CN110298815B - Method for detecting and evaluating skin pores - Google Patents
Method for detecting and evaluating skin pores Download PDFInfo
- Publication number
- CN110298815B CN110298815B CN201910239582.1A CN201910239582A CN110298815B CN 110298815 B CN110298815 B CN 110298815B CN 201910239582 A CN201910239582 A CN 201910239582A CN 110298815 B CN110298815 B CN 110298815B
- Authority
- CN
- China
- Prior art keywords
- skin
- image
- pores
- pore
- roughness
- 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.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0012—Biomedical image inspection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30004—Biomedical image processing
- G06T2207/30088—Skin; Dermal
Landscapes
- Engineering & Computer Science (AREA)
- Quality & Reliability (AREA)
- General Health & Medical Sciences (AREA)
- Medical Informatics (AREA)
- Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
- Radiology & Medical Imaging (AREA)
- Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Measurement Of The Respiration, Hearing Ability, Form, And Blood Characteristics Of Living Organisms (AREA)
- Measuring And Recording Apparatus For Diagnosis (AREA)
Abstract
The invention discloses a method for detecting and evaluating skin pores, which is used for detecting the pores and solving the problem of difficulty in detecting the pores by combining multi-color-layer images according to the pigment distribution characteristics of skin characteristics on different color layers. Firstly, setting reasonable threshold values for SURF and SIFT algorithms according to skin feature significance difference and K-means clustering to detect skin features on different pigment layers; then, on the basis of the difference of skin features on different skin pigment layers, euclidean distances are introduced to describe the similarity of position information of detection points of different pigment layers, and the optimal scale is used as a threshold value to effectively screen out interference items; and finally, constructing a skin pore roughness evaluation index by using the optimal scale in the SIFT algorithm on the basis of pore detection. The method improves the accuracy of pore detection, and the constructed pore evaluation index is more accurate and stable.
Description
Technical Field
The invention belongs to the field of facial pore detection, and particularly relates to a method for detecting and evaluating skin pores.
Background
The facial pore detection plays an important role in daily skin detection, medical cosmetology, cosmetic development, face recognition and facial reconstruction. Firstly, more and more beauty-conscious people pay attention to the personal facial skin condition, the pore problem is one of the prominent problems of the facial skin, and the premise for effectively solving the problem is scientific detection and evaluation of facial pores. The existing pore detection method basically depends on professional equipment such as a skin mirror and the like, so that the method is high in cost and large in place limitation, and brings inconvenience for daily skin detection, cosmetic development and skin medical research. Meanwhile, the lack of an objective evaluation method of pore conditions further adds difficulty to pore research. Secondly, because the characteristics of skin pores of people are unique, the pores can be used as a more unique and accurate characteristic in the fields of face recognition, face reconstruction and the like. Therefore, how to conveniently and accurately identify facial pores by using a common digital image and establish pore evaluation indexes adapting to the image is a difficult point of current research.
The existing methods for detecting pores by using common digital images mainly fall into two categories:
in the first category, pore detection is performed directly on the basis of skin images. Malathi, meena and the like utilize SIFT (Scale-invariant feature transform) algorithm to carry out fingerprint pore identification, and a new idea is provided for facial pore detection. Song et al used a special device to photograph and identify the pores of the nose by controlling a specific light, but this method has too high a requirement for the device light to be widespread. Li and the like refer to a fingerprint pore identification method, and directly act on facial skin images in a Bosphorus face library by using an SIFT algorithm to perform pore detection, so that the pore detection can be realized in a common digital image, and the application space of the pore detection is improved. However, this method does not remove other facial interference features outside the pores, such as nevi, pocks, pox marks, etc. Meanwhile, the method has the limitation that the selected pore detection area is a flat cheek area, so that no reflection or shadow exists, and in the actual situation, the image imaging can generate local reflection or shadow under the action of light, so that pores are difficult to accurately identify in the areas.
Second, pore detection is based on the melanin layer of the facial skin. Through research, researchers have found that the representation of skin tones on a defined color space (RGB, CIELAB, etc.) is not a true physical quantity, it is an abstract and derived human visual effect. The color of the skin is mainly determined by absorption and scattering of skin melanin, hemoglobin. Obtaining images of melanin and hemoglobin layers from skin images in a non-invasive manner allows more realistic study of skin features. Tsumura et al successfully separated a visible melanin layer hemoglobin layer image from a single human face image and eliminated the influence of facial shadows in the optical density field for the first time using Independent Component Analysis (ICA) technology. Carlos et al found that pores are mainly distributed in the melanin layer based on the absorption of light by different pigment layers, and performed pore detection using the separated melanin layer according to the existing research results. This method identifies facial pores more accurately, but still suffers from deficiencies. The facial part features are not only composed of single pigments, such as acne marks, pigmented nevi and the like, and in the process of separating the melanin layer, large features are broken into small features close to pores, so that new interference is introduced in pore identification, and inaccuracy still exists when pore detection is carried out by using the melanin layer.
In addition, some dermatological studies have given some universal pore evaluation criteria and ranges for pore evaluation methods. Francois et al propose pore sizes varying from 50 μm to 500 μm, with a range of 250 μm to 500 μm visible through the image. On this basis, flament et al suggested an index of the proportion of pores in the skin area and showed a maximum value of 25%. However, the existing pore evaluation method still needs to rely on professional equipment such as a skin mirror, and the method for evaluating the pores based on common digital images and the dermatology theory is few.
Disclosure of Invention
The invention aims to provide a method for detecting and evaluating skin pores, which fully utilizes the skin characteristic differences on different skin pigment layers, solves the problem that pleochroic layers are difficult to combine due to the differences, improves the pore detection accuracy and provides a pore roughness evaluation index suitable for images.
In order to achieve the purpose, the technical scheme of the invention is realized as follows:
a method of skin pore detection, comprising:
s1, separating a melanin layer image and a hemoglobin layer image based on an initial skin image by using an ICA (independent component analysis) method based on an image channel difference;
s2, detecting skin characteristics, screening out interference terms by utilizing characteristic significance differences of different skin characteristics on different pigment layer images, and finally detecting pores.
Further, the specific method of step S1 includes:
s11, eliminating upper epidermis reflection generated when the initial skin image is shot by a lens by controlling external conditions;
s12, constructing an equation of the initial skin image pixel color value and the pigment concentration on the basis of the Lambert-Beer law;
and S13, acquiring the pigment concentration and the separation matrix thereof by using ICA through the channel difference of R, G and B in the equation.
Further, the method for detecting skin features in step S2 includes:
s21, detecting all feature points including pores on a melanin layer image by using an SIFT algorithm;
s22, detecting a skin interference item with high significance on the initial skin image by using an SURF algorithm;
and S23, detecting all characteristic points containing hemoglobin on the hemoglobin layer image by using a SURF algorithm.
Furthermore, in step S22, using the response values representing the skin feature saliency in the SURF algorithm, clustering the response values of all the features on the initial skin image by K-means in the initial skin image, and selecting a reasonable threshold value according to the obvious difference of the response value distribution.
Further, the specific method for screening out the interference items in step S2 includes:
s31, screening interference items containing hemoglobin by using Euclidean distances of detection points on the initial skin image and the hemoglobin layer image;
s32, calculating the remarkable distribution range of the melanin at the characteristic point by using the optimal scale of the SURF algorithm characteristic point according to the Euclidean distance between the reserved melanin interference item and the detection point on the melanin layer, screening out the melanin interference item, and reserving pores.
Based on the detection method, the invention also provides a pore evaluation method, which comprises the following steps:
s41, obtaining the optimal size based on an SIFT algorithm according to the Tamura texture roughness calculation principle;
and S42, calculating the integral roughness of pores.
Further, the calculation principle of the Tamura texture roughness in step S41 includes:
s411, 2 of each pixel in the effective range is calculated k Mean gray value in neighborhood;
s412, calculating the average gray difference of the non-coincident neighborhood of each pixel in the vertical direction and the horizontal direction;
s413, calculating an optimal size parameter for each pixel that maximizes the average gray level difference;
and S414, calculating the roughness of the whole image.
Further, the optimal size method in step S41 includes:
s421, establishing images of the initial skin image under different scales, constructing a DOG pyramid, comparing each sampling point with all adjacent points on the basis, and searching for an extreme point in a scale space by judging whether the sampling point is larger or smaller than the adjacent points of an image domain and a scale domain;
and S422, determining the position and the scale of the key point through fitting of a ternary quadratic function according to the detected discrete extreme point.
Further, in step S42, the calculation formula of the pore overall roughness is:
wherein: p is crs Representing the overall roughness of facial pores in the selected region; m n is a pixel of the skin image; sigma S pbest Is the optimal size of the key point;
P crs larger indicates a coarser skin pore overall.
Compared with the prior art, the invention has the following beneficial effects:
the invention provides a pore detection algorithm integrating skin pigment distribution characteristics and an optimal scale, effectively improves the interference of other skin characteristics on pore detection, improves the accuracy of pore detection on common digital images, and provides a more reliable data base for pore research.
Drawings
FIG. 1 is a schematic view of a selected area of facial skin according to an embodiment of the present invention;
FIG. 2a is a face image 1 according to an embodiment of the present invention;
FIG. 2b is a skin selection of a face image 1&2 according to an embodiment of the present invention;
FIG. 2c is a face image 2 according to an embodiment of the present invention;
FIG. 2d is a diagram illustrating the result of pore detection directly based on skin images according to an embodiment of the present invention;
FIG. 2e is a schematic diagram illustrating the result of pore detection based on the melanin layer image according to an embodiment of the present invention;
FIG. 2f is a schematic illustration of the results of the pore detection algorithm of the present invention in an embodiment of the present invention;
FIG. 3a is a diagram illustrating an embodiment of the present invention in which a skin image with a significant skin pore size is extracted from an image set;
FIG. 3B is a B-skin image in an embodiment of the present invention;
FIG. 3C is a C skin image in an embodiment of the present invention;
FIG. 3D is a D-skin image in an embodiment of the present invention;
FIG. 3E is an E skin image in an embodiment of the present invention;
fig. 4 is an image of skin under illumination transformation in an embodiment of the invention.
Detailed Description
It should be noted that the embodiments and features of the embodiments may be combined with each other without conflict.
The technical scheme of the invention is further explained in detail by combining the attached drawings:
aiming at the problems of complex skin interference terms and difficulty in processing in the existing pore detection, on the basis of acquiring a skin pigment layer by an ICA (independent component analysis) method, a single skin image is not used any more, and a skin image with a multi-pigment layer is combined, so that a pore detection algorithm which integrates the distribution characteristics of skin pigments and the optimal scale is provided, the skin characteristic differences on different skin pigment layers are fully utilized, the problem that the multi-pigment layers are difficult to combine due to the differences is solved, and the accuracy of pore detection is improved. Firstly, the method utilizes the significance difference of different skin characteristics and the K-means clustering algorithm as SURF (Speeded-Up Robust Features) algorithm [20] And a SIFT algorithm sets a reasonable threshold value to detect the skin features. On the basis, euclidean distance is introduced to describe the similarity of the detected position information of the skin features on different pigment layers, and the skin feature information on different pigment layers is connected together; and then, an interference term from all skin features to pores is constructed by using pigment distribution characteristics and an optimal scale setting suprem threshold value in the SURF algorithm to screen out a funnel, so that the pores are detected more accurately. On the basis of the detection of the pores,aiming at the problem of scarce pore evaluation indexes, a pore roughness evaluation index suitable for an image, namely the integral roughness of skin pores, is provided by utilizing the optimal scale in the SIFT algorithm.
1. Skin pigment layer separation based on skin imaging model
The precondition for pore detection of facial skin is to clearly understand the imaging law of the skin. The skin is a complex multi-layered structure, and it is the multi-layered structure that determines the optical properties and imaging model of the skin.
In the imaging process of the skin, incident light irradiates the skin and firstly contacts the horny layer, a small part of the incident light generates upper epidermis reflection, the rest light penetrates through the horny layer to enter the epidermal layer, a large amount of melanin is contained in the epidermal layer, a part of the light is absorbed by the melanin, a part of the light is scattered and transmitted to enter the dermis layer, a large amount of hemoglobin is contained in the dermis layer, the light is absorbed by the hemoglobin and scattered and transmitted, and lower epidermis reflection is generated due to the fiber structure of the dermis layer and is reflected to the surface of the skin. The absorption of light by melanin and hemoglobin largely determines the true color of the skin. The color of the digital image is composed of the colors of three channels of R, G and B, and the value of the three channels reflects the amount of light which is reflected by the surface of an object and enters a lens, so that the RGB digital image is easily influenced by the change of illumination conditions to generate light reflection and shadow, and the real color of the skin is difficult to display. Therefore, in digital images taken under different lighting conditions, the detection of facial skin features of the same person at the same time is different, and many facial pores are difficult to detect due to reflection and shadows. However, this problem can be solved by obtaining a stable true pigment layer, i.e. melanin and hemoglobin.
An equation of the color value of the image pixel and the concentration of the pigment can be constructed on the basis of the traditional Lambert-Beer law:
wherein C is the logarithm of the difference of R, G and B channels, V represents the linear combination of the relative light absorption vectors of the melanin layer and the hemoglobin layer, C is the skin pigment concentration,
is a minimal amount related to the amount of incident spectral radiation.
C is the channel difference of R, G and B channels, is the linear combination of two pigments and is a mixed signal. Obtaining the unknown pigment concentration and its separation matrix is a blind separation problem with known differences of R, G, B channels. Since melanin and hemoglobin are independent and non-gaussian, this blind separation problem can be solved with ICA. The image channel difference-based ICA method is used herein to separate the pigment layers.
Notably, a relationship between the hypodermal reflectance and the pigment concentration is established here. Whereas the epithelium reflex is present in the skin image but it accounts for only 5% of the incident light, so in order to obtain an accurate skin image of the melanin layer and the hemoglobin layer, the epithelium reflex is eliminated by controlling the external conditions. The vast majority of the incident light can be controlled without directing the light directly at the face, a more efficient way is to place the polarizer for the lens, which can remove all the upper skin reflections.
2. Pore detection algorithm integrating skin pigment distribution characteristics and optimal scale
By separating the melanin layer and the hemoglobin layer, variable facial reflection and shadows generated by illumination are eliminated, more real facial skin color is obtained, and more accurate identification of facial pores is facilitated. On the basis, the related algorithm of the local feature saliency detection can detect pores, but the facial skin has not only one skin feature of the pores, but also has other complex and various skin features, such as acne, red blood streak, color spots and the like. Only those interference terms that are relative to the sweat pores can be screened out to accurately detect sweat pores.
According to the research of dermatology, the pore size ranges from 50 μm to 500 μm [15] At the initial skin imageThe feature significance is much less than the intact skin features of the majority of the face. The exception of the method is red blood streak, although the whole range of the red blood streak is generally larger, the local significance is basically close to facial pores due to the uneven distribution of the red blood streak, sporadic distribution of the red blood streak causes the local significance to be basically close to the facial pores, so that the red blood streak and the facial pores are difficult to distinguish by directly using an untreated skin image, the red blood streak is only distributed in a hemoglobin layer of the skin, and when the pore is detected by using the melanin layer, the red blood streak can be effectively screened.
The existing research and a large number of experiments prove that pores appear on a melanin layer, but other skin characteristics still exist on the melanin layer, and some characteristics consist of two pigments together. Therefore, the initial skin image, the melanin layer image and the hemoglobin layer image are used as the input of pore detection, interference items are screened out by using the characteristic significance difference of different skin characteristics on different pigment layers, and the pores are finally detected.
Firstly, basic facial skin features are classified into corresponding skin images, as shown in table 1, a pore detection algorithm which integrates skin pigment distribution characteristics and an optimal scale is constructed according to pigment distribution of the skin features, and the pore detection algorithm is composed of two parts, namely skin feature detection and interference item screening.
TABLE 1 basic skin characteristics of the face
2.1 detection of skin characteristics
On the basis of skin classification, according to the local feature significance difference of pores and other facial skin features, reasonable threshold values are set for SIFT and SURF algorithms to carry out pore identification and identification of interference terms. The SIFT algorithm is more accurate in positioning of micro features and more suitable for pore detection, but the running time is long, the time complexity of the whole algorithm is influenced by identifying interference items by using the SIFT algorithm, the SURF algorithm replaces a Gaussian second-order gradient template with a box function to position feature points, although the accuracy of pore detection is lower than that of SIFT, the detection of the interference items with remarkable features is still accurate, and the time complexity is far superior to SIFT. Therefore, SIFT is used to detect all feature points including pores on the melanin layer image, SURF is used to detect highly significant skin noise items on the initial skin image, and all feature points including hemoglobin are detected on the hemoglobin layer. In addition, pores and their distractors appear as dark spots on the image, so only the largest local feature points detected by the SIFT and SURF algorithms are retained.
Since all feature points including pores are detected on the melanin layer image using SIFT and all feature points including hemoglobin are detected on the hemoglobin layer image using SURF, the initial threshold value ω is an initial threshold value 2 、ω 3 Is set to 0. The SURF algorithm is used for detecting the obvious skin features on the initial skin image, the response values representing the skin feature significance in the algorithm are used for clustering the response values of all the features on the skin image by K-means in the skin image set due to the fact that the significance differences of pores, red blood streaks and other skin features are large and the significance degrees of the pores, the red blood streaks and other skin features are close to each other, and the reasonable threshold is selected according to the obvious difference of the response value distribution, and the specific algorithm is as follows:
and (4) inputting. K =2 in the initial skin image set, K-means.
And (6) outputting. Threshold ω for feature detection with SURF on initial skin images 1 ,。
And Step1, carrying out skin characteristic detection on the image set, and calculating a skin characteristic response value.
Step2, clustering the skin characteristic response value corresponding to each skin image by using K-means, calculating a clustering center set U with strong significance and a clustering center set V with weak significance, and on the basis, calculating a clustering radius set R of the region with weak significance and a maximum value set M of the region with weak significance, wherein the formula is as follows.
R＝V-A (2)
M＝V+R (3)
Where a is the minimum response value of the skin feature.
Step3, selecting the maximum value in the set M and setting the maximum value as a thresholdValue omega 1 。
2.2 interference item screening
On the basis of accurately identifying pores and interference terms, euclidean distances among three different skin image detection points are introduced to describe the similarity of the interference terms or the position information of the interference terms and the pores among different skin layers, and flexible supremum threshold values are set for the two different similarities by utilizing the skin pigment distribution characteristics to screen out the interference terms and accurately identify the pores.
Firstly, the Euclidean distance of detection points on an initial skin image and a hemoglobin layer image is utilized to effectively screen interference items containing hemoglobin. However, since some interference terms are composed of a plurality of pigments and the pigments are not uniformly distributed, in order to screen out only the interference at the hemoglobin position and to retain the melanin interference term to the maximum, the euclidean distance between the hemoglobin layer and the initial skin detection point is made smaller than a minimum value. On the basis, the retained melanin interference item and the Euclidean distance between detection points on a melanin layer are utilized to accurately screen out the melanin interference item and retain pores, the range of significant distribution of melanin at the characteristic points is calculated according to the optimal scale of the characteristic points of the SURF algorithm, and the melanin interference item is screened out as a flexible upper definite threshold.
The specific implementation algorithm is as follows:
inputting: initial skin, melanin layer, hemoglobin layer images;
and (3) outputting: detecting pores on the initial skin image;
step1, threshold selection based on pore detection in 2.1, set ω 3 Acquiring all position information L of detected points in a melanin layer image by using SIFT m ＝[x m ,y m ](ii) a Detecting the interference term on the initial skin image and the hemoglobin layer image by using SURF, and respectively setting a threshold value omega 1 And ω 2 The acquired interference item position information of the initial skin image and the hemoglobin layer image are respectively L s ＝[x s ,y s ]And L h ＝[x h ,y h ](ii) a And obtaining the optimal scale S of the interference item detected on the initial skin image according to the SURF algorithm s Features on the initial skin imageAll detection information of feature points is denoted as S = (L) s ；S s )。
Step2, screening interference items:
step2.1, using the location information L detected on the skin image s And position information L detected on the hemoglobin layer image h Removing interference terms containing hemoglobin and retaining interference terms containing melanin, so that L is introduced s And L h Euclidean distance between them
When d is satisfied E (L s (i),L h (j) ζ is less than or equal to i, j =1,2,3 \ 8230, the corresponding feature point is deleted and S is updated to S ', S' = (L) s '；S s '),L s '＝[x s ',y s ']Wherein L is s (i)∈L s ,L h (j)∈L h (ii) a ζ represents the maximum value of the similarity of the position information between the hemoglobin layer and the initial skin detection point, and since the pore sizes are all smaller than 1 pixel, the range of ζ is taken as (0, 1);
step2.2, using the updated location information L s ' and position information L detected on the hemoglobin layer image m Removing interference items on a melanin layer, and accurately positioning pores; introduction of L s ' and L m Euclidean distance between them
When d is satisfied E (L s '(p),L m (q))≤S s ' (p), p, q =1,2,3 \8230thatthe final sweat pore position information L is obtained by deleting the corresponding position information m Wherein, L s '(p)∈L s ',S s '(p)∈S s ',L m (q)∈L m ；S s ' (p) is L s The optimal scale of the corresponding melanin-containing interference term at the' p position.
Step3, using position information L m ' marking the location of the sweat pores on the initial skin image, fromThe pores are round and deep-colored and concave, so the positions of the pores are marked by using a point shape.
3. Skin pore integral roughness constructed based on Tamura texture roughness and optimal scale
The basic purpose of pore detection is to help researchers and ordinary people to visually recognize pores more intuitively, but it is difficult to objectively evaluate the pores only by vision, the existing pore detection methods all need skin detection, the pore size in the index can only be measured by a professional instrument, and the pore size calculated by the method is only the observation size due to different acquisition distances and resolutions of images. Furthermore, since pores are very fine concave pores, when one person visually perceives facial pores, the pores do not appear as single pores but as a whole the roughness of the pores in a certain region of the skin. Because human feelings are sensitive and inaccurate, in order to evaluate the roughness degree of facial pores better and objectively, the invention provides a more stable evaluation index, namely 'the integral roughness of skin pores', and the roughness condition of facial pores is more accurately known by quantized numbers. The skin pore overall roughness is established by taking Tamura texture roughness as reference and utilizing the optimal scale at each pore position in SIFT algorithm used in pore detection.
3.1 Tamura texture roughness
The Tamura roughness is calculated according to the following principle
First, 2 for each pixel in the effective range is calculated k Mean gray value in neighborhood
Then, for each pixel, the average gray-scale difference of its non-coincident neighbors in the vertical and horizontal directions is calculated. The difference value formula in the horizontal, vertical and diagonal directions is as follows:
E k,h (x,y)＝|A k (x+2 k-1 ,y)-A k (x-2 k-1 ,y)|
E k,v (x,y)＝|A k (x,y+2 k-1 )-A k (x,y-2 k-1 )|
E k,d (x,y)＝|A k (x+2 k-1 ,y+2 k-1 )-A k (x-2 k-1 ,y-2 k-1 )| (9)
on this basis, an optimal size parameter k that maximizes E (in any direction) is calculated for each pixel, as in equation (10).
S best (x,y)＝2 k
E k ＝E max ＝max(E 1 ,E 2 ,...,E L ) (10)
Finally, the roughness is calculated over the entire image of m x n, as in equation (11).
Actually, the Tamura roughness is calculated as the integral visual effect formed by accumulating the local feature saliency of the image, and the local feature saliency is the proportion of the optimal size range of each pixel point to the total number of the pixels of the image. According to the principle, the optimal size of the pore position in the SIFT algorithm, which can represent the significance of local features, is calculated, and then the overall roughness of the facial pores can be further calculated.
3.2 obtaining the optimal size based on SIFT algorithm
The scale space in SIFT describes the visual effect of the same object under the condition of continuously reducing the size, and the optimal size of the feature points can be obtained by utilizing the optimal scale calculation. Therefore, the optimal size of the pores can be calculated, and the integral roughness of the skin pores can be calculated.
The SIFT algorithm is used for positioning pore positions by using a DOG scale-space (DOG) generated by convolving different scales of a Gaussian difference kernel with an image, and the DOG is as follows.
D (x, y, α) represents the gaussian difference scale space of the skin image, L (x, y, α) represents the gaussian scale space of the skin image, (x, y) represents spatial coordinates, and α represents scale coordinates.
And establishing images of the skin image under different scales, and establishing a DOG pyramid. On the basis, each sampling point is compared with all the adjacent points to see whether the sampling point is larger or smaller than the adjacent points of the image domain and the scale domain, and the extreme point of the scale space is found.
The detected extreme points are discrete, and the positions and the scales of the key points can be accurately determined through ternary quadratic function fitting. Quadratic expansion of a scale space function D (x, y, int vl) centered at a key point is given by equation (13)
Wherein the first D on the right of the equal sign is the gray value at a certain key point, X = (X, y, int vl) T The offset amount centered at this point is obtained by a differential method since D (x) is discrete. Let the derivative of D (X) be zero, and the offset to obtain the precise extremum position is:
then substituting the obtained delta X into D (X) to obtain a scale space function at the accurate extreme point
On the basis, interference item points are eliminated, an optimal Scale value Scl at the pore position is obtained, the formula (16) is shown, the reciprocal of Scl represents the amplified proportion of the size, the optimal size Scale can be represented as the formula (17) because the extreme point is determined and is compared with adjacent pixels pairwise, and the optimal Scale function can be more accurately described by a circular formula as the pore shape is basically circular, and the formula (18) is shown.
Scl＝α 0 *2 ((intvl+Δx-1)/intvls) (16)
α 0 Is an initialScale coordinates, int vl being the number of layers corresponding to a feature point, int vls being the total number of layers
S pbest (x,y)＝πScale 2 (18)
3.3 skin pore Overall roughness
On the basis of Tamura roughness and the optimal scale value obtained by utilizing SIFT algorithm, the P is used crs The roughness of the skin pores can be reasonably, effectively and intuitively evaluated according to the index representing the overall roughness of the facial pores in the selected region, which is expressed by the formula (19).
m x n is the pixel of the skin image, P crs The larger the skin pore, the coarser the overall condition.
The overall roughness of skin pores is calculated by the proportion of the range of all local pore characteristic saliency salient to the skin area, namely, the proportion of pores to the skin area proposed by Flament and the like is calculated by using a digital image.
4. Results and analysis of the experiments
Because no standard data set of facial skin features exists at present, in order to facilitate a comparison experiment, the invention selects a Bosphorus face library [24] The front face image in (1) is used as the experimental image set. The frontal face image acquisition range of the Bosphorus face library is that the height is from the highest point of the forehead to the lowest point of the chin; the width is the maximum range from the left zygomatic arch to the right zygomatic arch; the pixels are around 1400 × 1400. The acquisition environment of the Bosphorus face library is that the face is acquired under indoor uniform illumination and has no epidermis reflection. The above conditions are suitable for the pore detection environment without epidermal reflection and high image definition proposed in the present aspect, so that pore detection is performed by using the frontal face image of the Bosphorus face library.
For more effective comparison test results, the area with the most gathered facial pores, namely the area with the width from the highest point of the human middle to the lowest point of the lower eyelid and the length from the left outer canthus to the right outer canthus, is selected, and is shown in the dotted line frame of fig. 1. And carrying out a comparison test on the area and manually marking the skin interference item in the area. And (3) manually marking the skin interference items in the experimental image set by 30 people according to different categories, and finally marking the interference which is determined consistently.
4.1 accuracy of pore detection
In order to verify that the algorithm improves the accuracy of pore detection and more effectively solves the problem that the facial skin interference item is difficult to screen, the pore detection algorithm based on the initial skin image and the pore detection algorithm utilizing the skin melanin layer are compared with the algorithm. The invention adopts the screening rate F (filter) of the skin interference term to evaluate the pore detection algorithm, which is defined as
Wherein, I is the number of certain skin interference items detected by using the pore detection algorithm, and M is the number of certain skin interference items marked manually.
As can be seen from Table 2, the screening rate of various interference items by using the pore detection algorithm of the invention is obviously higher than that of the existing two pore detection algorithms, which shows that the algorithm of the invention effectively solves the problem that other skin interference items are difficult to screen in pore detection, and improves the accuracy of pore detection. In the screening rate of the algorithm of the present invention, the screening rate of the color spots was 80. 3571% of the total number of the colored spots was less than that of the other interfering items, because some of the colored spots had larger areas and were not too prominent compared to the surrounding skin, and pores still existed in the colored spots, making it difficult to clearly distinguish the colored spots, and the screening rate was low.
TABLE 2 screening rate of each skin interference term under different pore detection algorithms
Fig. 2 shows the accuracy of pore detection more intuitively, with half of the skin area cut and the key part enlarged for easy detail viewing. Comparing fig. 2d and 2e, it can be seen that fig. 2e can better identify sweat pores in the shadow region than fig. 2d due to the elimination of the shadow effect, but neither effectively solves the problem of interference of other skin features with sweat pores. By comparing fig. 2d, 2e and 2f, fig. 2f can identify the sweat pores at the shadow, effectively remove the interference of skin features such as pits and pigmented nevus, and identify the sweat pores more accurately.
4.2 pore bulk roughness
The invention provides a pore evaluation index-integral roughness of skin pores, which is suitable for digital images, and the skin pore condition is evaluated more objectively. Therefore, the section verifies the correctness and stability of the indexes through experiments. And the verification is assisted by the dermatological indexes that the size of the real pores visible by the image is 250-500 mu m, the maximum proportion of the pores in the skin area is 25 percent, and the indexes measured by the image are the observed pore size and Tamura texture roughness.
4.2.1 verification of accuracy of pore roughness integrity
Selecting 9 pictures with obvious skin pore fineness degree from the image set, as shown in fig. 3a, line A to line C are skin images with fine pores, line D to line F are skin images with coarse pores, and calculating the integral roughness P of the pores respectively crs Recorded in Table 3, P of the skin image with fine pores can be seen crs P of skin image with rough pores, both less than 20 crs Are all more than 20, which basically indicates the overall pore roughness P crs The accuracy of (2).
In order to further verify the accuracy of the pore overall roughness, 4 face images with obviously different pore roughness in the skin area are selected for pore detection. Skin images are shown in fig. 3b to 3e, and it can be seen from observation that the skin pore roughness decreases from fig. 3b to 3e, and roughness ranks are recorded in table 4, and 1 to 4 are respectively marked from coarse to fine. Respectively calculating the integral roughness P of pores crs Average pore size d observed from the image, tamura texture roughness F crs The results are reported in table 2. By observing the data in the table, first, the overall pore roughness P in the table crs And average pore size d were consistent with the observed roughness order, but calculated for the imageThe obtained average pore size d does not accord with the range of the actual pore size of 250-mum to 500-mum, and the actual skin pore condition is difficult to reflect; and P is crs Then both are less than 25, below the maximum value of the proportion of pores in the skin area. Roughness F for Tamura texture crs Since it calculates the roughness of all the textures, all the facial features including the skin noise term and the nose are calculated, and therefore it does not conform to the observed roughness of the pores and cannot be used as an index for pore evaluation. In conclusion, the pore roughness P is described crs The method is close to the visual effect of people, accords with the dermatological index, and can be used as a more objective index to evaluate the roughness of pores.
Table 3 verification of pore roughness accuracy in whole body 1
Table 4 pore overall roughness accuracy verification 2
4.2.2 pore Overall roughness stability verification
The detected integral roughness of the pores has better stability under different illumination conditions without epidermis reflection, and the following experiment is carried out to change the light color and the light intensity of the skin image as shown in figure 4. In order to compare the stability of the integral pore roughness under the illumination change, the stability index Tamura texture roughness F for measuring the roughness is used crs The comparison was performed and the results are reported in table 5, on which a normalization process was performed and the variance was calculated to measure the stability. By observing the experimental results, the magnitude of the variance of the roughness of the whole pores and the variance of the roughness of Tamura textures is the same, but the fact that the roughness of the whole pores is basic stability can be shown.
TABLE 5 comparison of pore bulk roughness stability
The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents, improvements and the like that fall within the spirit and principle of the present invention are intended to be included therein.
Claims (7)
1. A method of skin pore detection, comprising:
s1, separating a melanin layer image and a hemoglobin layer image based on an initial skin image by using an ICA (independent component analysis) method based on an image channel difference;
s2, detecting skin characteristics, screening out interference items by utilizing characteristic significance differences of different skin characteristics on different pigment layer images, and finally detecting pores;
the specific method of step S1 includes:
s11, eliminating upper epidermis reflection generated when the initial skin image is shot by the lens by controlling external conditions;
s12, constructing an equation of the initial skin image pixel color value and the pigment concentration on the basis of the Lambert-Beer law;
s13, obtaining the pigment concentration and a separation matrix thereof by using ICA through channel differences of R, G and B in an equation;
the method for detecting the skin characteristics in the step S2 comprises the following steps:
s21, detecting all feature points including pores on a melanin layer image by using an SIFT algorithm;
s22, detecting a skin interference item with high significance on the initial skin image by using an SURF algorithm;
and S23, detecting all characteristic points containing hemoglobin on the hemoglobin layer image by using the SURF algorithm.
2. The method according to claim 1, wherein in step S22, the response values of all the features on the initial skin image are clustered by K-means in the initial skin image using the response values representing the skin feature saliency in the SURF algorithm, and a reasonable threshold is selected by the significant difference of the response value distribution.
3. The method according to claim 1, wherein the specific method for screening out the interference terms in step S2 comprises:
s31, screening interference items containing hemoglobin by using Euclidean distances of detection points on the initial skin image and the hemoglobin layer image;
s32, calculating the remarkable melanin distribution range at the characteristic point by using the retained Euclidean distance between the melanin interference item and the detection point on the melanin layer and the optimal scale of the characteristic point of the SURF algorithm, screening out the melanin interference item and retaining pores.
4. A pore evaluation method based on the detection method according to any one of claims 1 to 3, comprising:
s41, obtaining the optimal size based on an SIFT algorithm according to a Tamura texture roughness calculation principle;
and S42, calculating the integral roughness of pores.
5. The pore evaluation method according to claim 4, wherein the Tamura texture roughness calculation principle of step S41 comprises:
s411, 2 of each pixel in the effective range is calculated k Average gray value in neighborhood;
s412, calculating the average gray difference of the non-coincident neighborhood of each pixel in the vertical direction and the horizontal direction;
s413, calculating an optimal size parameter for each pixel that maximizes the average gray level difference;
and S414, calculating the roughness of the whole image.
6. The pore evaluation method according to claim 4, wherein the optimal size method of step S41 comprises:
s421, establishing images of the initial skin image under different scales, establishing a DOG pyramid, comparing each sampling point with all adjacent points on the basis, and finding an extreme point of a scale space according to whether the sampling point is larger or smaller than the adjacent points of the image domain and the scale domain;
and S422, determining the position and the scale of the key point through the fitting of a ternary quadratic function by the detected discrete extreme point.
7. The method for evaluating sweat pores according to claim 4, wherein in step S42, the calculation formula of the roughness of the whole sweat pores is:
wherein: p is crs Representing the overall roughness of facial pores in the selected region; m n is a pixel of the skin image; sigma s pbest Is the optimal size of the key point;
P crs larger indicates a coarser skin pore overall.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910239582.1A CN110298815B (en) | 2019-03-27 | 2019-03-27 | Method for detecting and evaluating skin pores |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910239582.1A CN110298815B (en) | 2019-03-27 | 2019-03-27 | Method for detecting and evaluating skin pores |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110298815A CN110298815A (en) | 2019-10-01 |
| CN110298815B true CN110298815B (en) | 2023-04-14 |
Family
ID=68026484
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910239582.1A Active CN110298815B (en) | 2019-03-27 | 2019-03-27 | Method for detecting and evaluating skin pores |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110298815B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111539932B (en) * | 2020-04-22 | 2023-03-14 | 四川省肿瘤医院 | Hemoglobin measuring instrument and method |
| CN113487573A (en) * | 2021-07-08 | 2021-10-08 | 杭州德肤修生物科技有限公司 | Cosmetic efficacy quantitative evaluation method based on accurate image comparison |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107169960A (en) * | 2017-05-15 | 2017-09-15 | 精诚工坊电子集成技术（北京）有限公司 | A kind of skin surface pore size detection method based on color of image space |
| CN109325938A (en) * | 2018-08-23 | 2019-02-12 | 武汉嫦娥医学抗衰机器人股份有限公司 | A kind of pigment separated method of the centralization image channel difference based on light intensity spatial |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2014027522A1 (en) * | 2012-08-17 | 2014-02-20 | ソニー株式会社 | Image processing device, image processing method, program, and image processing system |
-
2019
- 2019-03-27 CN CN201910239582.1A patent/CN110298815B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107169960A (en) * | 2017-05-15 | 2017-09-15 | 精诚工坊电子集成技术（北京）有限公司 | A kind of skin surface pore size detection method based on color of image space |
| CN109325938A (en) * | 2018-08-23 | 2019-02-12 | 武汉嫦娥医学抗衰机器人股份有限公司 | A kind of pigment separated method of the centralization image channel difference based on light intensity spatial |
Non-Patent Citations (1)
| Title |
|---|
| "Design and learn distinctive features from pore-scale facial keypoints";Dong Li etc.;《Pattern Recognition》;20141016;全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110298815A (en) | 2019-10-01 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111524080A (en) | Face skin feature identification method, terminal and computer equipment | |
| Boyce et al. | Multispectral iris analysis: A preliminary study51 | |
| CN109961426B (en) | Method for detecting skin of human face | |
| CN108549884A (en) | A kind of biopsy method and device | |
| CN104123543B (en) | A kind of eye movement recognition methods based on recognition of face | |
| US20040267102A1 (en) | Diagnostic feature extraction in dermatological examination | |
| CN110110637A (en) | A kind of method of face wrinkle of skin automatic identification and wrinkle severity automatic classification | |
| CN112396573A (en) | Facial skin analysis method and system based on image recognition | |
| Cula et al. | Bidirectional imaging and modeling of skin texture | |
| CA3204854A1 (en) | Systems and methods for evaluating pigmented tissue lesions | |
| CN110298815B (en) | Method for detecting and evaluating skin pores | |
| Pathan et al. | Classification of benign and malignant melanocytic lesions: A CAD tool | |
| KR101436988B1 (en) | Method and Apparatus of Skin Pigmentation Detection Using Projection Transformed Block Coefficient | |
| Zheng | Static and dynamic analysis of near infra-red dorsal hand vein images for biometric applications | |
| CN108288040A (en) | Multi-parameter face identification system based on face contour | |
| CN112790750A (en) | Fear and tension emotion recognition method based on video eye movement and heart rate analysis | |
| Wang et al. | Facial pore detection based on characteristics of skin pigment distribution | |
| Li | Biometric Person Identification Using Near-infrared Hand-dorsa Vein Images | |
| Liu et al. | Biological indexes based reflectional asymmetry for classifying cutaneous lesions | |
| Wang et al. | Image-based facial pore detection and visualization in skin health evaluation | |
| Rimskaya et al. | Development Of A Mobile Application For An Independent Express Assessment Of Pigmented Skin Lesions | |
| Madooei et al. | 2 A Bioinspired Color Representation for Dermoscopy Analysis | |
| Joshi Manisha et al. | Computerized pragmatic assessment of Prakriti Dosha using tongue images-Pilot study | |
| CN110796638A (en) | Pore detection method | |
| Husemann et al. | Computerised diagnosis of skin cancer using neural networks |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |