OpenCV——7

25像素重映射

把像素点P(x,y)重新映射到一个新的位置P’(x’, y’)

像素重映射函数

cv.remap(src, map1, map2, interpolation[, dst[, borderMode[, borderValue]]] ) ->dst

​ •src表示图像

​ •map1表示x,y方向映射规则，或者x方向映射

​ •Map2如果map1表示x,y映射时为空，否则表示y

​ •表示映射时候的像素插值方法 支持：INTER_NEAREST 、NTER_LINEAR 、NTER_CUBIC

#25像素重映射
def remap_demo():
    image = cv.imread("123.jpg")
    cv.namedWindow("remap-demo", cv.WINDOW_AUTOSIZE)
    cv.createTrackbar("remap-type", "remap-demo", 0, 3, trackbar_callback)
    h, w, c = image.shape
    cv.imshow("input", image)
    map_x = np.zeros((h, w), dtype=np.float32)
    map_y = np.zeros((h, w), dtype=np.float32)

    while True:
        pos = cv.getTrackbarPos("remap-type", "remap-demo")
        if pos == 0:  # 倒立
            for i in range(map_x.shape[0]):
                map_x[i, :] = [x for x in range(map_x.shape[1])]
            for j in range(map_y.shape[1]):
                map_y[:, j] = [map_y.shape[0] - y for y in range(map_y.shape[0])]
        elif pos == 1:  # 镜像
            for i in range(map_x.shape[0]):
                map_x[i, :] = [map_x.shape[1] - x for x in range(map_x.shape[1])]
            for j in range(map_y.shape[1]):
                map_y[:, j] = [y for y in range(map_y.shape[0])]
        elif pos == 2:  # 对象线对称
            for i in range(map_x.shape[0]):
                map_x[i, :] = [map_x.shape[1] - x for x in range(map_x.shape[1])]
            for j in range(map_y.shape[1]):
                map_y[:, j] = [map_y.shape[0] - y for y in range(map_y.shape[0])]
        elif pos == 3:  # 放大两倍
            for i in range(map_x.shape[0]):
                map_x[i, :] = [int(x/2) for x in range(map_x.shape[1])]
            for j in range(map_y.shape[1]):
                map_y[:, j] = [int(y/2) for y in range(map_y.shape[0])]

        dst = cv.remap(image, map_x, map_y, cv.INTER_LINEAR)
        cv.imshow("remap-demo", dst)
        c = cv.waitKey(100)
        if c == 27:
            break
    cv.destroyAllWindows()

26图像二值化

图像二值化定义

•只有两个像素值0、1(0表示黑色，1-255表示白色)，黑色表示背景，白色表示对象（规则）

图像二值化方法

•cv.mean,计算灰度图像均值m

•inRange方法分割

二值化函数

cv.threshold(src, thresh,maxval, type[,dst]) ->retval, dst

​ src表示输入图像

​ thresh表示阈值

​ maxval表示最大值

​ type表示 二值化THRESH_BINARY 或者 二值化反THRESH_BINARY_INV

​ retval表示返回阈值，dst返回的二值图像

# 26图像二值化
def binary_demo():
    image = cv.imread("123.jpg")
    gray = cv.cvtColor(image, cv.COLOR_BGR2GRAY)
    cv.imshow("gray", gray)

    # 手动阈值，二值化
    ret, binary = cv.threshold(gray, 127, 255, cv.THRESH_BINARY)
    cv.imshow("binary",binary)
    cv.waitKey(0)

    # 求均值，二值化
    m = cv.mean(gray)[0]
    ret, binary = cv.threshold(gray, m, 255, cv.THRESH_BINARY)
    cv.imshow("binary", binary)
    cv.waitKey(0)

    cv.destroyAllWindows()

27全局与自适应二值化

全局二值化

（1）大津法（针对两峰）：0~5六个灰度级别，根据直方图分布，以每个灰度等级分割直方图分布为两个部分，分别求取均值跟方差，如图示，最小方法差和对应的灰度值为，分割阈值.

（2）三角法（针对单峰）

α和β角都为45°，最长的d对应的点偏移0.2即为阈值点。

两种方法都是基于直方图分布

全局二值化函数

cv.threshold(src, thresh,maxval, type[,dst]) ->retval,dst

​ •type表示二值化

​ •THRESH_BINARY | THRESH_OTSU 全局自动阈值＋二值化（大津）

​ •THRESH_BINARY | THRESH_TRIANGLE 全局自动阈值＋二值化（三角）

​ •THRESH_BINARY_INV | THRESH_OTSU

​ 表示不同的全局二值化方法

自适应二值化

•模糊图像 D（可以为均值模糊/高斯模糊）

•原图S + 加上偏置常量C

•T = S –D > -C ? 255 : 0

自适应二值化函数

cv.adaptiveThreshold(src, maxValue, adaptiveMethod, thresholdType, blockSize, C[,dst] ) ->dst

​ •cv.ADAPTIVE_THRESH_MEAN_C 均值

​ cv.ADAPTIVE_THRESH_GAUSSIAN_C 高斯

​ •blockSize必须为奇数

​ •C表示要减去的权重，可以是正数，负数，0

# 27全局与自适应二值化
def binarier_demo():
    image = cv.imread("123.jpg")
    gray = cv.cvtColor(image, cv.COLOR_BGR2GRAY)
    cv.imshow("gray", gray)

    # 手动阈值，大津法
    ret, binary = cv.threshold(gray, 0, 255, cv.THRESH_BINARY | cv.THRESH_OTSU)
    cv.imshow("binary1", binary)
    cv.waitKey(0)

    # 手动阈值，三角法
    ret, binary = cv.threshold(gray, 0, 255, cv.THRESH_BINARY | cv.THRESH_TRIANGLE)
    cv.imshow("binary2", binary)
    cv.waitKey(0)

    # 自适应法
    binary = cv.adaptiveThreshold(gray, 255, cv.ADAPTIVE_THRESH_GAUSSIAN_C, cv2.THRESH_BINARY, 25, 10)
    cv.imshow("binary3", binary)
    cv.waitKey(0)

    cv.destroyAllWindows()

28实时人脸检测

OpenCV4 DNN模块

•来自另外一个开源项目tiny dnn

•OpenCV3.3正式发布

•最新版本OpenCV4.5.5

•支持后台硬件加速机制 CPU/GPU等

•支持多种任务(分类、检测、分割、风格迁移、场景文字检测等)

•只支持推理（模型部署），不支持模型训练

•支持主流的深度学习框架生成模型，OpenCV加载

•推荐使用pytorch/tensorflow

OpenCV人脸检测支持演化

•OpenCV3.3之前基于HAAR/LBP级联检测

•OpenCV3.3开始支持深度学习人脸检测

•支持人脸检测模型caffe/tensorflow

•OpenCV4.5.4 支持人脸检测+landmark

•模型下载地址：

•https://gitee.com/opencv_ai/opencv_tutorial_data

DNN相关函数

•读取模型：readNetFromTensorflow

•转换为blob对象：blobFromImage

•设置输入：setInput

•推理预测：forward

人脸检测显示

•模型输入:1x3x300x300

•模型输出:1xN（张人脸）x7（个数据）

​ 人脸检测框坐标（左上右下） – 后面四个值

​ 预测置信度（score） – 第三个值

​ class_id（类别） – 第一个值

# 人脸检测
#下载下面两个模型到项目地址（https://gitee.com/opencv_ai/opencv_tutorial_data）
model_bin = "opencv_face_detector_uint8.pb"
config_text = "opencv_face_detector.pbtxt";

# 视频人脸检测
def video_detection():
    font = cv.FONT_HERSHEY_SIMPLEX
    font_scale = 0.5
    thickness = 1

    #load tensorflow model
    net = cv.dnn.readNetFromTensorflow(model_bin, config=config_text)
    capture = cv.VideoCapture(0) #获取摄像头图像
    # 人脸检测
    while True:
        e1 = cv.getTickCount()
        ret, frame = capture.read()
        frame = cv.flip(frame, 1)
        if ret is not True:
            break
        h, w, c = frame.shape
        blobImage = cv.dnn.blobFromImage(frame, 1.0, (300, 300), (104.0, 177.0, 123.0), False, False);
        net.setInput(blobImage)
        cvOut = net.forward()
        print(cvOut.shape)

        # Put efficiency information
        t, _ = net.getPerfProfile()
        label = 'Inference time: %.2f ms' % (t * 1000.0 / cv.getTickFrequency())

        # 绘制检测矩阵
        for detection in cvOut[0, 0, :, :]:
            score = float(detection[2])
            objIndex = int(detection[1])
            if score > 0.5:
                left = detection[3] * w
                top = detection[4] * h
                right = detection[5] * w
                bottom = detection[6] * h

                # 绘制矩形框
                cv.rectangle(frame, (int(left), int(top)), (int(right), int(bottom)), (255, 0, 0),thickness=2)

                # 绘制类别跟得分
                label_txt = "score: %.2f"%score
                (fw, uph), dh = cv.getTextSize(label_txt, font, font_scale, thickness)
                cv.rectangle(frame, (int(left), int(top) - uph - dh), (int(left) + fw, int(top)), (255, 255, 255), -1, 8)
                cv.putText(frame, label_txt, (int(left), int(top) - dh), font, font_scale, (255, 0, 255), thickness)

        e2 = cv.getTickCount()
        fps = cv.getTickFrequency() / (e2 - e1)
        cv.putText(frame, label + (" FPS: %.2f"%fps), (10, 50), cv.FONT_HERSHEY_SIMPLEX, 1.0, (0, 0, 255), 2)
        cv.imshow('face-detection-demo', frame)
        c = cv.waitKey(1)
        if c == 27:
            break
    cv.destroyAllWindows()

# 图片人脸检测
def image_detection():
    font = cv.FONT_HERSHEY_SIMPLEX
    font_scale = 0.5
    thickness = 1

    #load tensorflow model
    net = cv.dnn.readNetFromTensorflow(model_bin, config=config_text)
    capture = cv.VideoCapture(0)
    # 人脸检测
    e1 = cv.getTickCount()
    frame = cv.imread("face.png")
    h, w, c = frame.shape
    blobImage = cv.dnn.blobFromImage(frame, 1.0, (300, 300), (104.0, 177.0, 123.0), False, False);
    net.setInput(blobImage)
    cvOut = net.forward()
    print(cvOut.shape)

    # Put efficiency information
    t, _ = net.getPerfProfile()
    label = 'Inference time: %.2f ms' % (t * 1000.0 / cv.getTickFrequency())

    # 绘制检测矩阵
    for detection in cvOut[0, 0, :, :]:
        score = float(detection[2])
        objIndex = int(detection[1])
        if score > 0.5:
            left = detection[3] * w
            top = detection[4] * h
            right = detection[5] * w
            bottom = detection[6] * h

            # 绘制矩形框
            cv.rectangle(frame, (int(left), int(top)), (int(right), int(bottom)), (255, 0, 0), thickness=2)

            # 绘制类别跟得分
            label_txt = "score: %.2f" % score
            (fw, uph), dh = cv.getTextSize(label_txt, font, font_scale, thickness)
            cv.rectangle(frame, (int(left), int(top) - uph - dh), (int(left) + fw, int(top)), (255, 255, 255), -1, 8)
            cv.putText(frame, label_txt, (int(left), int(top) - dh), font, font_scale, (255, 0, 255), thickness)

    e2 = cv.getTickCount()
    fps = cv.getTickFrequency() / (e2 - e1)
    cv.putText(frame, label + (" FPS: %.2f" % fps), (10, 50), cv.FONT_HERSHEY_SIMPLEX, 1.0, (0, 0, 255), 2)
    cv.imshow('face-detection-demo', frame)
    c = cv.waitKey(0)
    cv.destroyAllWindows()

​