-01-22 | 深圳市捷顺科技实业股份有限公司 | Method and related device for generating human face visible light pattern |
| CN113551772A (en) * | 2020-04-07 | 2021-10-26 | 武汉高德智感科技有限公司 | Infrared temperature measurement method, infrared temperature measurement system and storage medium |
| CN114742586A (en) * | 2022-04-11 | 2022-07-12 | 中科强基科技(北京)有限公司 | Advertisement charging statistical method based on intelligent display terminal |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20090003661A1 (en) * | 2007-02-28 | 2009-01-01 | Fotonation Vision Limited | Separating a Directional Lighting Variability In Statistical Face Modelling Based On Texture Space Decomposition |
| CN101404060A (en) * | 2008-11-10 | 2009-04-08 | 北京航空航天大学 | Human face recognition method based on visible light and near-infrared Gabor information amalgamation |
| CN101853397A (en) * | 2010-04-21 | 2010-10-06 | 中国科学院半导体研究所 | Bionic human face detection method based on human visual characteristics |
-
2012
- 2012-03-13 CN CN2012100644917A patent/CN102622589A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20090003661A1 (en) * | 2007-02-28 | 2009-01-01 | Fotonation Vision Limited | Separating a Directional Lighting Variability In Statistical Face Modelling Based On Texture Space Decomposition |
| CN101404060A (en) * | 2008-11-10 | 2009-04-08 | 北京航空航天大学 | Human face recognition method based on visible light and near-infrared Gabor information amalgamation |
| CN101853397A (en) * | 2010-04-21 | 2010-10-06 | 中国科学院半导体研究所 | Bionic human face detection method based on human visual characteristics |
Cited By (44)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103914679B (en) * | 2013-01-05 | 2017-07-21 | 联想(北京)有限公司 | A kind of image-recognizing method, device and electronic equipment |
| CN103914679A (en) * | 2013-01-05 | 2014-07-09 | 联想(北京)有限公司 | Image identification method, device and electronic device |
| CN103745461B (en) * | 2013-03-17 | 2016-09-07 | 北京理工大学 | A kind of printing image defect detection method based on areas combine feature |
| CN103745461A (en) * | 2013-03-17 | 2014-04-23 | 北京理工大学 | Printing image defect detection method based on area combination feature |
| CN103208103B (en) * | 2013-04-15 | 2016-06-29 | 中国科学院苏州纳米技术与纳米仿生研究所 | A kind of enhancement method of low-illumination image based on GPU |
| CN103208103A (en) * | 2013-04-15 | 2013-07-17 | 中国科学院苏州纳米技术与纳米仿生研究所 | Graphic processing unit (GPU)-based low-luminance image enhancement method |
| CN104573640A (en) * | 2013-10-23 | 2015-04-29 | 想象技术有限公司 | Face detection |
| CN104573640B (en) * | 2013-10-23 | 2019-04-30 | 想象技术有限公司 | Data processing system and its generating device, type of face detection method |
| CN103528990B (en) * | 2013-10-31 | 2017-07-28 | 天津工业大学 | A kind of multi-model Modeling Method of near infrared spectrum |
| CN103528990A (en) * | 2013-10-31 | 2014-01-22 | 天津工业大学 | Method for establishing multiple models of near infrared spectrums |
| WO2015070764A1 (en) * | 2013-11-13 | 2015-05-21 | 智慧城市系统服务（中国）有限公司 | Face positioning method and device |
| CN103646235A (en) * | 2013-11-25 | 2014-03-19 | 华为技术有限公司 | Method, device and server for target detection in multi-channel video detection system |
| CN103646235B (en) * | 2013-11-25 | 2016-10-05 | 华为技术有限公司 | Object detection method, device and server in multi-channel video detecting system |
| CN103685268A (en) * | 2013-12-10 | 2014-03-26 | 华东理工大学 | Network intrusion detection method based on GPU and SVM |
| CN105426848A (en) * | 2014-11-03 | 2016-03-23 | 倪蔚民 | Imaging method for improving success rate of biological recognition |
| CN105426848B (en) * | 2014-11-03 | 2020-12-18 | 苏州思源科安信息技术有限公司 | Imaging method for improving success rate of biological recognition |
| CN106922192A (en) * | 2014-12-10 | 2017-07-04 | 英特尔公司 | Using the type of face detection method and device of look-up table |
| CN106922192B (en) * | 2014-12-10 | 2021-08-24 | 英特尔公司 | Face detection method and apparatus using lookup table |
| CN105138956B (en) * | 2015-07-22 | 2019-10-15 | 小米科技有限责任公司 | Method for detecting human face and device |
| CN105138956A (en) * | 2015-07-22 | 2015-12-09 | 小米科技有限责任公司 | Face detection method and device |
| CN105224911A (en) * | 2015-08-27 | 2016-01-06 | 湖北文理学院 | A kind of various visual angles pedestrian detection method and system in real time |
| CN107153807A (en) * | 2016-03-03 | 2017-09-12 | 重庆信科设计有限公司 | A kind of non-greedy face identification method of two-dimensional principal component analysis |
| CN106503687B (en) * | 2016-11-09 | 2019-04-05 | 合肥工业大学 | Merge the monitor video system for identifying figures and its method of face multi-angle feature |
| CN106503687A (en) * | 2016-11-09 | 2017-03-15 | 合肥工业大学 | The monitor video system for identifying figures of fusion face multi-angle feature and its method |
| CN106707293B (en) * | 2016-12-01 | 2019-10-29 | 百度在线网络技术（北京）有限公司 | Obstacle recognition method and device for vehicle |
| CN106707293A (en) * | 2016-12-01 | 2017-05-24 | 百度在线网络技术（北京）有限公司 | Obstacle recognition method and device for vehicles |
| CN107194351B (en) * | 2017-05-22 | 2020-06-23 | 天津科技大学 | Face recognition feature extraction method based on Weber local symmetric graph structure |
| CN107194351A (en) * | 2017-05-22 | 2017-09-22 | 天津科技大学 | Face recognition features' extraction algorithm based on weber Local Symmetric graph structure |
| CN109753848A (en) * | 2017-11-03 | 2019-05-14 | 杭州海康威视数字技术股份有限公司 | Execute the methods, devices and systems of face identifying processing |
| CN109753848B (en) * | 2017-11-03 | 2021-01-26 | 杭州海康威视数字技术股份有限公司 | Method, device and system for executing face recognition processing |
| CN108062521A (en) * | 2017-12-12 | 2018-05-22 | 深圳大学 | Method for detecting human face, device, terminal and medium based on convolutional neural networks |
| CN108268839A (en) * | 2018-01-05 | 2018-07-10 | 北京万相融通科技股份有限公司 | A kind of live body verification method and its system |
| CN108090477A (en) * | 2018-01-23 | 2018-05-29 | 北京易智能科技有限公司 | A kind of face identification method and device based on Multi-spectral image fusion |
| CN108416326A (en) * | 2018-03-27 | 2018-08-17 | 百度在线网络技术（北京）有限公司 | Face identification method and device |
| CN108416326B (en) * | 2018-03-27 | 2021-07-16 | 百度在线网络技术（北京）有限公司 | Face recognition method and device |
| CN108415037B (en) * | 2018-05-14 | 2023-06-13 | 珠海一微半导体股份有限公司 | Obstacle avoidance device with infrared and visual characteristics and control method |
| CN108415037A (en) * | 2018-05-14 | 2018-08-17 | 珠海市微半导体有限公司 | A kind of infrared and visual signature obstacle avoidance apparatus and control method |
| CN110110606A (en) * | 2019-04-12 | 2019-08-09 | 天津大学 | The fusion method of visible light neural network based and infrared face image |
| CN110175562B (en) * | 2019-05-24 | 2024-01-16 | 深圳职业技术大学 | Parallel computing-based gpu acceleration pedestrian detection algorithm |
| CN110175562A (en) * | 2019-05-24 | 2019-08-27 | 深圳职业技术学院 | A kind of gpu acceleration pedestrian detection algorithm based on parallel computation |
| CN113551772A (en) * | 2020-04-07 | 2021-10-26 | 武汉高德智感科技有限公司 | Infrared temperature measurement method, infrared temperature measurement system and storage medium |
| CN113551772B (en) * | 2020-04-07 | 2023-09-15 | 武汉高德智感科技有限公司 | Infrared temperature measurement method, infrared temperature measurement system and storage medium |
| CN112258592A (en) * | 2020-09-17 | 2021-01-22 | 深圳市捷顺科技实业股份有限公司 | Method and related device for generating human face visible light pattern |
| CN114742586A (en) * | 2022-04-11 | 2022-07-12 | 中科强基科技(北京)有限公司 | Advertisement charging statistical method based on intelligent display terminal |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN102622589A (en) | Multispectral face detection method based on graphics processing unit (GPU) | |
| US10929649B2 (en) | Multi-pose face feature point detection method based on cascade regression | |
| CN102799901B (en) | Method for multi-angle face detection | |
| Yang et al. | Face detection and gesture recognition for human-computer interaction | |
| CN104268583B (en) | Pedestrian re-recognition method and system based on color area features | |
| Tsao et al. | A data mining approach to face detection | |
| US20100027892A1 (en) | System and method for circling detection based on object trajectory | |
| CN110956099B (en) | Dynamic gesture instruction identification method | |
| CN111158457A (en) | Vehicle-mounted HUD (head Up display) human-computer interaction system based on gesture recognition | |
| CN111158491A (en) | Gesture recognition man-machine interaction method applied to vehicle-mounted HUD | |
| Maksymiv et al. | Video-based flame detection using LBP-based descriptor: influences of classifiers variety on detection efficiency | |
| Hassaballah et al. | Facial features detection and localization | |
| Kheirkhah et al. | A hybrid face detection approach in color images with complex background | |
| Zhang et al. | Hand gesture recognition based on HOG-LBP feature | |
| Proenca et al. | Soft biometrics: Globally coherent solutions for hair segmentation and style recognition based on hierarchical mrfs | |
| Naseer et al. | Multimodal Objects Categorization by Fusing GMM and Multi-layer Perceptron | |
| Gilani et al. | Towards large-scale 3D face recognition | |
| Sarma et al. | Hand detection by two-level segmentation with double-tracking and gesture recognition using deep-features | |
| Huang et al. | High-efficiency face detection and tracking method for numerous pedestrians through face candidate generation | |
| Bhuiyan | Towards face recognition using eigenface | |
| Wu et al. | Partially occluded head posture estimation for 2D images using pyramid HoG features | |
| Xu et al. | Patch-wise skin segmentation of human body parts via deep neural networks | |
| Kounavis | Fingertip detection without the use of depth data, color information, or large training data sets | |
| Gupta et al. | Hand skin classification from other skin objects using multi-direction 3D color-texture feature and cascaded neural network classifier | |
| Gupta et al. | Design and Analysis of an Expert System for the Detection and Recognition of Criminal Faces |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| ASS | Succession or assignment of patent right |
Owner name: CHEN YUCHUN Free format text: FORMER OWNER: BEIJING WISLAND TECHNOLOGY PTE LTD. Effective date: 20130903 |
|
| C41 | Transfer of patent application or patent right or utility model | ||
| COR | Change of bibliographic data |
Free format text: CORRECT: ADDRESS; FROM: 100080 HAIDIAN, BEIJING TO: 100101 CHAOYANG, BEIJING |
|
| TA01 | Transfer of patent application right |
Effective date of registration: 20130903 Address after: 100101 Beijing city Chaoyang District Anxiang 47 floor 5 door 202 Applicant after: Chen Yuchun Address before: 100080 Beijing City, Haidian District Zhongguancun West strongchina technology building 607A Applicant before: Wisland Technology (Beijing) Pte Ltd. |
|
| C12 | Rejection of a patent application after its publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20120801 |