Détection de visage avec Lambda (Python) + Rekognition
Détection de visage avec Lambda (Python) + Rekognition
Cet article est un article pratique pour passer en revue et corriger les connaissances acquises en développant Serverless Web App Mosaic. C'est l'un des w2or3w / items / 87b57dfdbcf218de91e2).
Ce serait bien de lire cet article après avoir regardé ce qui suit.
- Traitement d'image avec Lambda + OpenCV (création d'image grise)
- Utilisez AppSync à l'avant et à l'arrière
introduction
OpenCV est la première personne à proposer un moyen de réaliser la détection de visage, mais il est assez difficile de le rendre convaincant avec OpenCV. J'ai donc essayé d'utiliser cette AWS Rekognition et j'ai été impressionné. Très précis et rapide, quel que soit l'angle ou la rotation!
contenu
Autoriser Lambda à utiliser Rekognition
Pour que Rekognition soit disponible à partir de Lambda, vous devez attacher une stratégie au rôle IAM de la fonction Lambda.
Sélectionnez le rôle IAM souhaité dans AWS Console> IAM> Rôles
J'ai sélectionné ʻAmazonRekognitionFullAccess` sur l'écran affiché en appuyant sur le bouton "Joindre la politique" et je l'ai joint.
Mise à jour de la fonction Lambda (Python3)
lambda_function.py
# coding: UTF-8
import boto3
import os
from urllib.parse import unquote_plus
import numpy as np
import cv2
import logging
logger = logging.getLogger()
logger.setLevel(logging.INFO)
s3 = boto3.client("s3")
rekognition = boto3.client('rekognition')
from gql import gql, Client
from gql.transport.requests import RequestsHTTPTransport
ENDPOINT = "https://**************************.appsync-api.ap-northeast-1.amazonaws.com/graphql"
API_KEY = "da2-**************************"
_headers = {
"Content-Type": "application/graphql",
"x-api-key": API_KEY,
}
_transport = RequestsHTTPTransport(
headers = _headers,
url = ENDPOINT,
use_json = True,
)
_client = Client(
transport = _transport,
fetch_schema_from_transport = True,
)
def lambda_handler(event, context):
bucket = event["Records"][0]["s3"]["bucket"]["name"]
key = unquote_plus(event["Records"][0]["s3"]["object"]["key"], encoding="utf-8")
logger.info("Function Start (deploy from S3) : Bucket={0}, Key={1}" .format(bucket, key))
fileName = os.path.basename(key)
dirPath = os.path.dirname(key)
dirName = os.path.basename(dirPath)
orgFilePath = "/tmp/" + fileName
if (not key.startswith("public") or key.startswith("public/processed/")):
logger.info("don't process.")
return
apiCreateTable(dirName, key)
keyOut = key.replace("public", "public/processed", 1)
dirPathOut = os.path.dirname(keyOut)
try:
s3.download_file(Bucket=bucket, Key=key, Filename=orgFilePath)
orgImage = cv2.imread(orgFilePath)
grayImage = cv2.cvtColor(orgImage, cv2.COLOR_RGB2GRAY)
processedFileName = "gray-" + fileName
processedFilePath = "/tmp/" + processedFileName
uploadImage(grayImage, processedFilePath, bucket, os.path.join(dirPathOut, processedFileName), dirName)
detectFaces(bucket, key, fileName, orgImage, dirName, dirPathOut)
except Exception as e:
logger.exception(e)
raise e
finally:
if os.path.exists(orgFilePath):
os.remove(orgFilePath)
def uploadImage(image, localFilePath, bucket, s3Key, group):
logger.info("start uploadImage({0}, {1}, {2}, {3})".format(localFilePath, bucket, s3Key, group))
try:
cv2.imwrite(localFilePath, image)
s3.upload_file(Filename=localFilePath, Bucket=bucket, Key=s3Key)
apiCreateTable(group, s3Key)
except Exception as e:
logger.exception(e)
raise e
finally:
if os.path.exists(localFilePath):
os.remove(localFilePath)
def apiCreateTable(group, path):
logger.info("start apiCreateTable({0}, {1})".format(group, path))
try:
query = gql("""
mutation create {{
createSampleAppsyncTable(input:{{
group: \"{0}\"
path: \"{1}\"
}}){{
group path
}}
}}
""".format(group, path))
_client.execute(query)
except Exception as e:
logger.exception(e)
raise e
def detectFaces(bucket, key, fileName, image, group, dirPathOut):
logger.info("start detectFaces ({0}, {1}, {2}, {3}, {4})".format(bucket, key, fileName, group, dirPathOut))
try:
response = rekognition.detect_faces(
Image={
"S3Object": {
"Bucket": bucket,
"Name": key,
}
},
Attributes=[
"DEFAULT",
]
)
name, ext = os.path.splitext(fileName)
imgHeight = image.shape[0]
imgWidth = image.shape[1]
index = 0
for faceDetail in response["FaceDetails"]:
index += 1
faceFileName = "face_{0:03d}".format(index) + ext
box = faceDetail["BoundingBox"]
x = max(int(imgWidth * box["Left"]), 0)
y = max(int(imgHeight * box["Top"]), 0)
w = int(imgWidth * box["Width"])
h = int(imgHeight * box["Height"])
logger.info("BoundingBox({0},{1},{2},{3})".format(x, y, w, h))
faceImage = image[y:min(y+h, imgHeight-1), x:min(x+w, imgWidth)]
localFaceFilePath = os.path.join("/tmp/", faceFileName)
uploadImage(faceImage, localFaceFilePath, bucket, os.path.join(dirPathOut, faceFileName), group)
cv2.rectangle(image, (x, y), (x+w, y+h), (0, 0, 255), 3)
processedFileName = "faces-" + fileName
processedFilePath = "/tmp/" + processedFileName
uploadImage(image, processedFilePath, bucket, os.path.join(dirPathOut, processedFileName), group)
except Exception as e:
logger.exception(e)
raise e
Oui, ça ressemble à ça. Il n'y a pas d'endroit pour le ramasser et l'expliquer selon le code, mais il semble que ce soit trop d'omission.
La séquence de traitement ressemble à ce qui suit. ・ Image DL de S3 ・ Créer une image grise, télécharger sur S3, notifier par AppSync ・ Détection de visage avec Rekognition ・ Tournez une boucle pour chaque visage détecté Créer une image recadrée, télécharger sur S3, notifier avec AppSync, Dessiner le ROI sur l'image d'origine ・ Téléchargez l'image avec le ROI dessiné sur le visage vers S3 et notifiez-la avec AppSync
Le code est ci-dessous. https://github.com/ww2or3ww/sample_lambda_py_project/tree/work5
Appel de reconnaissance
Dans les deux lignes ci-dessous, vous pouvez obtenir le résultat de la détection de visage par JSON (réponse). C'est trop facile à utiliser.
rekognition = boto3.client('rekognition')
response = rekognition.detect_faces(Image= { "S3Object" : { "Bucket" : bucket, "Name" : key, } }, Attributes=[ "DEFAULT", ])
À la suite de la détection de visage pour cette image, le json suivant a été obtenu.
{
"FaceDetails":
[
{
"BoundingBox": {"Width": 0.189957395195961, "Height": 0.439284086227417, "Left": 0.1840812712907791, "Top": 0.41294121742248535},
"Landmarks":
[
{"Type": "eyeLeft", "X": 0.21208296716213226, "Y": 0.5631930232048035},
{"Type": "eyeRight", "X": 0.24809405207633972, "Y": 0.5793104767799377},
{"Type": "mouthLeft", "X": 0.2103935033082962, "Y": 0.7187585234642029},
{"Type": "mouthRight", "X": 0.23671720921993256, "Y": 0.7346710562705994},
{"Type": "nose", "X": 0.18016678094863892, "Y": 0.643562912940979}
],
"Pose": {"Roll": 6.634916305541992, "Yaw": -62.60176086425781, "Pitch": -6.222261905670166},
"Quality": {"Brightness": 73.63239288330078, "Sharpness": 86.86019134521484},
"Confidence": 99.99996185302734
},
{
"BoundingBox": {"Width": 0.19120590388774872, "Height": 0.3650752902030945, "Left": 0.6294564008712769, "Top": 0.18926405906677246},
"Landmarks":
[
{"Type": "eyeLeft", "X": 0.6734799146652222, "Y": 0.30800101161003113},
{"Type": "eyeRight", "X": 0.757991373538971, "Y": 0.33103394508361816},
{"Type": "mouthLeft", "X": 0.661914587020874, "Y": 0.4431125521659851},
{"Type": "mouthRight", "X": 0.7317981719970703, "Y": 0.4621959924697876},
{"Type": "nose", "X": 0.6971173882484436, "Y": 0.37951982021331787}
],
"Pose": {"Roll": 7.885405540466309, "Yaw": -19.28563690185547, "Pitch": 4.210807800292969},
"Quality": {"Brightness": 60.976707458496094, "Sharpness": 92.22801208496094},
"Confidence": 100.0
}
],
"ResponseMetadata":
{
"RequestId": "189aac7c-4357-4293-a424-fc024feeded0", "HTTPStatusCode": 200, "HTTPHeaders":
{
"content-type": "application/x-amz-json-1.1",
"date": "Sat, 04 Jan 2020 14:30:47 GMT",
"x-amzn-requestid": "189aac7c-4357-4293-a424-fc024feeded0",
"content-length": "1322",
"connection": "keep-alive"
},
"RetryAttempts": 0
}
}
Dans l'exemple de programme, ʻAttributes était spécifié comme DEFAULT, mais lorsque ʻALL était spécifié, les informations suivantes pouvaient être obtenues.
{
"FaceDetails":
[
{
"BoundingBox": {"Width": 0.189957395195961, "Height": 0.439284086227417, "Left": 0.1840812712907791, "Top": 0.41294121742248535},
"AgeRange": {"Low": 22, "High": 34},
"Smile": {"Value": false, "Confidence": 99.91419982910156},
"Eyeglasses": {"Value": false, "Confidence": 97.5216293334961},
"Sunglasses": {"Value": false, "Confidence": 98.94334411621094},
"Gender": {"Value": "Male", "Confidence": 99.5092544555664},
"Beard": {"Value": true, "Confidence": 87.53535461425781},
"Mustache": {"Value": false, "Confidence": 73.32454681396484},
"EyesOpen": {"Value": true, "Confidence": 98.92841339111328},
"MouthOpen": {"Value": false, "Confidence": 98.00538635253906},
"Emotions":
[
{"Type": "FEAR", "Confidence": 0.03440825268626213},
{"Type": "SURPRISED", "Confidence": 0.13240031898021698},
{"Type": "DISGUSTED", "Confidence": 0.03342699632048607},
{"Type": "ANGRY", "Confidence": 0.29975441098213196},
{"Type": "HAPPY", "Confidence": 0.022920485585927963},
{"Type": "CALM", "Confidence": 85.07475280761719},
{"Type": "CONFUSED", "Confidence": 1.6896910667419434},
{"Type": "SAD", "Confidence": 12.712653160095215}
],
"Landmarks":
[
{"Type": "eyeLeft", "X": 0.21208296716213226, "Y": 0.5631930232048035},
{"Type": "eyeRight", "X": 0.24809405207633972, "Y": 0.5793104767799377},
{"Type": "mouthLeft", "X": 0.2103935033082962, "Y": 0.7187585234642029},
{"Type": "mouthRight", "X": 0.23671720921993256, "Y": 0.7346710562705994},
{"Type": "nose", "X": 0.18016678094863892, "Y": 0.643562912940979},
{"Type": "leftEyeBrowLeft", "X": 0.2109173983335495, "Y": 0.5323911309242249},
{"Type": "leftEyeBrowRight", "X": 0.20237770676612854, "Y": 0.5220629572868347},
{"Type": "leftEyeBrowUp", "X": 0.20012125372886658, "Y": 0.5176519751548767},
{"Type": "rightEyeBrowLeft", "X": 0.22496788203716278, "Y": 0.5295209288597107},
{"Type": "rightEyeBrowRight", "X": 0.2825181782245636, "Y": 0.5552548170089722},
{"Type": "rightEyeBrowUp", "X": 0.24639180302619934, "Y": 0.5279281139373779},
{"Type": "leftEyeLeft", "X": 0.21062742173671722, "Y": 0.5640645027160645},
{"Type": "leftEyeRight", "X": 0.21973173320293427, "Y": 0.5715448260307312},
{"Type": "leftEyeUp", "X": 0.2089911699295044, "Y": 0.5593260526657104},
{"Type": "leftEyeDown", "X": 0.21014972031116486, "Y": 0.5721304416656494},
{"Type": "rightEyeLeft", "X": 0.24421700835227966, "Y": 0.5806354284286499},
{"Type": "rightEyeRight", "X": 0.2665697932243347, "Y": 0.5854082107543945},
{"Type": "rightEyeUp", "X": 0.2504902184009552, "Y": 0.5750172138214111},
{"Type": "rightEyeDown", "X": 0.25109195709228516, "Y": 0.5880314707756042},
{"Type": "noseLeft", "X": 0.19916994869709015, "Y": 0.6648411154747009},
{"Type": "noseRight", "X": 0.21807684004306793, "Y": 0.6632155179977417},
{"Type": "mouthUp", "X": 0.20222291350364685, "Y": 0.6922502517700195},
{"Type": "mouthDown", "X": 0.20738232135772705, "Y": 0.7338021993637085},
{"Type": "leftPupil", "X": 0.21208296716213226, "Y": 0.5631930232048035},
{"Type": "rightPupil", "X": 0.24809405207633972, "Y": 0.5793104767799377},
{"Type": "upperJawlineLeft", "X": 0.27225449681282043, "Y": 0.5730943083763123},
{"Type": "midJawlineLeft", "X": 0.2593783736228943, "Y": 0.7156036496162415},
{"Type": "chinBottom", "X": 0.22620755434036255, "Y": 0.8010575771331787},
{"Type": "midJawlineRight", "X": 0.3367012143135071, "Y": 0.74432772397995},
{"Type": "upperJawlineRight", "X": 0.36771708726882935, "Y": 0.6083425879478455}
],
"Pose": {"Roll": 6.634916305541992, "Yaw": -62.60176086425781, "Pitch": -6.222261905670166},
"Quality": {"Brightness": 73.63239288330078, "Sharpness": 86.86019134521484},
"Confidence": 99.99996185302734
},
{
"BoundingBox": {"Width": 0.19120590388774872, "Height": 0.3650752902030945, "Left": 0.6294564008712769, "Top": 0.18926405906677246},
"AgeRange": {"Low": 20, "High": 32},
"Smile": {"Value": false, "Confidence": 99.19612884521484},
"Eyeglasses": {"Value": false, "Confidence": 97.284912109375},
"Sunglasses": {"Value": false, "Confidence": 99.13030242919922},
"Gender": {"Value": "Female", "Confidence": 99.6273422241211},
"Beard": {"Value": false, "Confidence": 99.83914184570312},
"Mustache": {"Value": false, "Confidence": 99.87841033935547},
"EyesOpen": {"Value": true, "Confidence": 98.84789276123047},
"MouthOpen": {"Value": false, "Confidence": 95.55352783203125},
"Emotions":
[
{"Type": "FEAR", "Confidence": 0.3591834008693695},
{"Type": "SURPRISED", "Confidence": 0.5032361149787903},
{"Type": "DISGUSTED", "Confidence": 0.15358874201774597},
{"Type": "ANGRY", "Confidence": 2.0029523372650146},
{"Type": "HAPPY", "Confidence": 0.6409074664115906},
{"Type": "CALM", "Confidence": 89.09111022949219},
{"Type": "CONFUSED", "Confidence": 0.8823814988136292},
{"Type": "SAD", "Confidence": 6.366642475128174}
],
"Landmarks":
[
{"Type": "eyeLeft", "X": 0.6734799146652222, "Y": 0.30800101161003113},
{"Type": "eyeRight", "X": 0.757991373538971, "Y": 0.33103394508361816},
{"Type": "mouthLeft", "X": 0.661914587020874, "Y": 0.4431125521659851},
{"Type": "mouthRight", "X": 0.7317981719970703, "Y": 0.4621959924697876},
{"Type": "nose", "X": 0.6971173882484436, "Y": 0.37951982021331787},
{"Type": "leftEyeBrowLeft", "X": 0.6481514573097229, "Y": 0.2714482247829437},
{"Type": "leftEyeBrowRight", "X": 0.6928644776344299, "Y": 0.2690320312976837},
{"Type": "leftEyeBrowUp", "X": 0.6709408164024353, "Y": 0.2575661838054657},
{"Type": "rightEyeBrowLeft", "X": 0.7426562905311584, "Y": 0.28226032853126526},
{"Type": "rightEyeBrowRight", "X": 0.7986495494842529, "Y": 0.31319472193717957},
{"Type": "rightEyeBrowUp", "X": 0.7705841064453125, "Y": 0.28441154956817627},
{"Type": "leftEyeLeft", "X": 0.6606857180595398, "Y": 0.30426955223083496},
{"Type": "leftEyeRight", "X": 0.6901771426200867, "Y": 0.31324538588523865},
{"Type": "leftEyeUp", "X": 0.6742243766784668, "Y": 0.3005616068840027},
{"Type": "leftEyeDown", "X": 0.6734598278999329, "Y": 0.313093900680542},
{"Type": "rightEyeLeft", "X": 0.7402892112731934, "Y": 0.32695692777633667},
{"Type": "rightEyeRight", "X": 0.7727544903755188, "Y": 0.33527684211730957},
{"Type": "rightEyeUp", "X": 0.757353663444519, "Y": 0.32352718710899353},
{"Type": "rightEyeDown", "X": 0.7553724646568298, "Y": 0.33583202958106995},
{"Type": "noseLeft", "X": 0.6838077902793884, "Y": 0.39679819345474243},
{"Type": "noseRight", "X": 0.7161107659339905, "Y": 0.4051041901111603},
{"Type": "mouthUp", "X": 0.6949385404586792, "Y": 0.43140000104904175},
{"Type": "mouthDown", "X": 0.6908546686172485, "Y": 0.472693532705307},
{"Type": "leftPupil", "X": 0.6734799146652222, "Y": 0.30800101161003113},
{"Type": "rightPupil", "X": 0.757991373538971, "Y": 0.33103394508361816},
{"Type": "upperJawlineLeft", "X": 0.6373797655105591, "Y": 0.3141503930091858},
{"Type": "midJawlineLeft", "X": 0.6338266730308533, "Y": 0.46012476086616516},
{"Type": "chinBottom", "X": 0.6859143972396851, "Y": 0.5467866659164429},
{"Type": "midJawlineRight", "X": 0.7851454615592957, "Y": 0.5020546913146973},
{"Type": "upperJawlineRight", "X": 0.8258264064788818, "Y": 0.3661481738090515}
],
"Pose": {"Roll": 7.885405540466309, "Yaw": -19.28563690185547, "Pitch": 4.210807800292969},
"Quality": {"Brightness": 60.976707458496094, "Sharpness": 92.22801208496094},
"Confidence": 100.0
}
],
"ResponseMetadata":
{
"RequestId": "77b4dbdb-e76b-4940-927e-7548f3e0b602", "HTTPStatusCode": 200, "HTTPHeaders":
{
"content-type": "application/x-amz-json-1.1",
"date": "Sat, 04 Jan 2020 14:48:25 GMT",
"x-amzn-requestid": "77b4dbdb-e76b-4940-927e-7548f3e0b602",
"content-length": "6668", "connection": "keep-alive"
},
"RetryAttempts": 0
}
}
Les informations ont considérablement augmenté. Vous pouvez obtenir des informations telles que le sexe, si vous riez, l'âge prévu et les sentiments. Les types de points de repère augmentent également.
Contrôle de fonctionnement
Lorsque vous téléchargez un fichier à partir de l'application Web, vous verrez une image grise, une image recadrée du visage et une image montrant le retour sur investissement de tous les visages.
C'est un peu comme une application.
Épilogue
OpenCV est certainement une bibliothèque de traitement d'image très utile et puissante. Cela fait environ 10 ans, mais j'ai pensé que ce serait extrêmement pratique de l'utiliser pour les raisons suivantes. Calibrage de la caméra, correction de la distorsion, extraction de points caractéristiques, acquisition d'informations de parallaxe à partir d'images stéréo, correspondance de motifs, etc.
Cependant, encore une fois, j'ai senti que ces bibliothèques pratiques deviendraient plus faciles à utiliser en devenant des services Web. En outre, des modèles appris par machine sont utilisés de l'autre côté du service Web, et des résultats rapides, de haute précision et de haute qualité peuvent être attendus. C'est merveilleux, n'est-ce pas?
Faites attention à l'environnement et aux conditions lors de l'acquisition de données, recherchez des paramètres qui vous donneront les résultats attendus par essais et erreurs, et cela renverra un beau résultat sans un tel effort, c'est donc comme par magie C'est comme ressentir.
Au fait, il y a environ 10 ans, je pense que les seuls échantillons d'OpenCV étaient du C ++, mais maintenant j'ai l'impression qu'il y a plus d'échantillons Python. C'est un oncle qui a ressenti le temps même dans un tel endroit.
Recommended Posts