MySQL 人脸向量，欧几里得距离相似查询

前言

如标题，就是通过提取的人脸特征向量，写一个欧几里得 SQL 语句，查询数据库里相似度排前 TOP_K 个的数据记录。做法虽然另类，业务层市面上有现成的面部检索 API，技术层现在有向量数据库。

用 MySQL 关系型存储 128 维人脸向量，先是进行欧式距离计算就要对每维循环，开根号后还要排序。数据一旦特别多的时候，这查询速度可想而知。但是可能就是有特别小众的需求，下面就从特征提取到 SQL 拼写几个步骤开始梳理。

环境

python 3.8

opencv-python 4.8

dlib 19.24.2

mysql 5.7

特征向量提取

其实这一步可以跳过，但是没有这个，随便造的人脸向量数据不标准。这里就以读取图片人脸方式，用 dlib 库的人脸关键点检测器和面部识别模型，生成一个 128D 向量。这里有一个注意的是在生成向量时，windows 可能会出现下面错误。

Could not locate zlibwapi.dll. Please make sure it is in your library path!

确实 zlibwapi 库，这里要先下载，这里我已经放在 gitee.com/gaoxingqufuhchao/opencv_demo 上面。解压后 zlibwapi.lib 文件放到 CUDA 安装位置的 lib 中，zlibwapi.dll 文件放到 CUDA 安装位置的 bin 中，最后执行下面代码。

import cv2
import dlib
import os

images_path = os.path.join("./imgs/", "62.jpg")
# 返回图片的像素RGB值数组
images = cv2.imread(images_path)

# 人脸检测器
face_detector = dlib.get_frontal_face_detector()
faces = face_detector(images, 1)
face = faces[0]  # 取第一张脸

# 人脸的上下左右坐标
face_left = face.left()
face_right = face.right()
face_top = face.top()
face_bottom = face.bottom()

# 绘制图片上人脸矩形位置
# rectangle_img = images
# cv2.rectangle(rectangle_img, (face_left, face_top), (face_right, face_bottom), (255, 0, 0), 2)
# cv2.imwrite('./imgs/draw_imgs/rectangle_test.png', rectangle_img)

# 人脸关键点预测器
predictor = dlib.shape_predictor("./lib/dlib/shape_predictor_68_face_landmarks.dat")
shape = predictor(images, face)

# 绘制图片上人脸特征点位置
point_img = images
for p in range(0, 68):
    cv2.circle(point_img, (shape.part(p).x, shape.part(p).y), 2, (0, 255, 0))
# cv2.imwrite('./imgs/draw_imgs/point_test.png', point_img)

# 面部识别模型
recognition_model = dlib.face_recognition_model_v1("./lib/dlib/dlib_face_recognition_resnet_model_v1.dat")
features = recognition_model.compute_face_descriptor(point_img, shape)
features_list = list(features)

print(features_list)

cv2.imshow("frame", point_img)

# 无限期等待用户按键，因此窗口会保持打开
cv2.waitKey(0)

cv2.destroyAllWindows()

向量数据采集入库 (表结构)

CREATE TABLE `face_data` (
  `face_id` bigint(20) NOT NULL COMMENT '编号',
  `face_name` varchar(255) COLLATE utf8_unicode_ci NOT NULL COMMENT '人脸名称',
  `feature_vector` varchar(10240) COLLATE utf8_unicode_ci DEFAULT NULL COMMENT '特征向量',
  `update_time` datetime DEFAULT NULL ON UPDATE CURRENT_TIMESTAMP COMMENT '更新时间',
  `create_time` datetime DEFAULT NULL COMMENT '创建时间',
  `data_id` bigint(20) NOT NULL AUTO_INCREMENT COMMENT '数据主键',
  PRIMARY KEY (`data_id`) USING BTREE
) ENGINE=MyISAM AUTO_INCREMENT=6 DEFAULT CHARSET=utf8 COLLATE=utf8_unicode_ci ROW_FORMAT=DYNAMIC;

自定义欧氏距离计算函数

由于欧几里得距离需要将两个点的每维数值求差，在 MySQL 中也就是要循环向量被逗号分割的每个数值，然后分别求差在平方累加，最后开根号就是距离值，数越小越相似。以下语句中用到了 SUBSTRING_INDEX 函数通过循环 1 到 128 取出指定长度串，最后的 SUBSTRING_INDEX 是取出最后的数值就是循环每维数值。

CREATE DEFINER=`root`@`localhost` FUNCTION `euclidean_distance`(vector1 VARCHAR(128), vector2 VARCHAR(128)) RETURNS float
    DETERMINISTIC
BEGIN
	DECLARE len1 INT;  
	DECLARE len2 INT;  
	DECLARE i INT default 1;  
	DECLARE sum FLOAT;  
	SET len1 = LENGTH(vector1);  
	SET len2 = LENGTH(vector2);  
	SET sum = 0;  
	while i<129 do
			SET sum = sum + POW(SUBSTRING_INDEX(SUBSTRING_INDEX(vector1, ',', i), ',', -1) - SUBSTRING_INDEX(SUBSTRING_INDEX(vector2, ',', i), ',', -1), 2);  
			SET i = i + 1;
	END while;
	RETURN SQRT(sum);  
END

SQL 语句

查询就要调用上面函数，通过指定向量与每条库里的向量值算距离，随后按顺序排列查询 TOP5

1. 测试计算两个2维向量点的距离

select euclidean_distance("-0.0781729444861412,0.110044464468956", "-0.12611718475818634,0.13194166123867035")


2. 查询指定向量最相似的前5条

SELECT 
face_id,
euclidean_distance('-0.0781729444861412,0.110044464468956,0.06956595182418823,-0.04195471480488777,-0.13797412812709808,-0.01051325537264347,-0.08628914505243301,-0.17037545144557953,0.10026334226131439,-0.12803488969802856,0.23731382191181183,-0.094626285135746,-0.196399986743927,-0.06725674867630005,-0.08001142740249634,0.15153078734874725,-0.14427754282951355,-0.18382087349891663,-0.05251418426632881,-0.031287774443626404,0.1022581234574318,0.0790780633687973,0.008861299604177475,0.020882003009319305,-0.06397232413291931,-0.2706142067909241,-0.107462577521801,-0.03476380184292793,0.043968893587589264,-0.05694590508937836,-0.06642886996269226,0.04123011231422424,-0.18131156265735626,-0.03576948121190071,0.04664576053619385,0.1326863318681717,0.003225848078727722,-0.047573719173669815,0.1651453822851181,-0.029888691380620003,-0.21578450500965118,0.08537548780441284,0.0766519159078598,0.2540043294429779,0.22466638684272766,0.06360028684139252,-0.019438330084085464,-0.1689058244228363,0.043127596378326416,-0.11864039301872253,0.15812252461910248,0.19117693603038788,0.09921862185001373,0.05779439955949783,0.020246407017111778,-0.15069229900836945,-0.029495960101485252,0.1517685353755951,-0.09771490097045898,0.055743880569934845,0.18273179233074188,-0.05583849176764488,0.004194958135485649,-0.11965519189834595,0.17633725702762604,0.07909074425697327,-0.09237003326416016,-0.22499722242355347,0.13512863218784332,-0.16307398676872253,-0.16911545395851135,0.07714464515447617,-0.16546982526779175,-0.1413039118051529,-0.266265332698822,0.022507959976792336,0.4201160669326782,0.12732553482055664,-0.2231403887271881,0.036990948021411896,0.010170978493988514,-0.00629305187612772,0.1687457412481308,0.10473113507032394,0.00399409607052803,-0.013406611979007721,-0.12538184225559235,0.005433365702629089,0.2502303421497345,-0.02349778823554516,-0.07866109907627106,0.21128004789352417,0.0021615102887153625,-0.0009294161573052406,0.06593659520149231,0.031842511147260666,0.008746307343244553,0.05120106786489487,-0.19502216577529907,-0.037121932953596115,-0.027497289702296257,-0.0628679022192955,-0.06642049551010132,0.09603795409202576,-0.10845164209604263,0.13225442171096802,0.005377473309636116,0.04529388248920441,-0.05451689288020134,-0.038580723106861115,-0.12509524822235107,0.016826599836349487,0.20309969782829285,-0.22045192122459412,0.2263873666524887,0.12779565155506134,0.118931345641613,0.12429559975862503,0.16083121299743652,0.15235833823680878,0.04558601975440979,-0.07488320767879486,-0.21370939910411835,-0.07309035211801529,0.0741523876786232,-0.013888441026210785,0.16118189692497253,0.10571793466806412', feature_vector) 
AS distance
from face_data 
order by distance asc limit 5

其他

在实际应用中，可能是有一个预登陆账号，然后人脸身份验证类似验证码，所以可能是先有了用户标识，也是取出库里已有的用户人脸特征，然后用库里的人脸向量和摄像头采集的向量做对比。下面就用 numpy 演示两个人脸向量的欧氏距离计算和 OBS 虚拟摄像头人脸采集。

欧氏距离

import os
import dlib
import glob
import numpy as np
import cv2

# 计算人脸特征向量的欧式距离
face_01 = [-0.11144012212753296,0.18558962643146515,0.0016858093440532684,-0.030448582023382187,-0.12307003140449524,-0.03573177754878998,-0.09033556282520294,-0.11640644073486328,0.1467275768518448,-0.020302172750234604,0.3062277138233185,-0.0453588105738163,-0.19605347514152527,-0.0651734471321106,-0.058096982538700104,0.16225126385688782,-0.21905581653118134,-0.1029929518699646,0.013068988919258118,0.04029736667871475,0.05545109510421753,0.048230767250061035,0.0634097084403038,0.08036009967327118,-0.024840131402015686,-0.30172669887542725,-0.07167031615972519,-0.04815453290939331,0.09843403100967407,-0.07043944299221039,-0.11155916005373001,0.06715665757656097,-0.1751995086669922,-0.09061796963214874,0.07401317358016968,0.06398393213748932,-0.05072097107768059,-0.021325048059225082,0.2341330200433731,-0.02269885316491127,-0.15249019861221313,0.05262574926018715,0.08381971716880798,0.3157000243663788,0.20683002471923828,0.07494638860225677,0.000228429795242846,-0.12610206007957458,0.046763330698013306,-0.1762111932039261,0.12894554436206818,0.18361130356788635,0.102451391518116,0.04922708123922348,0.006102066021412611,-0.17877379059791565,-0.04263974353671074,0.11318269371986389,-0.11839815974235535,0.07388490438461304,0.11303036659955978,-0.03952416777610779,0.013598351739346981,-0.09794041514396667,0.27298909425735474,0.08994753658771515,-0.11253070831298828,-0.18397797644138336,0.07603368163108826,-0.10888601094484329,-0.1495959311723709,0.06924907863140106,-0.13828963041305542,-0.162084698677063,-0.3005405366420746,0.017705515027046204,0.33277949690818787,0.1113307923078537,-0.21170896291732788,0.064120352268219,-0.011443777941167355,-0.08527453988790512,0.07820181548595428,0.11378934234380722,-0.055745672434568405,-0.016978316009044647,-0.15705107152462006,-0.021412618458271027,0.22962769865989685,-0.03235346078872681,-0.06656813621520996,0.22491349279880524,0.012072622776031494,0.029958534985780716,0.048272471874952316,0.0701068788766861,-0.043136656284332275,-0.02105512097477913,-0.20309726893901825,-0.07898224890232086,-0.0016847627703100443,-0.059206850826740265,-0.0697859525680542,0.14457839727401733,-0.17818854749202728,0.10272862762212753,0.012656494975090027,0.026718538254499435,-0.04024617373943329,-0.01651758886873722,-0.09254957735538483,-0.024667389690876007,0.13368789851665497,-0.22851689159870148,0.2718657851219177,0.10549997538328171,0.13540413975715637,0.13176411390304565,0.11028678715229034,0.06738141179084778,0.020261352881789207,-0.025652553886175156,-0.11222656071186066,-0.08695542067289352,0.0743352621793747,-0.005103417672216892,0.15722909569740295,0.08712321519851685]
face_02 = [-0.06928954273462296,0.11818723380565643,0.040963172912597656,-0.03380622714757919,-0.125632643699646,0.014028718695044518,-0.09829825907945633,-0.13999347388744354,0.10254613310098648,-0.13002724945545197,0.2407195270061493,-0.07027815282344818,-0.25125718116760254,-0.032886065542697906,-0.023477450013160706,0.13404496014118195,-0.16187764704227448,-0.18711647391319275,-0.03598808869719505,-0.023680226877331734,0.06534431874752045,0.06392928957939148,-0.02046673744916916,0.026214662939310074,-0.07153397798538208,-0.3270975947380066,-0.12721474468708038,-0.011338643729686737,0.04254578426480293,-0.07672133296728134,-0.05328984931111336,0.05578911304473877,-0.19925925135612488,-0.0457644984126091,0.05274736136198044,0.10184930264949799,-0.010766003280878067,-0.04257964715361595,0.14281891286373138,-0.0182943232357502,-0.21731841564178467,0.04613770544528961,0.09813840687274933,0.291894793510437,0.19322234392166138,0.052529361099004745,-0.04576069116592407,-0.12129618227481842,0.024770550429821014,-0.12463519722223282,0.14057938754558563,0.19306844472885132,0.11266665160655975,0.03293035924434662,0.032030634582042694,-0.09783944487571716,-0.01481545064598322,0.15577413141727448,-0.1594703644514084,0.05806709825992584,0.1744152009487152,-0.06528271734714508,-0.012003951705992222,-0.09977130591869354,0.15240584313869476,0.07093586027622223,-0.0664982721209526,-0.2397862672805786,0.117387555539608,-0.1688736081123352,-0.12494431436061859,0.11812026053667068,-0.15014362335205078,-0.17380347847938538,-0.25874194502830505,0.0031708646565675735,0.4290091395378113,0.12104129791259766,-0.1617836207151413,0.06288817524909973,0.02954815700650215,-0.07710471004247665,0.17532427608966827,0.08251089602708817,0.007482852786779404,-0.034130796790122986,-0.12541988492012024,0.07749714702367783,0.24795527756214142,0.007125057280063629,-0.06481210887432098,0.2083437591791153,-0.013705004006624222,-0.030975420027971268,0.07249104231595993,0.06181362271308899,-0.010871338658034801,0.0194531362503767,-0.1996041238307953,-0.08055830746889114,-0.018493879586458206,-0.047192126512527466,-0.06708985567092896,0.11301972717046738,-0.1046549454331398,0.10359721630811691,0.03471799194812775,0.04366254806518555,0.004359336569905281,-0.01676999032497406,-0.1265866607427597,-0.05440763011574745,0.17379546165466309,-0.25932636857032776,0.14294737577438354,0.11041377484798431,0.07096583396196365,0.06915411353111267,0.08656369149684906,0.10719767212867737,0.02909727394580841,-0.11240153759717941,-0.24186159670352936,-0.06347794830799103,0.07049560546875,-0.013076946139335632,0.1388940066099167,0.04212008789181709]
face_03 = [-0.12067002058029175,0.17179107666015625,0.005495276302099228,-0.0025373362004756927,-0.12682169675827026,-0.046837806701660156,-0.06272760033607483,-0.11881549656391144,0.131977841258049,-0.007651656866073608,0.30675050616264343,-0.06511575728654861,-0.19687950611114502,-0.05430234596133232,-0.061561085283756256,0.15335164964199066,-0.1840490847826004,-0.10880403220653534,0.003372782375663519,0.020364956930279732,0.07791364192962646,0.04028210788965225,0.06327678263187408,0.09058713912963867,-0.041233912110328674,-0.2671954035758972,-0.07270435988903046,-0.05469353869557381,0.10929089039564133,-0.07597753405570984,-0.12340617924928665,0.060622505843639374,-0.19064907729625702,-0.11399897933006287,0.09063999354839325,0.06096772477030754,-0.07348572462797165,-0.03499014303088188,0.2246333211660385,-0.04379507154226303,-0.14929009974002838,0.049476295709609985,0.09826625883579254,0.3019770681858063,0.1847151517868042,0.08934237062931061,-0.01426877174526453,-0.15356388688087463,0.03547971695661545,-0.16697046160697937,0.14273102581501007,0.18264640867710114,0.109134241938591,0.06253807246685028,0.01051153801381588,-0.16892385482788086,-0.024090711027383804,0.09763640910387039,-0.12262223660945892,0.07233531028032303,0.11958196014165878,-0.03347306326031685,0.018053846433758736,-0.11581023037433624,0.24381747841835022,0.07473541796207428,-0.09524650871753693,-0.18174171447753906,0.07709110528230667,-0.12344227731227875,-0.1506771594285965,0.09598767012357712,-0.11854802072048187,-0.1694779396057129,-0.3036326766014099,0.01272799912840128,0.355130672454834,0.11748693883419037,-0.1921916902065277,0.06253601610660553,-0.008992472663521767,-0.08943237364292145,0.07774092257022858,0.09980118274688721,-0.06231308355927467,-0.023926906287670135,-0.15287624299526215,-0.012958861887454987,0.23298338055610657,-0.044136106967926025,-0.06738908588886261,0.2136751115322113,0.020884856581687927,0.019170459359884262,0.05304054170846939,0.07699457556009293,-0.04709392040967941,-0.013687841594219208,-0.1895902156829834,-0.07537028193473816,-0.009934772737324238,-0.056975316256284714,-0.06765390932559967,0.13928769528865814,-0.17901428043842316,0.08535695821046829,0.014542719349265099,0.011990601196885109,-0.04733775928616524,-0.022746175527572632,-0.11427918076515198,-0.02713892236351967,0.11226294934749603,-0.2461158186197281,0.2587936222553253,0.10381574183702469,0.14032721519470215,0.11607213318347931,0.10593312233686447,0.061977699398994446,0.03079364448785782,-0.03547735884785652,-0.12035234272480011,-0.08648461103439331,0.09446684271097183,-0.008687352761626244,0.14321735501289368,0.08460132777690887]
# 融合黄
face_04 = [-0.0781729444861412,0.110044464468956,0.06956595182418823,-0.04195471480488777,-0.13797412812709808,-0.01051325537264347,-0.08628914505243301,-0.17037545144557953,0.10026334226131439,-0.12803488969802856,0.23731382191181183,-0.094626285135746,-0.196399986743927,-0.06725674867630005,-0.08001142740249634,0.15153078734874725,-0.14427754282951355,-0.18382087349891663,-0.05251418426632881,-0.031287774443626404,0.1022581234574318,0.0790780633687973,0.008861299604177475,0.020882003009319305,-0.06397232413291931,-0.2706142067909241,-0.107462577521801,-0.03476380184292793,0.043968893587589264,-0.05694590508937836,-0.06642886996269226,0.04123011231422424,-0.18131156265735626,-0.03576948121190071,0.04664576053619385,0.1326863318681717,0.003225848078727722,-0.047573719173669815,0.1651453822851181,-0.029888691380620003,-0.21578450500965118,0.08537548780441284,0.0766519159078598,0.2540043294429779,0.22466638684272766,0.06360028684139252,-0.019438330084085464,-0.1689058244228363,0.043127596378326416,-0.11864039301872253,0.15812252461910248,0.19117693603038788,0.09921862185001373,0.05779439955949783,0.020246407017111778,-0.15069229900836945,-0.029495960101485252,0.1517685353755951,-0.09771490097045898,0.055743880569934845,0.18273179233074188,-0.05583849176764488,0.004194958135485649,-0.11965519189834595,0.17633725702762604,0.07909074425697327,-0.09237003326416016,-0.22499722242355347,0.13512863218784332,-0.16307398676872253,-0.16911545395851135,0.07714464515447617,-0.16546982526779175,-0.1413039118051529,-0.266265332698822,0.022507959976792336,0.4201160669326782,0.12732553482055664,-0.2231403887271881,0.036990948021411896,0.010170978493988514,-0.00629305187612772,0.1687457412481308,0.10473113507032394,0.00399409607052803,-0.013406611979007721,-0.12538184225559235,0.005433365702629089,0.2502303421497345,-0.02349778823554516,-0.07866109907627106,0.21128004789352417,0.0021615102887153625,-0.0009294161573052406,0.06593659520149231,0.031842511147260666,0.008746307343244553,0.05120106786489487,-0.19502216577529907,-0.037121932953596115,-0.027497289702296257,-0.0628679022192955,-0.06642049551010132,0.09603795409202576,-0.10845164209604263,0.13225442171096802,0.005377473309636116,0.04529388248920441,-0.05451689288020134,-0.038580723106861115,-0.12509524822235107,0.016826599836349487,0.20309969782829285,-0.22045192122459412,0.2263873666524887,0.12779565155506134,0.118931345641613,0.12429559975862503,0.16083121299743652,0.15235833823680878,0.04558601975440979,-0.07488320767879486,-0.21370939910411835,-0.07309035211801529,0.0741523876786232,-0.013888441026210785,0.16118189692497253,0.10571793466806412]
face_01_arr = np.array(face_01)
face_02_arr = np.array(face_02)
face_03_arr = np.array(face_03)
face_04_arr = np.array(face_04)

# 一般小于0.4大概率是同一个人
# distance = np.linalg.norm(face_02_arr - face_04_arr)
distance = np.linalg.norm(np.array([10,51]) - np.array([15,78]))
print(distance)
exit()

OBS 虚拟摄像头人脸采集

import cv2

indices = 0
# 获取OBS虚拟摄像头，真实摄像头填写IP地址
cap = cv2.VideoCapture(0)
face_xml = cv2.CascadeClassifier("models/haarcascade_frontalface_default.xml") #导入XML文件
while True:
    ret, frame = cap.read()
    if not ret:
        break

    gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)  # 转换为灰度图
    face = face_xml.detectMultiScale(gray, 1.3, 10)  # 检测人脸，并返回人脸位置信息

    if indices % 3000 == 0:
        print(face)

    indices += 1
    for (x, y, w, h) in face:
        cv2.rectangle(frame, (x, y), (x + w, y + h), (255, 0, 0), 2)

    cv2.imshow("frame", frame)

    if cv2.waitKey(1) & 0xFF == ord('q'):
        break

# 释放资源
cap.release()
cv2.destroyAllWindows()