face_aligned_photo_generator

 이번 포스팅에서는 저번 포스팅에 이어서 사진을 정면을 보게끔 조정한뒤 얼굴 부분만 잘라내어서 face_recognition 의 성능을 향상시킬수 있는 방법를 정리해보겠다. 다음 코드를 통해 특정 폴더에 위치한 사진들을 불러와서 전처리 해준 다음 새로운 폴더에 저장할수 있게 된다.

import numpy as np
import dlib
import cv2

RIGHT_EYE=list(range(36,42))
LEFT_EYE=list(range(42,48))
EYES=list(range(36,48))

dataset_paths=['./image/tedy-front/','./image/son-front/','./image/unknown-front/']
output_paths=['./image/tedy-align/','./image/son-align/','./image/unknown-align/']
number_images=20
image_type='.jpg'

predictor_file='./model/shape_predictor_68_face_landmarks.dat'
MARGIN_RATIO=1.5
# 마지막에 얻어지는 사진의 size
OUTPUT_SIZE=(96,96) 

detector=dlib.get_frontal_face_detector()
predictor=dlib.shape_predictor(predictor_file)

def getFaceDimension(rect):
    #얼굴을 가리키는 사각형의 왼쪽 상단 (꼭지점 좌표, 너비, 높이)
    return (rect.left(),rect.top(),rect.right()-rect.left(),rect.bottom()-rect.top())

def getCropDimension(rect, center):
    width = (rect.right() - rect.left())
    half_width = width // 2
    (centerX, centerY) = center
    startX = centerX - half_width
    endX = centerX + half_width
    startY = rect.top()
    endY = rect.bottom() 
    return (startX, endX, startY, endY) 

for (i,dataset_path) in enumerate(dataset_paths):
    # extract the person name from names
    output_path=output_paths[i]

    for idx in range(number_images):
        input_file=dataset_path + str(idx+1) + image_type

        # get RGB image from BGR, OpenCV format
        image=cv2.imread(input_file)
        image_origin=image.copy()

        (image_height,image_width) = image.shape[:2]
        gray=cv2.cvtColor(image,cv2.COLOR_BGR2GRAY)

        rects=detector(gray,1)

        for (i,rect) in enumerate(rects): #검출된 사람의 숫자만큼 반복
            (x,y,w,h) = getFaceDimension(rect)
            cv2.rectangle(image,(x,y),(x+w,y+h),(0,255,0),2)

            points=np.matrix([[p.x,p.y] for p in predictor(gray,rect).parts()]) #68개의 점의 좌표
            show_parts=points[EYES] #눈에 해당하는 점

            right_eye_center=np.mean(points[RIGHT_EYE],axis=0).astype('int')
            left_eye_center=np.mean(points[LEFT_EYE],axis=0).astype('int')

            eye_delta_x=right_eye_center[0,0] - left_eye_center[0,0]
            eye_delta_y=right_eye_center[0,1] - left_eye_center[0,1]
            degree=np.degrees(np.arctan2(eye_delta_y,eye_delta_x)) - 180

            eye_distance=np.sqrt((eye_delta_x ** 2) + (eye_delta_y ** 2))
            # 사진을 돌린 이후 눈 사이의 거리 조정
            aligned_eye_distance=left_eye_center[0,0] - right_eye_center[0,0]
            scale=aligned_eye_distance / eye_distance

            eyes_center=((left_eye_center[0,0] + right_eye_center[0,0])//2,
            (left_eye_center[0,1] + right_eye_center[0,1])//2)

            matrix=cv2.getRotationMatrix2D(eyes_center,degree,scale)
            wraped=cv2.warpAffine(image_origin,matrix,(image_width,image_height),
            flags=cv2.INTER_CUBIC)

            (startX,endX,startY,endY) = getCropDimension(rect,eyes_center)

            croped=wraped[startY:endY,startX:endX]
            output=cv2.resize(croped,OUTPUT_SIZE)

            output_file=output_path + str(idx+1) + image_type
            cv2.imshow(output_file,output)
            cv2.imwrite(output_file,output)

cv2.waitKey(0)
cv2.destroyAllWindows()