OCR -- 训练数据扩增的方法

结交了一些大神级的人物， 每次交流都有很多收获，感谢一路相伴，感谢带我一起成长。

（1）透视变换

（2）gauss_blur

（3）norm_blur

（4）模糊图像，模拟小图片放大的效果

（5）颜色翻转、滤波等等

具体代码实现如下：

（1）透视变换 （具体原理可查看：仿射变换，透视变换：二维坐标到二维坐标之间的线性变换，可用于landmark人脸矫正）

#!/usr/env/bin python3
from functools import reduce
import numpy as np
import cv2
import math
import random


# http://planning.cs.uiuc.edu/node102.html
def get_rotate_matrix(x, y, z):
    """
    按照 zyx 的顺序旋转，输入角度单位为 degrees, 均为顺时针旋转
    :param x: X-axis
    :param y: Y-axis
    :param z: Z-axis
    :return:
    """
    x = math.radians(x)
    y = math.radians(y)
    z = math.radians(z)

    c, s = math.cos(y), math.sin(y)
    M_y = np.matrix([[c, 0., s, 0.],
                     [0., 1., 0., 0.],
                     [-s, 0., c, 0.],
                     [0., 0., 0., 1.]])

    c, s = math.cos(x), math.sin(x)
    M_x = np.matrix([[1., 0., 0., 0.],
                     [0., c, -s, 0.],
                     [0., s, c, 0.],
                     [0., 0., 0., 1.]])

    c, s = math.cos(z), math.sin(z)
    M_z = np.matrix([[c, -s, 0., 0.],
                     [s, c, 0., 0.],
                     [0., 0., 1., 0.],
                     [0., 0., 0., 1.]])

    return M_x * M_y * M_z


def cliped_rand_norm(mu=0, sigma3=1):
    """
    :param mu: 均值
    :param sigma3: 3 倍标准差， 99% 的数据落在 (mu-3*sigma, mu+3*sigma)
    :return:
    """
    # 标准差
    sigma = sigma3 / 3
    dst = sigma * np.random.randn() + mu
    dst = np.clip(dst, 0 - sigma3, sigma3)
    return dst


def warpPerspective(src, M33, sl, gpu):
    if gpu:
        from libs.gpu.GpuWrapper import cudaWarpPerspectiveWrapper
        dst = cudaWarpPerspectiveWrapper(src.astype(np.uint8), M33, (sl, sl), cv2.INTER_CUBIC)
    else:
        dst = cv2.warpPerspective(src, M33, (sl, sl), flags=cv2.INTER_CUBIC)
    return dst


# https://stackoverflow.com/questions/17087446/how-to-calculate-perspective-transform-for-opencv-from-rotation-angles
# https://nbviewer.jupyter.org/github/manisoftwartist/perspectiveproj/blob/master/perspective.ipynb
# http://planning.cs.uiuc.edu/node102.html
class PerspectiveTransform(object):
    def __init__(self, x, y, z, scale, fovy):
        self.x = x
        self.y = y
        self.z = z
        self.scale = scale
        self.fovy = fovy

    def transform_image(self, src, gpu=False):
        if len(src.shape) > 2:
            H, W, C = src.shape
        else:
            H, W = src.shape

        M33, sl, _, ptsOut = self.get_warp_matrix(W, H, self.x, self.y, self.z, self.scale, self.fovy)
        sl = int(sl)

        dst = warpPerspective(src, M33, sl, gpu)

        return dst, M33, ptsOut

    def transform_pnts(self, pnts, M33):
        """
        :param pnts: 2D pnts, left-top, right-top, right-bottom, left-bottom
        :param M33: output from transform_image()
        :return: 2D pnts apply perspective transform
        """
        pnts = np.asarray(pnts, dtype=np.float32)
        pnts = np.array([pnts])
        dst_pnts = cv2.perspectiveTransform(pnts, M33)[0]

        return dst_pnts

    def get_warped_pnts(self, ptsIn, ptsOut, W, H, sidelength):
        ptsIn2D = ptsIn[0, :]
        ptsOut2D = ptsOut[0, :]
        ptsOut2Dlist = []
        ptsIn2Dlist = []

        for i in range(0, 4):
            ptsOut2Dlist.append([ptsOut2D[i, 0], ptsOut2D[i, 1]])
            ptsIn2Dlist.append([ptsIn2D[i, 0], ptsIn2D[i, 1]])

        pin = np.array(ptsIn2Dlist) + [W / 2., H / 2.]
        pout = (np.array(ptsOut2Dlist) + [1., 1.]) * (0.5 * sidelength)
        pin = pin.astype(np.float32)
        pout = pout.astype(np.float32)

        return pin, pout

    def get_warp_matrix(self, W, H, x, y, z, scale, fV):
        fVhalf = np.deg2rad(fV / 2.)
        d = np.sqrt(W * W + H * H)
        sideLength = scale * d / np.cos(fVhalf)
        h = d / (2.0 * np.sin(fVhalf))
        n = h - (d / 2.0)
        f = h + (d / 2.0)

        # Translation along Z-axis by -h
        T = np.eye(4, 4)
        T[2, 3] = -h

        # Rotation matrices around x,y,z
        R = get_rotate_matrix(x, y, z)

        # Projection Matrix
        P = np.eye(4, 4)
        P[0, 0] = 1.0 / np.tan(fVhalf)
        P[1, 1] = P[0, 0]
        P[2, 2] = -(f + n) / (f - n)
        P[2, 3] = -(2.0 * f * n) / (f - n)
        P[3, 2] = -1.0

        # pythonic matrix multiplication
        M44 = reduce(lambda x, y: np.matmul(x, y), [P, T, R])

        # shape should be 1,4,3 for ptsIn and ptsOut since perspectiveTransform() expects data in this way.
        # In C++, this can be achieved by Mat ptsIn(1,4,CV_64FC3);
        ptsIn = np.array([[
            [-W / 2., H / 2., 0.],
            [W / 2., H / 2., 0.],
            [W / 2., -H / 2., 0.],
            [-W / 2., -H / 2., 0.]
        ]])
        ptsOut = cv2.perspectiveTransform(ptsIn, M44)

        ptsInPt2f, ptsOutPt2f = self.get_warped_pnts(ptsIn, ptsOut, W, H, sideLength)

        # check float32 otherwise OpenCV throws an error
        assert (ptsInPt2f.dtype == np.float32)
        assert (ptsOutPt2f.dtype == np.float32)
        M33 = cv2.getPerspectiveTransform(ptsInPt2f, ptsOutPt2f).astype(np.float32)

        return M33, sideLength, ptsInPt2f, ptsOutPt2f

def apply_perspective_transform(img, text_box_pnts, max_x, max_y, max_z, gpu=False):
    """
    Apply perspective transform on image
    :param img: origin numpy image
    :param text_box_pnts: four corner points of text
    :param x: max rotate angle around X-axis
    :param y: max rotate angle around Y-axis
    :param z: max rotate angle around Z-axis
    :return:
        dst_img:
        dst_img_pnts: points of whole word image after apply perspective transform
        dst_text_pnts: points of text after apply perspective transform
    """

    x = math_utils.cliped_rand_norm(0, max_x)
    y = math_utils.cliped_rand_norm(0, max_y)
    z = math_utils.cliped_rand_norm(0, max_z)

    # print("x: %f, y: %f, z: %f" % (x, y, z))

    transformer = PerspectiveTransform(x, y, z, scale=1.0, fovy=50)

    dst_img, M33, dst_img_pnts = transformer.transform_image(img, gpu)
    dst_text_pnts = transformer.transform_pnts(text_box_pnts, M33)

    return dst_img, dst_img_pnts, dst_text_pnts

（2）gauss_blur

def apply_gauss_blur(img, ks=None):
    if ks is None:
        ks = [7, 9, 11, 13]
    ksize = random.choice(ks)

    sigmas = [0, 1, 2, 3, 4, 5, 6, 7]
    sigma = 0
    if ksize >= 3:
        sigma = random.choice(sigmas)
    img = cv2.GaussianBlur(img, (ksize, ksize), sigma)
    return img

（3）norm_blur

def apply_norm_blur(img, ks=None):
    # kernel == 1, the output image will be the same
    if ks is None:
        ks = [2, 3]
    kernel = random.choice(ks)
    img = cv2.blur(img, (kernel, kernel))
    return img

（4）模糊图像，模拟小图片放大的效果

def apply_prydown(img):
    """
    模糊图像，模拟小图片放大的效果
    """
    scale = random.uniform(1, self.cfg.prydown.max_scale)
    height = img.shape[0]
    width = img.shape[1]

    out = cv2.resize(img, (int(width / scale), int(height / scale)), interpolation=cv2.INTER_AREA)
    return cv2.resize(out, (width, height), interpolation=cv2.INTER_AREA)

（5）颜色翻转、滤波等等

def reverse_img(word_img):
    offset = np.random.randint(-10, 10)
    return 255 + offset - word_img


def apply_emboss(word_img):
    emboss_kernal = np.array([
        [-2, -1, 0],
        [-1, 1, 1],
        [0, 1, 2]
    ])
    return cv2.filter2D(word_img, -1, emboss_kernal)

def apply_sharp(word_img):
    sharp_kernel = np.array([
        [-1, -1, -1],
        [-1, 9, -1],
        [-1, -1, -1]
    ])
    return cv2.filter2D(word_img, -1, sharp_kernel)