Use existing labeled text information to make fake data

from PIL import Image, ImageDraw, ImageFont, ImageFilter
import random
import glob
import numpy as np
import os
import cv2
from nespaper_semantics import seg_str

'''
1. 从文字库随机选择10个字符
2. 生成图片
3. 随机使用函数
'''
# 从字库中随机选择n个字符
def sto_choice_from_info_str(info_str,quantity):
    start = random.randint(0, len(info_str)-(quantity+1))
    end = start + quantity
    random_word = info_str[start:end]
    return random_word

def random_word_color():
    font_color_choice = [[10, 10, 10], [5, 5, 5], [0, 0, 0]]
    font_color = random.choice(font_color_choice)
    noise = np.array([random.randint(0,2), random.randint(0,2), random.randint(0,2)])
    font_color = (np.array(font_color) + noise).tolist()
    return tuple(font_color)
# 生成一张背景图片
def create_an_image(bground_path, width, height):
    bground_list = os.listdir(bground_path)
    bground_choice = random.choice(bground_list)
    bground = cv2.imread(os.path.join(bground_path, bground_choice))
    x, y = random.randint(0, bground.shape[1]-width), random.randint(0, bground.shape[0]-height)
    bground = bground[y:y+height, x:x+width, :]
    return bground

# 随机选取文字贴合起始的坐标, 根据背景的尺寸和字体的大小选择
def random_x_y(bground_size, font_size):
    height, width, _ = bground_size
    print('bground_size:', bground_size)
    print('font_size:', font_size)
    # 为防止文字溢出图片，x，y要预留宽高
    # x = random.randint(0, width - font_size * 10)
    # y = random.randint(0, int((height-font_size)/4))
    """====notice notice===="""
    #10个字要减140 9个字要减100 8个字要减80 7个字要减40 6个字要减20 5个字以下不用减
    x = random.randint(3, int((width - font_size) / 2))
    y = random.randint(10, height - font_size * 7)

    return x, y

def random_font_size():
    font_size = random.randint(22,25)

    return font_size

def random_font(font_path):
    font_list = os.listdir(font_path)
    random_font = random.choice(font_list)

    return os.path.join(font_path, random_font)
def add_white_noise(image):
    rows, cols, dims = image.shape

    random_index=random.randint(10,100)
    for i in range(random_index):
        x = np.random.randint(2, rows)
        y = np.random.randint(2, cols)

        if random.getrandbits(1):
            image[x-1:x+1, y-1:y+1, :] = 180
        else:
            image[x, y, :] = 180
    return image

def main(infostr, save_path, num,words_length,width,height,back_path,font_path):
    # 随机选取5个字符
    random_word_ori = sto_choice_from_info_str(infostr, words_length)
    print('random_word_ori:', random_word_ori)
    random_word = ''.join([i+'\n' for i in random_word_ori if i not in [':','(',')','｢','｣']])
    print('random_word:\n', random_word)
    # # 生成一张背景图片，已经剪裁好，宽高为280*32
    raw_image = create_an_image(back_path, width, height)
    # 随机选取字体大小
    font_size = random_font_size()
    # 随机选取字体
    font_name = random_font(font_path)
    print('font_name:', font_name)
    # 随机选取字体颜色
    font_color = random_word_color()

    # 随机选取文字贴合的起始坐标 x,y
    draw_x, draw_y = random_x_y(raw_image.shape, font_size)

    # 将文本贴到背景图片
    raw_image = Image.fromarray(raw_image[..., ::-1])
    font = ImageFont.truetype(font_name, font_size)
    draw = ImageDraw.Draw(raw_image)
    draw.text((draw_x, draw_y), random_word, fill=font_color, font=font)
    raw_image = raw_image.rotate(0.3)
    image = add_white_noise(np.array(raw_image))
    # cv2.imwrite('image.jpg', image)

    # 保存文本信息和对应图片名称
    img_name = save_path+'/'+'fake_'+str(words_length)+'_'+str(num)+'.jpg'
    cv2.imwrite(img_name, image)
    with open(img_name.replace('.jpg','.txt'), 'w', encoding='utf-8') as file:
        [file.write(val) for val in random_word_ori]
def make_fake_data(total):
    #背景图片地址
    back_path = './background'
    #字体地址
    font_path = './font'
    #语料库 #从make_fake_word_library函数得到得到的语料库
    info_str = seg_str
    print('len(info_str):', len(info_str))
    # 生成的字符个数
    words_length = 3
    output_path = 'data_set/fake_one_'+str(words_length)
    if not os.path.exists(output_path):
        os.mkdir(output_path)

    if words_length < 6:
        width=32
        height=200
    elif 6<=words_length<7:
        width = 32
        height = 220

    elif 7<=words_length<8:
        width = 32
        height = 240
    elif 8 <= words_length < 9:
        width = 32
        height = 280
    elif 9 <= words_length < 10:
        width = 32
        height = 300
    else:
        width = 32
        height = 340

    for i,num in enumerate(range(0, total)):
        if i<1:
            main(info_str, output_path, num, words_length, width, height,back_path,font_path)
            if num % 1000 == 0:
                print('[%d/%d]' % (num, total))
#根据标注的报纸制作语义库
def make_fake_word_library():
    import pandas as pd
    train_path = './label/train.txt'
    val_path = './label/val.txt'

    train_names = np.array(pd.read_csv(train_path, header=None))
    val_names = np.array(pd.read_csv(val_path, header=None))

    Semantics_list=[]
    for i,train_name in enumerate(train_names):
            words = train_name[0].split(' ')[-1]
            Semantics_list.append(words)

    for i,val_name in enumerate(val_names):
            words = val_name[0].split(' ')[-1]
            Semantics_list.append(words)
    print(len(Semantics_list))
    print(Semantics_list[:2])
    Semantics_str=''.join(Semantics_list)
    print(len(Semantics_str))
    print(Semantics_str)
if __name__ == '__main__':
    #用报纸制作的假数据
    make_fake_data(total=1000)
    # make_fake_word_library()

background:

　

Font: https://download.csdn.net/download/fanzonghao/11866723

 

Output: