EasyCV/easycv/datasets/ocr/pipelines/rec_transform.py

698 lines
22 KiB
Python

# Modified from https://github.com/PaddlePaddle/PaddleOCR/tree/release/2.6/ppocr/data/imaug
import math
import random
import sys
import cv2
import imgaug
import imgaug.augmenters as iaa
import numpy as np
import pyclipper
from shapely.geometry import Polygon
from easycv.datasets.registry import PIPELINES
@PIPELINES.register_module()
class RecConAug(object):
"""concat multiple texts together for text recognition training
"""
def __init__(self,
prob=0.5,
image_shape=(32, 320, 3),
max_text_length=25,
**kwargs):
"""
Args:
prob (float, optional): the probability whether do data augmentation. Defaults to 0.5.
image_shape (tuple, optional): the output image shape. Defaults to (32, 320, 3).
max_text_length (int, optional): the max length of text label. Defaults to 25.
"""
self.prob = prob
self.max_text_length = max_text_length
self.image_shape = image_shape
self.max_wh_ratio = self.image_shape[1] / self.image_shape[0]
def merge_ext_data(self, data, ext_data):
ori_w = round(data['img'].shape[1] / data['img'].shape[0] *
self.image_shape[0])
ext_w = round(ext_data['img'].shape[1] / ext_data['img'].shape[0] *
self.image_shape[0])
data['img'] = cv2.resize(data['img'], (ori_w, self.image_shape[0]))
ext_data['img'] = cv2.resize(ext_data['img'],
(ext_w, self.image_shape[0]))
data['img'] = np.concatenate([data['img'], ext_data['img']], axis=1)
data['label'] += ext_data['label']
return data
def __call__(self, data):
rnd_num = random.random()
if rnd_num > self.prob:
return data
for idx, ext_data in enumerate(data['ext_data']):
if len(data['label']) + len(
ext_data['label']) > self.max_text_length:
break
concat_ratio = data['img'].shape[1] / data['img'].shape[
0] + ext_data['img'].shape[1] / ext_data['img'].shape[0]
if concat_ratio > self.max_wh_ratio:
break
data = self.merge_ext_data(data, ext_data)
data.pop('ext_data')
return data
@PIPELINES.register_module()
class RecAug(object):
"""data augmentation function for ocr recognition
"""
def __init__(self, use_tia=True, aug_prob=0.4, **kwargs):
"""
Args:
use_tia (bool, optional): whether make tia augmentation. Defaults to True.
aug_prob (float, optional): the probability were do data augmentation. Defaults to 0.4.
"""
self.use_tia = use_tia
self.aug_prob = aug_prob
def __call__(self, data):
img = data['img']
img = warp(img, 10, self.use_tia, self.aug_prob)
data['img'] = img
return data
def flag():
"""
flag
"""
return 1 if random.random() > 0.5000001 else -1
def cvtColor(img):
"""
cvtColor
"""
hsv = cv2.cvtColor(img, cv2.COLOR_BGR2HSV)
delta = 0.001 * random.random() * flag()
hsv[:, :, 2] = hsv[:, :, 2] * (1 + delta)
new_img = cv2.cvtColor(hsv, cv2.COLOR_HSV2BGR)
return new_img
def blur(img):
"""
blur
"""
h, w, _ = img.shape
if h > 10 and w > 10:
return cv2.GaussianBlur(img, (5, 5), 1)
else:
return img
def jitter(img):
"""
jitter
"""
w, h, _ = img.shape
if h > 10 and w > 10:
thres = min(w, h)
s = int(random.random() * thres * 0.01)
src_img = img.copy()
for i in range(s):
img[i:, i:, :] = src_img[:w - i, :h - i, :]
return img
else:
return img
def add_gasuss_noise(image, mean=0, var=0.1):
"""
Gasuss noise
"""
noise = np.random.normal(mean, var**0.5, image.shape)
out = image + 0.5 * noise
out = np.clip(out, 0, 255)
out = np.uint8(out)
return out
def get_crop(image):
"""
random crop
"""
h, w, _ = image.shape
top_min = 1
top_max = 8
top_crop = int(random.randint(top_min, top_max))
top_crop = min(top_crop, h - 1)
crop_img = image.copy()
ratio = random.randint(0, 1)
if ratio:
crop_img = crop_img[top_crop:h, :, :]
else:
crop_img = crop_img[0:h - top_crop, :, :]
return crop_img
class Config:
"""
Config
"""
def __init__(self, use_tia):
self.anglex = random.random() * 30
self.angley = random.random() * 15
self.anglez = random.random() * 10
self.fov = 42
self.r = 0
self.shearx = random.random() * 0.3
self.sheary = random.random() * 0.05
self.borderMode = cv2.BORDER_REPLICATE
self.use_tia = use_tia
def make(self, w, h, ang):
"""
make
"""
self.anglex = random.random() * 5 * flag()
self.angley = random.random() * 5 * flag()
self.anglez = -1 * random.random() * int(ang) * flag()
self.fov = 42
self.r = 0
self.shearx = 0
self.sheary = 0
self.borderMode = cv2.BORDER_REPLICATE
self.w = w
self.h = h
self.perspective = self.use_tia
self.stretch = self.use_tia
self.distort = self.use_tia
self.crop = True
self.affine = False
self.reverse = True
self.noise = True
self.jitter = True
self.blur = True
self.color = True
def rad(x):
"""
rad
"""
return x * np.pi / 180
def get_warpR(config):
"""
get_warpR
"""
anglex, angley, anglez, fov, w, h, r = \
config.anglex, config.angley, config.anglez, config.fov, config.w, config.h, config.r
if w > 69 and w < 112:
anglex = anglex * 1.5
z = np.sqrt(w**2 + h**2) / 2 / np.tan(rad(fov / 2))
# Homogeneous coordinate transformation matrix
rx = np.array(
[[1, 0, 0, 0], [0, np.cos(rad(anglex)), -np.sin(rad(anglex)), 0],
[0, -np.sin(rad(anglex)),
np.cos(rad(anglex)), 0], [0, 0, 0, 1]], np.float32)
ry = np.array([[np.cos(rad(angley)), 0,
np.sin(rad(angley)), 0], [0, 1, 0, 0],
[
-np.sin(rad(angley)),
0,
np.cos(rad(angley)),
0,
], [0, 0, 0, 1]], np.float32)
rz = np.array(
[[np.cos(rad(anglez)), np.sin(rad(anglez)), 0, 0],
[-np.sin(rad(anglez)),
np.cos(rad(anglez)), 0, 0], [0, 0, 1, 0], [0, 0, 0, 1]], np.float32)
r = rx.dot(ry).dot(rz)
# generate 4 points
pcenter = np.array([h / 2, w / 2, 0, 0], np.float32)
p1 = np.array([0, 0, 0, 0], np.float32) - pcenter
p2 = np.array([w, 0, 0, 0], np.float32) - pcenter
p3 = np.array([0, h, 0, 0], np.float32) - pcenter
p4 = np.array([w, h, 0, 0], np.float32) - pcenter
dst1 = r.dot(p1)
dst2 = r.dot(p2)
dst3 = r.dot(p3)
dst4 = r.dot(p4)
list_dst = np.array([dst1, dst2, dst3, dst4])
org = np.array([[0, 0], [w, 0], [0, h], [w, h]], np.float32)
dst = np.zeros((4, 2), np.float32)
# Project onto the image plane
dst[:, 0] = list_dst[:, 0] * z / (z - list_dst[:, 2]) + pcenter[0]
dst[:, 1] = list_dst[:, 1] * z / (z - list_dst[:, 2]) + pcenter[1]
warpR = cv2.getPerspectiveTransform(org, dst)
dst1, dst2, dst3, dst4 = dst
r1 = int(min(dst1[1], dst2[1]))
r2 = int(max(dst3[1], dst4[1]))
c1 = int(min(dst1[0], dst3[0]))
c2 = int(max(dst2[0], dst4[0]))
try:
ratio = min(1.0 * h / (r2 - r1), 1.0 * w / (c2 - c1))
dx = -c1
dy = -r1
T1 = np.float32([[1., 0, dx], [0, 1., dy], [0, 0, 1.0 / ratio]])
ret = T1.dot(warpR)
except:
ratio = 1.0
T1 = np.float32([[1., 0, 0], [0, 1., 0], [0, 0, 1.]])
ret = T1
return ret, (-r1, -c1), ratio, dst
def get_warpAffine(config):
"""
get_warpAffine
"""
anglez = config.anglez
rz = np.array(
[[np.cos(rad(anglez)), np.sin(rad(anglez)), 0],
[-np.sin(rad(anglez)), np.cos(rad(anglez)), 0]], np.float32)
return rz
def warp(img, ang, use_tia=True, prob=0.4):
"""
warp
"""
h, w, _ = img.shape
config = Config(use_tia=use_tia)
config.make(w, h, ang)
new_img = img
if config.distort:
img_height, img_width = img.shape[0:2]
if random.random() <= prob and img_height >= 20 and img_width >= 20:
new_img = tia_distort(new_img, random.randint(3, 6))
if config.stretch:
img_height, img_width = img.shape[0:2]
if random.random() <= prob and img_height >= 20 and img_width >= 20:
new_img = tia_stretch(new_img, random.randint(3, 6))
if config.perspective:
if random.random() <= prob:
new_img = tia_perspective(new_img)
if config.crop:
img_height, img_width = img.shape[0:2]
if random.random() <= prob and img_height >= 20 and img_width >= 20:
new_img = get_crop(new_img)
if config.blur:
if random.random() <= prob:
new_img = blur(new_img)
if config.color:
if random.random() <= prob:
new_img = cvtColor(new_img)
if config.jitter:
new_img = jitter(new_img)
if config.noise:
if random.random() <= prob:
new_img = add_gasuss_noise(new_img)
if config.reverse:
if random.random() <= prob:
new_img = 255 - new_img
return new_img
class WarpMLS:
def __init__(self, src, src_pts, dst_pts, dst_w, dst_h, trans_ratio=1.):
self.src = src
self.src_pts = src_pts
self.dst_pts = dst_pts
self.pt_count = len(self.dst_pts)
self.dst_w = dst_w
self.dst_h = dst_h
self.trans_ratio = trans_ratio
self.grid_size = 100
self.rdx = np.zeros((self.dst_h, self.dst_w))
self.rdy = np.zeros((self.dst_h, self.dst_w))
@staticmethod
def __bilinear_interp(x, y, v11, v12, v21, v22):
return (v11 * (1 - y) + v12 * y) * (1 - x) + (v21 *
(1 - y) + v22 * y) * x
def generate(self):
self.calc_delta()
return self.gen_img()
def calc_delta(self):
w = np.zeros(self.pt_count, dtype=np.float32)
if self.pt_count < 2:
return
i = 0
while 1:
if self.dst_w <= i < self.dst_w + self.grid_size - 1:
i = self.dst_w - 1
elif i >= self.dst_w:
break
j = 0
while 1:
if self.dst_h <= j < self.dst_h + self.grid_size - 1:
j = self.dst_h - 1
elif j >= self.dst_h:
break
sw = 0
swp = np.zeros(2, dtype=np.float32)
swq = np.zeros(2, dtype=np.float32)
new_pt = np.zeros(2, dtype=np.float32)
cur_pt = np.array([i, j], dtype=np.float32)
k = 0
for k in range(self.pt_count):
if i == self.dst_pts[k][0] and j == self.dst_pts[k][1]:
break
w[k] = 1. / ((i - self.dst_pts[k][0]) *
(i - self.dst_pts[k][0]) +
(j - self.dst_pts[k][1]) *
(j - self.dst_pts[k][1]))
sw += w[k]
swp = swp + w[k] * np.array(self.dst_pts[k])
swq = swq + w[k] * np.array(self.src_pts[k])
if k == self.pt_count - 1:
pstar = 1 / sw * swp
qstar = 1 / sw * swq
miu_s = 0
for k in range(self.pt_count):
if i == self.dst_pts[k][0] and j == self.dst_pts[k][1]:
continue
pt_i = self.dst_pts[k] - pstar
miu_s += w[k] * np.sum(pt_i * pt_i)
cur_pt -= pstar
cur_pt_j = np.array([-cur_pt[1], cur_pt[0]])
for k in range(self.pt_count):
if i == self.dst_pts[k][0] and j == self.dst_pts[k][1]:
continue
pt_i = self.dst_pts[k] - pstar
pt_j = np.array([-pt_i[1], pt_i[0]])
tmp_pt = np.zeros(2, dtype=np.float32)
tmp_pt[0] = np.sum(
pt_i * cur_pt) * self.src_pts[k][0] - np.sum(
pt_j * cur_pt) * self.src_pts[k][1]
tmp_pt[1] = -np.sum(pt_i * cur_pt_j) * self.src_pts[k][
0] + np.sum(pt_j * cur_pt_j) * self.src_pts[k][1]
tmp_pt *= (w[k] / miu_s)
new_pt += tmp_pt
new_pt += qstar
else:
new_pt = self.src_pts[k]
self.rdx[j, i] = new_pt[0] - i
self.rdy[j, i] = new_pt[1] - j
j += self.grid_size
i += self.grid_size
def gen_img(self):
src_h, src_w = self.src.shape[:2]
dst = np.zeros_like(self.src, dtype=np.float32)
for i in np.arange(0, self.dst_h, self.grid_size):
for j in np.arange(0, self.dst_w, self.grid_size):
ni = i + self.grid_size
nj = j + self.grid_size
w = h = self.grid_size
if ni >= self.dst_h:
ni = self.dst_h - 1
h = ni - i + 1
if nj >= self.dst_w:
nj = self.dst_w - 1
w = nj - j + 1
di = np.reshape(np.arange(h), (-1, 1))
dj = np.reshape(np.arange(w), (1, -1))
delta_x = self.__bilinear_interp(di / h, dj / w,
self.rdx[i, j], self.rdx[i,
nj],
self.rdx[ni, j], self.rdx[ni,
nj])
delta_y = self.__bilinear_interp(di / h, dj / w,
self.rdy[i, j], self.rdy[i,
nj],
self.rdy[ni, j], self.rdy[ni,
nj])
nx = j + dj + delta_x * self.trans_ratio
ny = i + di + delta_y * self.trans_ratio
nx = np.clip(nx, 0, src_w - 1)
ny = np.clip(ny, 0, src_h - 1)
nxi = np.array(np.floor(nx), dtype=np.int32)
nyi = np.array(np.floor(ny), dtype=np.int32)
nxi1 = np.array(np.ceil(nx), dtype=np.int32)
nyi1 = np.array(np.ceil(ny), dtype=np.int32)
if len(self.src.shape) == 3:
x = np.tile(np.expand_dims(ny - nyi, axis=-1), (1, 1, 3))
y = np.tile(np.expand_dims(nx - nxi, axis=-1), (1, 1, 3))
else:
x = ny - nyi
y = nx - nxi
dst[i:i + h,
j:j + w] = self.__bilinear_interp(x, y, self.src[nyi, nxi],
self.src[nyi, nxi1],
self.src[nyi1, nxi],
self.src[nyi1, nxi1])
dst = np.clip(dst, 0, 255)
dst = np.array(dst, dtype=np.uint8)
return dst
def tia_distort(src, segment=4):
img_h, img_w = src.shape[:2]
cut = img_w // segment
thresh = cut // 3
src_pts = list()
dst_pts = list()
src_pts.append([0, 0])
src_pts.append([img_w, 0])
src_pts.append([img_w, img_h])
src_pts.append([0, img_h])
dst_pts.append([np.random.randint(thresh), np.random.randint(thresh)])
dst_pts.append(
[img_w - np.random.randint(thresh),
np.random.randint(thresh)])
dst_pts.append(
[img_w - np.random.randint(thresh), img_h - np.random.randint(thresh)])
dst_pts.append(
[np.random.randint(thresh), img_h - np.random.randint(thresh)])
half_thresh = thresh * 0.5
for cut_idx in np.arange(1, segment, 1):
src_pts.append([cut * cut_idx, 0])
src_pts.append([cut * cut_idx, img_h])
dst_pts.append([
cut * cut_idx + np.random.randint(thresh) - half_thresh,
np.random.randint(thresh) - half_thresh
])
dst_pts.append([
cut * cut_idx + np.random.randint(thresh) - half_thresh,
img_h + np.random.randint(thresh) - half_thresh
])
trans = WarpMLS(src, src_pts, dst_pts, img_w, img_h)
dst = trans.generate()
return dst
def tia_stretch(src, segment=4):
img_h, img_w = src.shape[:2]
cut = img_w // segment
thresh = cut * 4 // 5
src_pts = list()
dst_pts = list()
src_pts.append([0, 0])
src_pts.append([img_w, 0])
src_pts.append([img_w, img_h])
src_pts.append([0, img_h])
dst_pts.append([0, 0])
dst_pts.append([img_w, 0])
dst_pts.append([img_w, img_h])
dst_pts.append([0, img_h])
half_thresh = thresh * 0.5
for cut_idx in np.arange(1, segment, 1):
move = np.random.randint(thresh) - half_thresh
src_pts.append([cut * cut_idx, 0])
src_pts.append([cut * cut_idx, img_h])
dst_pts.append([cut * cut_idx + move, 0])
dst_pts.append([cut * cut_idx + move, img_h])
trans = WarpMLS(src, src_pts, dst_pts, img_w, img_h)
dst = trans.generate()
return dst
def tia_perspective(src):
img_h, img_w = src.shape[:2]
thresh = img_h // 2
src_pts = list()
dst_pts = list()
src_pts.append([0, 0])
src_pts.append([img_w, 0])
src_pts.append([img_w, img_h])
src_pts.append([0, img_h])
dst_pts.append([0, np.random.randint(thresh)])
dst_pts.append([img_w, np.random.randint(thresh)])
dst_pts.append([img_w, img_h - np.random.randint(thresh)])
dst_pts.append([0, img_h - np.random.randint(thresh)])
trans = WarpMLS(src, src_pts, dst_pts, img_w, img_h)
dst = trans.generate()
return dst
@PIPELINES.register_module()
class RecResizeImg(object):
def __init__(
self,
image_shape,
infer_mode=False,
character_dict_path='./easycv/datasets/ocr/dict/ppocr_keys_v1.txt',
padding=True,
**kwargs):
self.image_shape = image_shape
self.infer_mode = infer_mode
self.character_dict_path = character_dict_path
self.padding = padding
def __call__(self, data):
img = data['img']
if self.infer_mode and self.character_dict_path is not None:
norm_img, valid_ratio = resize_norm_img_chinese(
img, self.image_shape)
else:
norm_img, valid_ratio = resize_norm_img(img, self.image_shape,
self.padding)
data['img'] = norm_img
data['valid_ratio'] = valid_ratio
return data
def resize_norm_img(img, image_shape, padding=True):
imgC, imgH, imgW = image_shape
h = img.shape[0]
w = img.shape[1]
if not padding:
resized_image = cv2.resize(
img, (imgW, imgH), interpolation=cv2.INTER_LINEAR)
resized_w = imgW
else:
ratio = w / float(h)
if math.ceil(imgH * ratio) > imgW:
resized_w = imgW
else:
resized_w = int(math.ceil(imgH * ratio))
resized_image = cv2.resize(img, (resized_w, imgH))
resized_image = resized_image.astype('float32')
if image_shape[0] == 1:
resized_image = resized_image / 255
resized_image = resized_image[np.newaxis, :]
else:
# resized_image = resized_image.transpose((2, 0, 1)) / 255
resized_image = resized_image / 255
resized_image -= 0.5
resized_image /= 0.5
# padding_im = np.zeros((imgC, imgH, imgW), dtype=np.float32)
# padding_im[:, :, 0:resized_w] = resized_image
padding_im = np.zeros((imgH, imgW, imgC), dtype=np.float32)
padding_im[:, 0:resized_w, :] = resized_image
valid_ratio = min(1.0, float(resized_w / imgW))
return padding_im, valid_ratio
def resize_norm_img_chinese(img, image_shape):
imgC, imgH, imgW = image_shape
# todo: change to 0 and modified image shape
max_wh_ratio = imgW * 1.0 / imgH
h, w = img.shape[0], img.shape[1]
ratio = w * 1.0 / h
max_wh_ratio = max(max_wh_ratio, ratio)
imgW = int(imgH * max_wh_ratio)
if math.ceil(imgH * ratio) > imgW:
resized_w = imgW
else:
resized_w = int(math.ceil(imgH * ratio))
resized_image = cv2.resize(img, (resized_w, imgH))
resized_image = resized_image.astype('float32')
if image_shape[0] == 1:
resized_image = resized_image / 255
resized_image = resized_image[np.newaxis, :]
else:
# resized_image = resized_image.transpose((2, 0, 1)) / 255
resized_image = resized_image / 255
resized_image -= 0.5
resized_image /= 0.5
# padding_im = np.zeros((imgC, imgH, imgW), dtype=np.float32)
# padding_im[:, :, 0:resized_w] = resized_image
padding_im = np.zeros((imgH, imgW, imgC), dtype=np.float32)
padding_im[:, 0:resized_w, :] = resized_image
valid_ratio = min(1.0, float(resized_w / imgW))
return padding_im, valid_ratio
@PIPELINES.register_module()
class ClsResizeImg(object):
def __init__(self, img_shape, **kwargs):
self.img_shape = img_shape
def __call__(self, data):
img = data['img']
norm_img, _ = resize_norm_img(img, self.img_shape)
data['img'] = norm_img
return data