CN103268499B - Human body skin detection method based on multispectral imaging - Google Patents

Abstract

The method utilizing calibration in multi-optical spectrum imaging system obtains object reflection characteristic under selected two wave bands and as feature, reflection characteristic is carried out skin detection.Described multi-optical spectrum imaging system includes the object of reference that reflection characteristic is known, and configuration shoots light source and the equipment of reception of selected two different-waveband hypographs.It is utilize the reflection characteristic of other objects in the reflection characteristic predicted picture of object of reference that the described method by calibrating obtains object reflection characteristic.Described skin detecting method is as feature using the different reflection characteristics of human body skin and non-skin object, and skin and noncutaneous detection are considered as two classification problems, utilizes the method training grader of machine learning, and detects.<!--1-->

Description

Human body skin detection method based on multispectral imaging

Two, technical field:

The present invention relates to computer vision and image processing field, particularly relate to the method utilizing computer vision to carry out human body skin detection.More particularly it relates to an cheap and real-time multi-optical spectrum imaging system is obtained image and is obtained out the reflection characteristic of objects in images by system, method design.Present invention relate to how detection human body skin under the complex background having dummy, mask simultaneously.

Three, background technology:

Skin detection is one important research direction of computer vision field.Such as, the detecting and tracking of face in video conference, the understanding of body language, the instruction analysis etc. in man-machine interaction, it is required for accurate skin detection algorithm.Traditional skin detection algorithm is mostly based on the colouring information of RGB camera acquisition and carries out human body skin tone testing, the principle of RGB camera is red by R(), G(green), B(blue) three kinds of Color Channel component combined color images, therefore modal skin detection algorithm is to reach testing goal by defining colour of skin interval on each component in color space.So, when the color-values of pixel drops in interval of definition, then it is assumed that this pixel is skin.About the select permeability of color space, Kakumanua et al. gives good summary.But, when detection crowd includes from not agnate people, or background color is complicated, and ambient light is according to time stronger, and the skin detection precision based on color is just relatively low.Although it addition, use the color space with illumination invariant, it is possible to reduce ambient light to a certain extent according to the impact on algorithm.But when extraneous illumination variation is violent, skin detection performance still reduces significantly.And, the skin detection algorithm based on the trichroism imaging of RGB can not effectively detect and remove the article close with human body complexion color, particularly cannot distinguish between true man and dummy's (including body model, mask, photo etc.).Therefore, the skin detection algorithm based on the trichroism imaging of RGB has significant limitation.

Human body skin reflectance is from physiology and the characteristic being experimentally proved its uniqueness: relatively low at ultraviolet portion reflectance, then increases with the increase reflectance of wavelength, has " W " shape at 575nm place.Skin reflectivity reaches peak value at 800nm place, there will be decline at 900nm place, after reaching minima again increase but increase amplitude is only small.Through reality test, such as figure mono-, dummy's reflectance is not possess this characteristic.Utilize this reflectivity Characteristics can distinguish corium skin and dummy and other article.But, spectrogrph is used in the acquisition of current reflectance mostly, and therefore length expensive, time-consuming can not be measured in real time.

[reference listing]

[non-patent literature]

1, ElliAngelopoulou, " Understandingthecolorofhumanskin ", ComputerScienceDept., StevensInstituteofTechnology, CastlePointonHudson, Hoboken, NJ07030, USA

The US6370260 of IoannisPavlidis illustrates the method that two kinds of near infrared band images of a kind of 800nm-1400nm and 1400nm-2200nm of use carry out human body skin detection, utilizes human body skin in the unique reflections characteristic of both wave bands such that it is able to distinguish true man and dummy.But, the imaging device of the two wave band is much more expensive, is difficult in usual civil applications and uses.

The 201080002426 of Sony propose skin detection system and method under a kind of use multiband near-infrared lighting condition, it is possible to the multiple care pixels in detection image, identify and predetermined corresponding object.The method uses two different near-infrared wavelengths of 870nm and 950nm to obtain two width images under corresponding wavelength, then generates binaryzation skin image according to the difference between two width images, utilizes binary image to extract skin image from original image.Extraction to the skin (hands) being sitting in the data inputting personnel before computer screen when the method is mainly used in computer data typing, it is impossible to be applied to the skin detection under usual scene.

Therefore case above is considered, it is necessary to find a kind of multi-optical spectrum imaging system cheap, that real-time is good and obtain multispectral image.Simultaneously need to a kind of skin detecting method that can effectively distinguish true man and dummy under usual scene.

Goal of the invention

The main purpose of the present invention is to provide a kind of multi-optical spectrum imaging system, and this system price is cheap, simple in construction, can obtain object reflection characteristic in the picture by this system.There is provided a kind of detection method to complex background (especially containing dummy) to carry out skin detection for the reflection characteristic present invention obtained.

A kind of method that it is a further object to provide calibration, obtains object reflection characteristic under each wavelength by this calibration steps.The effect of calibration is to eliminate imaging path (including light source, capture apparatus and optical filter) spectral response difference at different wavelengths, a marker that reflectance is known on different wave length is used to calibrate the spectral response of imaging path on different wave length, thus obtaining the true reflection characteristic of object.Calibration steps is relevant with the reflection characteristic of required acquisition.Reflection characteristic can refer to the gray value of pixel, reflectance and deformation thereof herein, for instance the deformation of gray value can make the difference between normalized gray value, gray value.

Present invention also offers a kind of skin detecting method, skin detection is considered as two classification problems by the method, using reflection characteristic as feature, utilizes the method training grader of machine learning to carry out skin detection.

Four, summary of the invention:

Before the system of the present invention, method and hardware realize description, it is necessary to explanation is to the invention is not limited in described several embodiments and detection method, and the present invention can have multiple possible embodiment not clearly stated in this article.

Detection method provided by the invention is to utilize narrow-band multispectral to irradiate lower skin and noncutaneous different reflection characteristics, carries out skin detection, such that it is able to effectively distinguish human body skin and anthropoid skin article (including body model, mask, photo etc.).

The invention provides the system of a kind of multispectral imaging, this system cost is low, real-time good.The image under interested wave band can be obtained by system provided by the present invention.Light source selected by system must cover selected multi light spectrum hands (can also adopt nature light during conditions permit), if such as selected wave band is 420nm and 850nm, then selected light source is necessary for comprising the two wave band.The capture apparatus that native system adopts is B/W camera, and capture apparatus can also select according to the difference of embodiment, and photodiode such as can be used as receptor.One embodiment of the present of invention, adopts a capture apparatus, obtains multispectral image by switching the narrow band pass filter of different-waveband before capture apparatus successively.Additionally can also adopting two capture apparatus, the narrow band pass filter installing different-waveband before each capture apparatus additional obtains multispectral image simultaneously.Install narrow band pass filter additional before can also using narrow-band light source or light source, capture apparatus obtain multispectral image.

Present invention also offers and a kind of obtain the method for reflection characteristic in multispectral image, reflection characteristic includes this pixel gray value under used wave band, reflectance and respective deformation thereof, such as grey scale pixel value difference under different-waveband, or normalized gray value, reflectance.The present invention mainly adopts reflectance.The present invention by arranging object of reference in the scene, the method for calibration obtain approximate reflectance, the difference according to embodiment, the implication of calibration is slightly different.

One provided by the invention obtains the embodiment of reflection characteristic (reflectance) in multispectral image:

Capture apparatus: B/W camera, selected wave band wave band such as: 420nm, 850nm.Before narrow band pass filter is put in video camera, place the object of reference of known reflectivity in the scene.In order to obtain more image detail, make the image averaging gray scale under each wave band between 180-220.The gray scale of P point in image:

$I_i^P=t\&Integral;s^P\left(\mathrm{\&lambda;}\right)c\left(\mathrm{\&lambda;}\right)[p_i^P\left(\mathrm{\&lambda;}\right)+A^P\left(\mathrm{\&lambda;}\right)]L\left(\mathrm{\&lambda;}\right)d\left(\mathrm{\&lambda;}\right)$

Wherein i: i-th wave band, t: camera exposure time, s:P point place object reflectance, c: camera responds, p: multispectral light is at the light distribution at P point place, A: ambient light is at the light distribution at P point place, L: filter spectral characteristic.What the present invention used is narrow band pass filter, the sign of integration of dieing:

$I_i^P=t{s}^P\left(\mathrm{\&lambda;}\right)c\left(\mathrm{\&lambda;}\right)[p_i^P+A^P\left(\mathrm{\&lambda;}\right)]L\left(\mathrm{\&lambda;}\right)$

Assume that multispectral and ambient light light distribution is uniform, then camera response time is definite value, and camera response c and filter spectral characteristic L is definite value under the same band, and its I of object of reference for known reflectivity is also detectable.Thus can obtain

$\mathrm{tc}\left(\mathrm{\&lambda;}\right)[p_i^P+A^P\left(\mathrm{\&lambda;}\right)]L\left(\mathrm{\&lambda;}\right)=\frac{I_{\mathrm{ref}}}{s_{\mathrm{ref}}}$

For pixel any in image, its reflectance:

$s=s_{\mathrm{ref}}\frac{I}{I_{\mathrm{ref}}}$

In the ideal case, such as uniform illumination, narrow band filter bandwidth are sufficiently narrow, this embodiment the reflectance obtained is approximately equal to real reflectance.But what the method provided obtains reflectance is mainly used in the skin detection in the present invention, therefore its degree of accuracy is acceptable.By the reflectance obtained or its deformation in skin detection, such as normalization, it is trained as feature.

Provided by the invention another obtains the embodiment of reflection characteristic (reflectance trend) in multispectral image:

Capture apparatus: B/W camera, selected wave band wave band: 420nm, 850nm.Before narrow band pass filter is put in multispectral light source, place the object of reference of known reflectivity in the scene for the calibration process before shooting.The gray scale of P point in captured image:

$I_i^P=t\&Integral;s^P\left(\mathrm{\&lambda;}\right)c\left(\mathrm{\&lambda;}\right)[p_i^P\left(\mathrm{\&lambda;}\right)L\left(\mathrm{\&lambda;}\right)+A^P\left(\mathrm{\&lambda;}\right)]d\left(\mathrm{\&lambda;}\right)$

When being left out ambient light

$I_{\mathrm{idark}}^P=t\&Integral;s^P\left(\mathrm{\&lambda;}\right)c\left(\mathrm{\&lambda;}\right)\left[p_i^P\left(\mathrm{\&lambda;}\right)L\left(\mathrm{\&lambda;}\right)\right]d\left(\mathrm{\&lambda;}\right)$

Calibration: by above formula, t is definite value, for the object of reference of known reflectivity, by regulating the light intensity regulating p of selected wave band light source^P(λ) value of L (λ), is ensureing that selected wave band hypograph average gray value is when 180-220, makes tc (λ) p under each wave band^P(λ) value of L (λ) is equal.So without ambient light when, grey scale pixel value trend just can represent reflectance trend, and even gray value just can represent the reflectance under its corresponding wave band after being multiplied by a weight factor w.But this kind of method is impracticable when there being ambient light, in order to solve this problem, this embodiment is made a return journey except the impact of ambient light by designing the deformation of reflection characteristic.

Calibration process complete after when there being ambient light, the impact of ambient light can be got rid of by following formula.

$I_i^P-I_j^P=t\&Integral;s^P\left(\mathrm{\&lambda;}\right)c\left(\mathrm{\&lambda;}\right)[p_i^P\left(\mathrm{\&lambda;}\right)L\left(\mathrm{\&lambda;}\right)+A^P\left(\mathrm{\&lambda;}\right)]d\left(\mathrm{\&lambda;}\right)-t\&Integral;s^P\left(\mathrm{\&lambda;}\right)c\left(\mathrm{\&lambda;}\right)[p_j^P\left(\mathrm{\&lambda;}\right)L\left(\mathrm{\&lambda;}\right)+A^P\left(\mathrm{\&lambda;}\right)]d\left(\mathrm{\&lambda;}\right)$ $=t\&Integral;s^P\left(\mathrm{\&lambda;}\right)c\left(\mathrm{\&lambda;}\right)[p_i^P\left(\mathrm{\&lambda;}\right)L\left(\mathrm{\&lambda;}\right)-p_j^P\left(\mathrm{\&lambda;}\right)L\left(\mathrm{\&lambda;}\right)]d\left(\mathrm{\&lambda;}\right)$

Using the pixel difference under the different-waveband characteristic as pixel, using this characteristic as training characteristics.In order to reduce the impact of ambient light further, it is possible to difference is normalized.

Can according to the multispectral image of system gained and then obtain the reflection characteristic of pixel in image by method provided by the invention.Difference according to embodiment, the characteristic of the pixel of required calculating is different.

A kind of method that invention also provides skin detection, extracts the reflection characteristic of just sample (skin) and negative sample (non-skin object) in multispectral image as feature, it is thus achieved that training set and test set.Method (such as SVM, ADABOOST) training utilizing machine learning obtains grader, then uses this grader to carry out skin detection.

Accompanying drawing illustrates:

Fig. 1 skin and the contrast of dummy's reflectance

Fig. 2 system constructs

Fig. 3 trains flow process

Fig. 4 shoots acquisition image, the positive and negative sample instance in 420nm, 800nm place

The reflectance of the actual acquisition of Fig. 5

Fig. 6 testing process

Fig. 7 testing result

Fig. 8 object of reference reflectance

Detailed description of the invention:

Equipment connects:

Processor 300 is connected with light source 100 to control light source switch and output brightness by controller 101, and processor 300 is connected with capture apparatus by capture card.Due to that shoot in order that obtain multispectral image, so processor 300 should ensure that when each wave band light source is opened, capture apparatus energy sync pulse jamming image.If the image such as obtained under two wave bands of 420nm, 850nm.Processor 300 should ensure that when the light source of 420nm wave band is opened, and capture apparatus starts shooting, and when the light source of 420nm wave band is closed, capture apparatus stops shooting.

Scene is arranged:

According to method provided by the invention, in order to be able to obtain the reflection characteristic of pixel in image, it is necessary to place the object of reference of known reflectivity in the scene for obtaining the reflection characteristic of pixel in image.The present invention uses PTFE plate as object of reference.

Owing to the invention provides various embodiments, and the detailed description of the invention of different embodiment is also slightly different, then this explains as far as the embodiment of above-mentioned two embodiments.The experiment carried out in the present invention, selected light source is the spectral distribution xenon lamp (can select nature light depending on embodiment difference) at 350nm-2500nm, and the wave band of selection is the narrow band pass filter of 420nm, 850nm, and bandwidth is 20nm, and capture apparatus is black and white camera.

Embodiment one:

Connect at equipment, after scene arranged, by optical filter 401(420nm) be put in camera 200 before.Then process it is calibrated: the intensity of illumination being regulated light source 100 by processor 300 makes to be taken scene gray value in picture between 180-220, and the image namely shot under wave band 420nm after having calibrated is denoted as pic1.The process of shooting 850nm image (being denoted as pic2) is identical with above procedure, it is only necessary to change optical filter.

Sample extraction: choose the pixel at skin and non-skin place respectively as positive negative sample.

Feature extraction: the method according to providing in [0016] calculates pixel reflectivity Characteristics under selected wave band, then this pixel characteristic value is a bivector being made up of the reflectance under wave band 420nm and wave band 800nm.

Training: calculate all eigenvalues the labelling of selected sample, obtains grader with SVM training.

Detection: obtain pic1, pic2 by same method, calculates the eigenvalue of wherein each pixel, and obtains final result with detection of classifier

Embodiment two:

Connect at equipment, after scene arranged, by optical filter 401(420nm) be put in light source 100 before.Then being calibrated, calibrating principle and method are as shown in [0021], and the image shot under wave band 420nm after having calibrated is denoted as pic1, it should be noted that the calibration of the present embodiment needs to carry out when dark (without ambient light).Obtain pic2 in the same way.

Sample extraction: choose the pixel at skin and non-skin place respectively as positive negative sample.

Feature extraction: the purpose of the present embodiment calibration is to characterize its reflectance trend by the gray value trend of pixel, then using the difference of the gray value after the normalization feature as this pixel.Namely the present embodiment is characterized by one-dimensional.

Training: calculate all eigenvalues the labelling of selected sample, obtains grader with SVM training.

Detection: obtain pic1, pic2 by same method, calculates the eigenvalue of wherein each pixel, and obtains final result with detection of classifier.

Claims (5)

1., by obtaining the method that in two wave band multispectral images, object reflectance carries out human body skin detection in multi-optical spectrum imaging system, said method comprising the steps of:

1a) shooting: be used for shooting multispectral image before the optical filter of selected two wave bands is put in capture apparatus camera lens respectively, and in the scene that is taken, place object of reference, for obtaining the reflection characteristic of each pixel in image, make the average gray of image to ensure not lose image information as far as possible between 180-220 by regulating multispectral light source controller change light-source brightness；

1b) calibration: utilize object of reference reflection characteristic obtain other object reflectance characteristics in image to implement step as follows:

1. processor obtains the gray value of certain pixel of object of reference in image or some region of average gray, and reads this reflection characteristic of object of reference reflectance preserved；

2. processor reads the grey scale pixel value of shooting objects in images；

3. processor calculates the reflectivity Characteristics obtaining objects in images according to below equation:

1c) training: extract the positive sample of skin and non-skin object negative sample from the picture of shooting, input as grader using the reflectivity Characteristics of each pixel, utilizes the method training grader of machine learning；

1d) detection: the grader obtained with the training stage carries out human body skin detection.

2., by obtaining the method that in two wave band multispectral images, object difference reflectance signature carries out human body skin detection in multi-optical spectrum imaging system, said method comprising the steps of:

2a) shooting: to obtain narrow-band light source before the optical filter of selected two wave bands is put in multispectral light source respectively, and in the scene that is taken, place object of reference, for obtaining the reflection characteristic of each pixel in image；

2b) calibration: utilize the step that realizes that object of reference reflection characteristic obtains objects in images difference reflectance signature to include:

1. under the dark surrounds without ambient light, processor obtains certain grey scale pixel value of object of reference position in the spectrum picture of first wave band or a certain area grayscale meansigma methods by capture apparatus, read this reflection characteristic of object of reference reflectance preserved, and the ratio both calculating, preserve result of calculation；

2. the intensity of illumination that processor changes under second band by regulating multispectral light source controller makes gray value under this wave band equal with the result under the ratio of reflectance and first wave band, it may be assumed that

3. after completing Illumination adjusting under dark surrounds, in the shooting environmental having ambient light, same pixel difference under two wave bands is utilized to make a return journey except the impact of ambient light, it is thus achieved that object difference reflectance signature, in order to difference is normalized by the impact weakening ambient light further uneven brought；

2c) training: extract the positive sample of skin and non-skin object negative sample from the picture of shooting, input as grader using the object difference reflectance signature of each pixel, utilizes the method training of machine learning to obtain grader；

2d) detection: the grader obtained with the training stage carries out human body skin detection.

3. method as described in claim 1 or 2, needs in the scene that is taken to place object of reference, and the reflection characteristic of object of reference must be known.

4. method as described in claim 1 or 2, multispectral image under acquired two wave bands, a wave band is at 420nm, and another wave band is between 780-850nm.

5. method as described in claim 1 or 2, considers skin detection as two classification problems, calculates the reflection characteristic of each pixel in multispectral image when detection, then obtains final result with grader classification.