CN103268499A - Human body skin detection method based on multi-spectral imaging - Google Patents

Abstract

The invention discloses a human body skin detection method based on multi-spectral imaging. Reflection characteristics of an object under two selected wave bands are obtained by a calibration method in a multi-spectral imaging system and are used as features for skin detection. The multi-spectral imaging system comprises a reference with the known reflection characteristic, and light source and receiving equipment which are configured to shoot the images under the two selected wave bands. The reflection characteristics of the object are obtained by using the calibration method, namely the reflection characteristics of other objects in the images are forecast by using the reflection characteristic of the reference. According to the skin detection method, different reflection characteristics of human body skin and non-skin objects are used as features, and skin detection and non-skin detection are regarded as two-classification problems, so that a classifier is trained by using a machine learning method and is used for detecting the human body skin.

Description

Human body skin detection method based on multispectral imaging

Two, technical field:

The present invention relates to computer vision and image processing field, relate in particular to and utilize computer vision to carry out the method that human body skin detects.More specifically, the present invention relates to a kind of cheap and real-time multi-optical spectrum imaging system obtains image and obtains reflected by objects characteristic in the picture of publishing picture by system, method design.How to the present invention relates to human body skin under the complex background that dummy, mask are arranged simultaneously.

Three, background technology:

Skin detection is important research direction of computer vision field.For example, the detection of people's face is followed the tracks of in the video conference, the understanding of body language, and the indication analyses in the man-machine interaction etc. all need accurate skin detection algorithm.Tradition skin detection algorithm carries out human body skin tone testing based on the colouring information that the RGB camera obtains mostly, the principle of RGB camera is red by R(), G(is green), B(indigo plant) three kinds of Color Channel component combined color images, therefore modal skin detection algorithm is to reach testing goal by definition colour of skin interval on each component in color space.Like this, when the color-values of pixel drops in the interval of definition, think that then this pixel is skin.About the selection problem of color space, people such as Kakumanua have provided good summary.But when the detection crowd comprises from not agnate people, or the background color complexity, ambient light is when strong, and is just lower based on the skin detection precision of color.In addition, although use the color space with illumination unchangeability, can reduce ambient light to a certain extent according to the influence to algorithm.But when extraneous illumination variation was violent, the skin detection performance still reduced significantly.And, based on the article that the skin detection algorithm can not effectively detect and removal is close with the human body complexion color of RGB three look imagings, particularly can not distinguish true man and dummy (comprising body model, mask, photo etc.).Therefore, the skin detection algorithm based on RGB three look imagings has significant limitation.

The human body skin reflectivity has reached from physiology and experimentally has been proved its unique characteristic: lower at the ultraviolet portion reflectivity, the increase reflectivity with wavelength increases then, and " W " shape is arranged at the 575nm place.Skin reflectivity reaches peak value at the 800nm place, can occur descend at the 900nm place, increases again after reaching minimum value but the increase amplitude is very little.Through actual test, as figure one, dummy's reflectivity is not possess this specific character.Utilize this reflectivity Characteristics can distinguish corium skin and dummy and other article.But obtaining mostly of reflectivity will be used spectrometer at present, and therefore expensive, time-consuming length can not be measured in real time.

[reference listing]

[non-patent literature]

1、Elli?Angelopoulou，“Understanding?the?color?of?human?skin”，Computer?Science?Dept.,Stevens?Institute?of?Technology,?Castle?Point?on?Hudson,?Hoboken,?NJ?07030,?USA

The US6370260 of Ioannis Pavlidis shows a kind of 800nm-1400nm of use and two kinds of near-infrared band images of 1400nm-2200nm carry out the method that human body skin detects, thereby utilizes human body skin can distinguish true man and dummy in unique reflection characteristic of these two kinds of wave bands.But, the imaging device of these two wave bands is very expensive, is difficult in the common civil applications and uses.

201080002426 of Sony has proposed skin detection system and method under a kind of use multiband near infrared lighting condition, can the interior a plurality of care pixels of detected image, identification and the corresponding object of being scheduled to.Two different near-infrared wavelengths of this method use 870nm and 950nm obtain two width of cloth images under the corresponding wavelength, generate the binaryzation skin image according to the difference between two width of cloth images then, utilize binary image to extract skin image from original image.When being mainly used in the computer data typing, this method to the extraction of the skin (hand) that is sitting in the data typing personnel before the computer screen, can't be applied to the skin detection under the common scene.

Therefore consider above situation, need to seek a kind of multi-optical spectrum imaging system cheap, that real-time is good and obtain multispectral image.Need a kind of skin detecting method that can under common scene, can effectively distinguish true man and dummy simultaneously.

Goal of the invention

Fundamental purpose of the present invention provides a kind of multi-optical spectrum imaging system, and this system price is cheap, simple in structure, can obtain the reflection characteristic of object in image by this system.The invention provides a kind of detection method to complex background (especially containing the dummy) to carry out skin detection at the reflection characteristic of obtaining.

Another object of the present invention provides a kind of Calibration Method, obtains the reflection characteristic of object under each wavelength by this calibration steps.The effect of calibration is in order to eliminate the spectral response difference of imaging path (comprising light source, capture apparatus and optical filter) under different wave length, use a spectral response of calibrating imaging path on the different wave length at the known marker of reflectivity on the different wave length, thereby obtain the true reflection characteristic of object.Calibration steps is relevant with the reflection characteristic of required acquisition.Reflection characteristic can refer to gray-scale value, reflectivity and the distortion thereof of pixel herein, and for example the distortion of gray-scale value can make poor between normalized gray-scale value, the gray-scale value.

The present invention also provides a kind of skin detecting method, and this method is considered as two classification problems with skin detection, as feature, utilizes the method training classifier of machine learning to carry out skin detection reflection characteristic.

Four, summary of the invention:

Before system of the present invention, method and hardware are realized describing, need to prove that the present invention is not limited to described several embodiment and detection method, the present invention can have a plurality of possible embodiment that do not offer some clarification in this article.

Detection method provided by the invention is to utilize skin and noncutaneous different reflection characteristics under the multispectral irradiation in arrowband, carries out skin detection, thereby can effectively distinguish human body skin and anthropoid skin article (comprising body model, mask, photo etc.).

The invention provides a kind of system of multispectral imaging, this system cost is low, real-time good.By system provided by the present invention can obtain be concerned about image under the wave band.The selected light source of system must cover selected multi light spectrum hands (also can adopt natural light during conditions permit), if be 420nm and 850nm such as selected wave band, so selected light source just must comprise this two wave bands.The capture apparatus that native system adopts is B, and capture apparatus also can be selected according to the difference of embodiment, such as using photodiode as receiver.One embodiment of the present of invention adopt a capture apparatus, obtain multispectral image by the narrow band pass filter that switches different-waveband before capture apparatus successively.Also can adopt two capture apparatus in addition, the narrow band pass filter that installs 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, obtain multispectral image by capture apparatus.

The present invention also provides a kind of method of obtaining reflection characteristic in the multispectral image, reflection characteristic comprises gray-scale value, reflectivity and separately the distortion thereof of this pixel under the use wave band, such as grey scale pixel value poor under different-waveband, perhaps normalized gray-scale value, reflectivity.The present invention mainly adopts reflectivity.The reflectivity of the present invention by arranging that in scene object of reference, Calibration Method obtain to be similar to, according to the difference of embodiment, the implication of calibration is slightly different.

An embodiment who obtains reflection characteristic in the multispectral image (reflectivity) provided by the invention:

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

:

$I_i^P=t{\∫s}^P\left(\mathrm{\λ}\right)c\left(\mathrm{\λ}\right)[p_i^P\left(\mathrm{\λ}\right)+A^P\left(\mathrm{\λ}\right)]L\left(\mathrm{\λ}\right)d\left(\mathrm{\λ}\right)$

I wherein: i wave band, t: the camera exposure time, s:P point place object reflectance, c: the camera response, p: multispectral light is in the light distribution at P point place, A: ambient light is in the light distribution at P point place, L: optical filter spectral characteristic.The present invention uses is narrow band pass filter, the sign of integration of dieing:

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

It is that the camera response time is definite value so uniformly that the light of supposing multispectral and ambient light distributes, and camera response c and optical filter spectral characteristic L are definite values under same wave band, also is detectable for its I of object of reference of known reflectivity.So just can obtain

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

For any pixel in the image, its reflectivity:

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

In the ideal case, such as illumination evenly, the narrow band filter bandwidth is enough narrow, the reflectivity that is obtained by this embodiment is approximately equal to real reflectance.But this method provides obtains reflectivity and is mainly used in skin detection among the present invention, so its degree of accuracy is acceptable.To the reflectivity that obtain or its distortion as normalization, be trained as feature in the skin detection.

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

Capture apparatus: B, selected wave band wave band: 420nm, 850nm.Before narrow band pass filter is put in multispectral light source, the calibration process before the object of reference of placing known reflectivity in scene is used for taking.The gray scale that P is ordered in the captured image

:

$I_i^P=t{\∫s}^P\left(\mathrm{\λ}\right)c\left(\mathrm{\λ}\right)[p_i^P\left(\mathrm{\λ}\right)L\left(\mathrm{\λ}\right)+A^P\left(\mathrm{\λ}\right)]d\left(\mathrm{\λ}\right)$

Under the situation of not considering ambient light

$I_{\mathrm{idark}}^P=t{\∫s}^P\left(\mathrm{\λ}\right)c\left(\mathrm{\λ}\right)\left[p_i^P\left(\mathrm{\λ}\right)L\left(\mathrm{\λ}\right)\right]d\left(\mathrm{\λ}\right)$

Calibration: by following 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 guaranteeing to make tc under each wave band (λ) p under the condition of selected wave band hypograph average gray value at 180-220 ^P(λ) value of L (λ) equates.Under the condition of no ambient light, grey scale pixel value trend just can be represented reflectivity trend like this, even gray-scale value multiply by the reflectivity that just can represent behind the weight factor w under its corresponding wave band.But this kind method is impracticable having under the condition of ambient light, and in order to address this problem, this embodiment removes the influence of ambient light by the distortion of design reflectivity characteristic.

Having under the situation of ambient light after calibration process is finished, can get rid of the influence of ambient light by following formula

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

$=t{\∫s}^P\left(\mathrm{\λ}\right)c\left(\mathrm{\λ}\right)[p_i^P\left(\mathrm{\λ}\right)L\left(\mathrm{\λ}\right)+p_j^P\left(\mathrm{\λ}\right)L\left(\mathrm{\λ}\right)]d\left(\mathrm{\λ}\right)$

With the characteristic of the difference of pixel under different-waveband as pixel, with this specific character as training characteristics.In order further to reduce the influence of ambient light, can carry out normalization to difference.

Can and then obtain the reflection characteristic of pixel in the image according to the multispectral image of system's gained by method provided by the invention.According to the difference of embodiment, the characteristic difference of the pixel of required calculating.

The present invention provides a kind of method of skin detection simultaneously, and the reflection characteristic of positive sample (skin) and negative sample (non-skin object) obtains training set and test set as feature in the extraction multispectral image.Utilize method (such as SVM, the ADABOOST) training of machine learning to obtain sorter, use this sorter to carry out skin detection then.

Description of drawings:

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

Fig. 2 system structure

Fig. 3 trains flow process

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

The actual reflectivity that obtains of Fig. 5

Fig. 6 testing process

Fig. 7 testing result

Fig. 8 object of reference reflectivity

Embodiment:

Equipment connects:

Processor 300 links to each other with light source 100 in order to control light source switch and output brightness by controller 101, and processor 300 links to each other with capture apparatus by capture card.Because the purpose of taking is in order to obtain multispectral image, so processor 300 should be able to guarantee that when each wave band light source is opened capture apparatus is photographic images synchronously.Such as if obtain image under two wave bands of 420nm, 850nm.When processor 300 should be able to guarantee that the light source of 420nm wave band is opened, capture apparatus began to take, and when the light source of 420nm wave band was closed, capture apparatus stopped to take.

Scene is arranged:

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

Because the invention provides various embodiments, and the embodiment of different embodiment is also slightly different, this only explains with regard to the embodiment of above-mentioned two embodiment again.The experiment of carrying out among the present invention, selected light source be spectral distribution at the xenon lamp (look the embodiment difference and can select natural light) of 350nm-2500nm, the wave band of selection is the narrow band pass filter of 420nm, 850nm, bandwidth is 20nm, capture apparatus is the black and white camera.

Embodiment one:

Connect at equipment, after scene arranges and finishes, with optical filter 401(420nm) be put in camera 200 before.Carry out calibration process then: the intensity of illumination of regulating light sources 100 by processor 300 makes and is taken the gray-scale value of scene in picture between 180-220, and the image note of namely taking after calibration is finished under the wave band 420nm is pic1.The process of taking 850nm image (note is pic2) is identical with above process, only need change optical filter.

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

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

Training: calculate all eigenwerts and the mark of selected sample, obtain sorter with the SVM training.

Detect: by obtaining pic1, pic2 with quadrat method, calculate the wherein eigenwert of each pixel, and obtain final result with the sorter detection

Embodiment two:

Connect at equipment, after scene arranges and finishes, with optical filter 401(420nm) be put in light source 100 before.Calibrate then, calibrating principle and method are as shown in [0021], and the image note of taking after calibration is finished under the wave band 420nm is pic1, it should be noted that the calibration of present embodiment need be carried out under the condition of dark (no 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 present embodiment calibration is that the gray-scale value trend with pixel characterizes its reflectivity trend, then with the difference of the gray-scale value after the normalization feature as this pixel.Be that the present embodiment feature is one dimension.

Training: calculate all eigenwerts and the mark of selected sample, obtain sorter with the SVM training.

Detect: by obtaining pic1, pic2 with quadrat method, calculate the wherein eigenwert of each pixel, and obtain final result with the sorter detection.

Claims (8)

1. one kind is obtained under selected two wave bands reflected by objects characteristic in the multispectral image by Calibration Method in multi-optical spectrum imaging system, and carries out the method that human body skin detects based on reflection characteristic, and described method comprises the following steps that realize with processor:

A) place object of reference in the scene that is taken, object of reference is used for obtaining the reflection characteristic of other pixels of image;

B) calibration: the embodiment difference, the reflection characteristic difference that need obtain, calibration steps is relevant with the reflection characteristic that need obtain.Calibration is that processor obtains other object reflection characteristics by the reflection characteristic of known object of reference;

C) training: from picture shot, extract positive sample (skin) and negative sample (non-skin object), utilize the method training of machine learning to obtain sorter;

D) detect: the sorter that obtains with the training stage carries out the human body skin detection.

2. multi-optical spectrum imaging system described in claim 1 need be arranged object of reference in scene, and the reflection characteristic of object of reference must be known.

3. method described in claim 1, reflection characteristic comprises gamma characteristic, reflectivity Characteristics, also comprises the distortion based on the two, as the normalization of reflectivity or difference each other.

4. two wave bands described in claim 1, first wave band is 420nm, second wave band can be between 780-850nm.

5. method described in claim 1 is utilized the object of reference reflection characteristic to obtain in the image other reflected by objects characteristics and need be passed through Calibration Method.An embodiment who obtains reflection characteristic by calibration steps comprises the following steps that realize with processor:

A) optical filter with selected two wave bands is put in the preceding shooting multispectral image that is used for of capture apparatus camera lens respectively, processor calculates the average gray value of the image of taking, and processor changes multispectral light source brightness by the controller of regulating multispectral light source makes the average gray of image not lose image information with assurance between 180-220 as far as possible;

B) processor obtains the gray-scale value of certain pixel of object of reference or the average gray in a certain zone by capture apparatus, and reads the object of reference reflection characteristic (reflectivity) of having preserved;

C) processor is by taking the grey scale pixel value that equipment obtains object in the image;

D) processor obtains reflected by objects characteristic in the image according to the information of the object of reference that obtains by calculating, for example:

。

6. method described in claim 1, another embodiment that obtains reflection characteristic by Calibration Method comprises the following steps that realize with processor:

A) under dark surrounds (no ambient light), the multispectral light source of selected wave band must be the arrowband, need not install optical filter additional before the capture apparatus camera lens.Processor obtains certain grey scale pixel value of object of reference position under first wave band of multispectral light source or the average gray in a certain zone by capture apparatus, processor reads the object of reference reflection characteristic (reflectivity) of having preserved, and processor calculates the ratio of the two and preserves result of calculation;

B) according to the step in last step, the same gray-scale value of object of reference and the ratio of reflectivity of calculating under second wave band, processor change intensity of illumination under this wave band by the controller of regulating multispectral light source makes gray-scale value and the ratio of reflectivity under this wave band equate with ratio under first wave band:

C) under dark surrounds, finish calibration after, when the influencing of ambient light arranged, be the multispectral light source of arrowband and before the capture apparatus camera lens, do not add optical filter because this embodiment adopts.So can utilize the difference of same pixel under different-waveband to remove the influence of ambient light.Can also carry out normalization to difference in order further to weaken the inhomogeneous influence that brings of ambient light.

7. as obtaining the method for reflection characteristic in claim 5 and 6, positive sample (skin) and negative sample (non-skin object) reflection characteristic as feature, are utilized the method for machine learning to train and obtained sorter.

8. method described in claim 1 is considered skin detection as two classification problems, calculates the reflection characteristic of each pixel in the multispectral image when detecting, and obtains net result with the sorter classification then.

Images