效果展示：
原图：（图片来自 b站up 借我300去洗牙）

dlib实现的特征点检测

瞳孔检测结果

完整代码：

# encoding:utf-8

import dlib
import numpy as np
import cv2

def rect_to_bb(rect): # 获得人脸矩形的坐标信息
    x = rect.left()
    y = rect.top()
    w = rect.right() - x
    h = rect.bottom() - y
    return (x, y, w, h)

def shape_to_np(shape, dtype="int"): # 将包含68个特征的的shape转换为numpy array格式
    coords = np.zeros((68, 2), dtype=dtype)
    for i in range(0, 68):
        coords[i] = (shape.part(i).x, shape.part(i).y)
    return coords


def resize(image, width=1200):  # 将待检测的image进行resize
    r = width * 1.0 / image.shape[1]
    dim = (width, int(image.shape[0] * r))
    resized = cv2.resize(image, dim, interpolation=cv2.INTER_AREA)
    return resized

def feature():
    image_file = r"F:\project_python\facenet\2.PNG"
    detector = dlib.get_frontal_face_detector()
    predictor = dlib.shape_predictor(r"F:\project_python\facenet\shape_predictor_68_face_landmarks.dat")
    image = cv2.imread(image_file)
    image = resize(image, width=1200)
    gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
    rects = detector(gray, 1)
    shapes = []
    for (i, rect) in enumerate(rects):
        shape = predictor(gray, rect)
        shape = shape_to_np(shape)
        shapes.append(shape)
        (x, y, w, h) = rect_to_bb(rect)
        cv2.rectangle(image, (x, y), (x + w, y + h), (0, 255, 0), 2)
        cv2.putText(image, "Face: {}".format(i + 1), (x - 10, y - 10), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 255, 0), 2)

    count1 = 0
    count2 = 0# encoding:utf-8

import dlib
import numpy as np
import cv2

def rect_to_bb(rect): # 获得人脸矩形的坐标信息
    x = rect.left()
    y = rect.top()
    w = rect.right() - x
    h = rect.bottom() - y
    return (x, y, w, h)

def shape_to_np(shape, dtype="int"): # 将包含68个特征的的shape转换为numpy array格式
    coords = np.zeros((68, 2), dtype=dtype)
    for i in range(0, 68):
        coords[i] = (shape.part(i).x, shape.part(i).y)
    return coords


def resize(image, width=1200):  # 将待检测的image进行resize
    r = width * 1.0 / image.shape[1]
    dim = (width, int(image.shape[0] * r))
    resized = cv2.resize(image, dim, interpolation=cv2.INTER_AREA)
    return resized

def feature():
    image_file = r"F:\project_python\facenet\2.PNG"
    detector = dlib.get_frontal_face_detector()
    predictor = dlib.shape_predictor(r"F:\project_python\facenet\shape_predictor_68_face_landmarks.dat")
    image = cv2.imread(image_file)
    image = resize(image, width=1200)
    gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
    rects = detector(gray, 1)
    shapes = []
    for (i, rect) in enumerate(rects):
        shape = predictor(gray, rect)
        shape = shape_to_np(shape)
        shapes.append(shape)
        (x, y, w, h) = rect_to_bb(rect)
        cv2.rectangle(image, (x, y), (x + w, y + h), (0, 255, 0), 2)
        cv2.putText(image, "Face: {}".format(i + 1), (x - 10, y - 10), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 255, 0), 2)

    count1 = 0
    count2 = 0
    image1=image.copy()
    
    eyes=[]
    eyes2=[]
    
    for shape in shapes:
        left_eye=[]
        right_eye=[]
        left_eye2=[]
        right_eye2=[]
        for (x, y) in shape:
            
            cv2.circle(image1, (x, y), 2, (0, 0, 255), -1)
            cv2.putText(image1, str(count1)+"-"+str(count2), (x-3, y-3), cv2.FONT_HERSHEY_COMPLEX, 0.4, (100, 200, 200), 1)

            if count2>=36 and count2<=41:
                left_eye.append([x,y])
                left_eye2.append((x,y))
            elif count2>=42 and count2<=47:
                right_eye.append([x,y])
                right_eye2.append((x,y))
            
            count2+=1
        count1+=1    
        eyes.append([left_eye,right_eye])
        eyes2.append([left_eye2,right_eye2])
    cv2.imshow("Output", image1)
    cv2.waitKey(0)
    
    
    image2=image.copy()
    
    for i in range(len(eyes)):
        e=eyes[i]
        for j in range(len(e)):
            points=e[j]
            
            
            # 六边形的顶点坐标
            pts = np.array(points, np.int32)
            pts = pts.reshape((-1, 1, 2))

            # 创建一个空白图像作为掩模
            mask = np.zeros_like(image)

            # 在掩模上绘制填充了白色的六边形
            cv2.fillPoly(mask, [pts], (255, 255, 255))

            # 创建一个白色背景图像
            white_background = np.ones_like(image) * 255

            # 在白色背景上绘制填充了原始图像的六边形
            cv2.fillPoly(white_background, [pts], (0, 0, 0))

            # 将原始图像和白色背景图像按位取反
            inverse_mask = cv2.bitwise_not(mask)

            # 将原始图像中六边形内的部分与白色背景中六边形外的部分相结合
            result = cv2.bitwise_and(image, mask) + cv2.bitwise_and(white_background, inverse_mask)
            
            # 显示最终结果
            cv2.imshow('Result', result)
            cv2.waitKey(0)
            
            
            # 找到最小长方形的左上角和右下角坐标
            min_x = min(point[0] for point in points)
            max_x = max(point[0] for point in points)
            min_y = min(point[1] for point in points)
            max_y = max(point[1] for point in points)
            cv2.rectangle(image2, (min_x, min_y), (max_x, max_y), (0, 255, 255), 2)
            
            # 获取最小长方形中的图像部分
            roi = result[min_y:max_y, min_x:max_x]

            # 将图像部分转换为灰度图
            gray_roi = cv2.cvtColor(roi, cv2.COLOR_BGR2GRAY)

            # define a threshold, 128 is the middle of black and white in grey scale
            thresh = 100
            
            # assign blue channel to zeros
            img_binary = cv2.threshold(gray_roi, thresh, 255, cv2.THRESH_BINARY)[1]
            
            cv2.imshow("Output", img_binary)
            cv2.waitKey(0)
            
            #处理眼睛反光形成的白点
            # 获取图像高度
            img_binary1=img_binary.copy()
            height = img_binary1.shape[0]

            # 遍历每一列像素
            for col in range(img_binary1.shape[1]):
                white_run_length = 0
                start_index = 0
                end_index = 0
                for row in range(height):
                    if img_binary1[row, col] == 255:  # 白色像素
                        if white_run_length == 0:
                            start_index = row
                        white_run_length += 1
                        end_index = row
                    else:  # 黑色像素
                        if white_run_length > 0 and white_run_length < height / 2 and (start_index == 0 or img_binary1[start_index - 1, col] == 0) and (end_index == height - 1 or img_binary1[end_index + 1, col] == 0):
                            img_binary1[start_index:end_index + 1, col] = 0  # 将连续的白色像素改为黑色像素
                        white_run_length = 0

            cv2.imshow("Output", img_binary1)
            cv2.waitKey(0)
            
            # 计算每一列的像素值总和
            column_sums = np.sum(img_binary1, axis=0)

            # 计算每一列与相邻的n列的总和
            neighbor_sums = np.convolve(column_sums, np.ones(1), mode='valid')

            # 找到结果最小的n列的列号
            min_indices = np.argpartition(neighbor_sums, 20)[:20]

            # 将列号按数字排序
            sorted_indices = np.sort(min_indices)

            # 找到这7列的最中间的列的列号
            middle_column = sorted_indices[len(sorted_indices) // 2]

            print("每一列与相邻的6列的总和:", neighbor_sums)
            print("结果最小的7列的列号（按数字排序）:", sorted_indices)
            print("这7列的最中间的列的列号:", middle_column)

            
            cv2.line(image2, (min_x+middle_column, min_y), (min_x+middle_column, max_y), (255, 0, 0))
            
            # 找出结果最小的列
            min_column = img_binary[:, middle_column]

            # 找出连续相邻的白色像素点
            white_pixels = np.where(min_column == 255)[0]

            # 找出连续相邻的白色像素点中最长的一段
            longest_start = 0
            longest_end = 0
            max_length = 0

            current_start = 0
            current_length = 0

            for i in range(1, len(white_pixels)):
                if white_pixels[i] == white_pixels[i-1] + 1:
                    current_length += 1
                else:
                    if current_length > max_length:
                        max_length = current_length
                        longest_start = white_pixels[current_start]
                        longest_end = white_pixels[i-1]
                    current_start = i
                    current_length = 0

            # 判断起始位置和结束位置的中点在图像的上半部分还是下半部分
            mid_point = (longest_start + longest_end) // 2
            height = img_binary.shape[0]
            if mid_point < height // 2:
                print("最长的白色像素段起始位置:", longest_start)
                print("最长的白色像素段结束位置:", longest_end)
                print("连续的长度:", max_length)
                print("中点在图像的上半部分")
                cv2.line(image2, (min_x,min_y+ height // 2-max_length), (max_x,min_y+ height // 2-max_length), (255, 0, 0))
            else:
                print("最长的白色像素段起始位置:", longest_start)
                print("最长的白色像素段结束位置:", longest_end)
                print("连续的长度:", max_length)
                print("中点在图像的下半部分")
                cv2.line(image2, (min_x,min_y+ height // 2+max_length), (max_x,min_y+ height // 2+max_length), (255, 0, 0))
  
    
    cv2.imshow("Output", image2)
    cv2.waitKey(0)


if __name__ == "__main__":

    feature()

文件下载：
1.shape_predictor_68_face_landmarks.dat 下载地址

1.官方下载地址(会比较慢) http://dlib.net/files/

2.我的网盘： 链接：https://pan.baidu.com/s/1ORhqLS1bkHyyYfzbmftNpg 提取码：va12

3.我的资源(免费下载)：https://download.csdn.net/download/qq_51985653/15122811

原文链接：https://blog.csdn.net/qq_51985653/article/details/113748025