DeepFace - Retinaface - DeepID

TestDeepFaceOneHot.py

#!/usr/bin/env python3
import os
import cv2
import json
import time
import pickle
from deepface import DeepFace
from deepface.modules import verification

# https://github.com/serengil/retinaface
# https://sefiks.com/2020/06/16/face-recognition-with-deepid-in-keras/
# https://sefiks.com/2020/02/23/face-alignment-for-face-recognition-in-python-within-opencv/

# https://en.wikipedia.org/wiki/Cosine_similarity
# https://en.wikipedia.org/wiki/Euclidean_distance

# something we forgot about ,
# if there are more than one faces found in the search ,
# then don't just take threshold cutoff , we need the lowest one in the whole set
# because they obviously can't appear more than once , and should be the lowest in the group

# add a thing that takes in multiple test inputs ? and uses same threshold
# like age 3 , 5 , 7 , 10 , 12 , 15 , 20 , etc
# so then we match at each age group ?

def write_pickle( file_path , python_object ):
	with open( file_path , 'wb') as f:
		pickle.dump( python_object , f , protocol=pickle.HIGHEST_PROTOCOL )

def read_pickle( file_path ):
	with open( file_path , 'rb' ) as f:
		return pickle.load( f )

def write_json( file_path , python_object ):
	with open( file_path , 'w' , encoding='utf-8' ) as f:
		json.dump( python_object , f , ensure_ascii=False , indent=4 )

def extract_faces( file_path ):
	faces = DeepFace.extract_faces(
		img_path=image_path ,
		detector_backend="retinaface" , # 'retinaface', 'mtcnn', 'ssd', 'dlib', 'mediapipe', 'yolov8', 'yolov11n',  'yolov11s', 'yolov11m', 'centerface'
		enforce_detection=False ,
		align=True ,
		normalize_face=False ,   # We'll handle normalization inside represent()
		# expand_percentage=0 ,
		# color_face='rgb' ,     # optional
		# grayscale=False ,      # optional
		# anti_spoofing=False ,  # optional
	)
	return faces

def get_ebedding( cropped_face ):
	representation = DeepFace.represent(
		img_path=cropped_face ,
		model_name="DeepID" , # "VGG-Face", "Facenet", "Facenet512", "OpenFace", "DeepFace", "DeepID", "ArcFace", "Dlib", "SFace", "GhostFaceNet"
		detector_backend="skip" , # NO detection; face is already cropped
		enforce_detection=False ,
		align=False , # we already aligned
		normalization="ArcFace" ,
		# anti_spoofing=False ,
	)
	if len( representation ) != 1:
		return []
	embedding = representation[ 0 ][ "embedding" ]
	return embedding

def get_test_ebedding( file_path ):
	representation = DeepFace.represent(
		img_path=file_path ,
		model_name="DeepID" , # "VGG-Face", "Facenet", "Facenet512", "OpenFace", "DeepFace", "DeepID", "ArcFace", "Dlib", "SFace", "GhostFaceNet"
		detector_backend="retinaface" , # NO detection; face is already cropped
		enforce_detection=True ,
		align=True ,
		normalization="ArcFace" ,
		# anti_spoofing=False ,
	)
	if len( representation ) != 1:
		return []
	embedding = representation[ 0 ][ "embedding" ]
	return embedding

def get_embedding_distance( alpha , beta ):
	distance = verification.find_distance( alpha , beta , "cosine" )
	return distance

SHOW_IMAGES = True
# image_path = "/Users/morpheous/WORKSPACE/PYTHON/FaceSorter/image-tests/Winner-2022/DSC_0031.JPG"
# image_path = "/Users/morpheous/WORKSPACE/PYTHON/FaceSorter/image-tests/Winner-2018/20181123_170259.jpg"
image_path = "/Users/morpheous/WORKSPACE/PYTHON/FaceSorter/image-tests/Winner-2019/20191130_165001.jpg"
test_path = "/Users/morpheous/WORKSPACE/PYTHON/FaceSorter/image-misc/1.png"

test_embedding = get_test_ebedding( test_path )
faces = extract_faces( image_path )
for i , face in enumerate( faces ):
	landmarks = face[ "facial_area" ]
	confidence = face[ "confidence" ]
	embedding = get_ebedding( face[ "face" ] )
	cosine_distance = verification.find_distance( test_embedding , embedding , "cosine" )
	euclidean_distance = verification.find_distance( test_embedding , embedding , "euclidean" )
	euclidean_l2_distance = verification.find_distance( test_embedding , embedding , "euclidean_l2" )
	print( f"Face #{i}:" )
	print( f"  Bounding box = {landmarks}" )
	print( f"  Confidence   = {confidence}" )
	print( f"  Embedding   = {len(embedding)}" )
	print( f"  Cosine Distance   = {cosine_distance}" )
	print( f"  Euclidean Distance   = {euclidean_distance}" )
	print( f"  Euclidean L2 Distance   = {euclidean_l2_distance}" )
	print( "" )

	if SHOW_IMAGES:
		# Convert RGB -> BGR for correct OpenCV color
		face_bgr = cv2.cvtColor( face[ "face" ] , cv2.COLOR_RGB2BGR )
		cv2.imshow( f"Face {i} (conf={confidence},cd={cosine_distance})" , face_bgr )

if SHOW_IMAGES:
	cv2.waitKey( 0 )
	cv2.destroyAllWindows()