人脸识别之人脸矫正

一般来说，使用 mtcnn 网络检测到人脸后，都需要进行矫正。而对于人脸矫正，最简单的可以通过使用仿射变换来实现。

思路：

• 通过 mtcnn 模型检测到人脸的 5 个特征点：左眼，右眼，鼻子，左嘴角，右嘴角；

bounding_boxes = {
        'box': [int(bounding_box[0]), int(bounding_box[1]),
                int(bounding_box[2]), int(bounding_box[3])],
        'confidence': bounding_box[-1],
        'keypoints': {
            'left_eye': (int(keypoints[0]), int(keypoints[5])),
            'right_eye': (int(keypoints[1]), int(keypoints[6])),
            'nose': (int(keypoints[2]), int(keypoints[7])),
            'mouth_left': (int(keypoints[3]), int(keypoints[8])),
            'mouth_right': (int(keypoints[4]), int(keypoints[9])),
        }
    }

bounding_box = bounding_boxes['box']
keypoints = bounding_boxes['keypoints']

• 计算双眼中心点的位置 eye_center 和双眼连线的倾斜角度 angle；
• 通过 eye_center 和 angle 得到仿射变换矩阵 rot_matrix；
• 将整张图片进行仿射变换得到 align_image；

def align_face(image, keypoints, scale=1.0):
    eye_center = (
            (keypoints['left_eye'][0] + keypoints['right_eye'][0]) * 0.5,
            (keypoints['left_eye'][1] + keypoints['right_eye'][1]) * 0.5,
            )
    dx = keypoints['right_eye'][0] - keypoints['left_eye'][0]
    dy = keypoints['right_eye'][1] - keypoints['left_eye'][1]

    angle = cv2.fastAtan2(dy, dx)
    rot_matrix = cv2.getRotationMatrix2D(eye_center, angle, scale=scale)
    rot_image = cv2.warpAffine(image, rot_matrix, dsize=(image.shape[1], image.shape[0]))
    return rot_image

align_image = align_face(image, keypoints)

• 根据从 原图检测的 2D 框从 align_image 图片中抠出人脸并保存。

xmin = bounding_box[0]
ymin = bounding_box[1]
xmax = bounding_box[2]
ymax = bounding_box[3]
crop_image = align_image[ymin:ymax, xmin:xmax, :]
cv2.imwrite("align_face.jpg", crop_image)