发票识别小工具

[color=rgba(0, 0, 0, 0.9)]有一个需求就是有几十上百个发票需要把关键信息：发票号、发票日期、发票金额、名称等信息提取出来，汇总到excel里面。我的需求很简单，就是把pdf里面的发票或指定尺寸的发票图片信息提取出来，自动保存到本地。
整体的思路是，通过代码把pdf文件转为标准格式的图片（convert_pdf_to_image方法），然后通过定位关键信息的位置，然后通过ddddocr进行识别，目前准确率有待提升。代码如下：

from PIL import Image as PI
import io
import os
import ddddocr
import pandas as pd
import easyocr
import cv2
import fitz  # PyMuPDF
from PIL import Image
ocr = ddddocr.DdddOcr()
 
#把pdf转为图片
def convert_pdf_to_image(pdf_path, output_image_path):
    # 打开PDF文件
    pdf_document = fitz.open(pdf_path)
 
    # 获取第一页
    first_page = pdf_document[0]
 
    # 设置DPI（每英寸点数）
    dpi = 100.0
 
    # 获取图像
    image = first_page.get_pixmap(matrix=fitz.Matrix(dpi / 72, dpi / 72))
 
    # 转换为PIL图像
    pil_image = Image.frombytes("RGB", [image.width, image.height], image.samples)
 
    # 调整图像大小
    #resized_image = pil_image.resize((image_width, image_height), Image.Resampling.LANCZOS)
 
    # 保存图像
    pil_image.save(output_image_path, "PNG")
 
    # 关闭PDF文件
    pdf_document.close()
 
# 读取发票
def readPic(img_url):
    #img_url = "pic/fp01.jpg"
    with open(img_url, 'rb') as f:
        a = f.read()
    new_img = PI.open(io.BytesIO(a))
    #new_img.show()
    return new_img;
 
#递归遍历文件夹
def traverse_directory(path,file_names):
    for root, dirs, files in os.walk(path):
        for filename in files:
            full_path = os.path.join(root, filename)
            if full_path.endswith('.png') or full_path.endswith('.jpeg') or full_path.endswith('.jpg'):
                file_names.append(full_path)
    return file_names
 
# 提取发票号码
 
def parse_invoice(new_img):
    # 748 *500   门诊收费票据
    # 827 * 552  增值类电子普通发票
    # result = reader.readtext(new_img)
    # print(result)
    w = new_img.width  # 图片的宽
    h = new_img.height
    print('识别图片宽、高为：',w,h)
    new_img = new_img.resize((827, int(new_img.size[1] * 827 / new_img.size[0])))
    w = new_img.width  # 图片的宽
    h = new_img.height
    print('调整后识别图片宽、高为：', w, h)
    tuple_voice_type = (236, 28, 583, 57)
    image_voice_type = new_img.crop(tuple_voice_type)
    #image_voice_type.show()
    voice_type = ocr.classification(image_voice_type)
    print('识别出的voice_type为：' + voice_type)
    if '医疗' in voice_type:
        type = 'menzhen'
    elif '普通' in voice_type:
        type = 'putong'
    else:
        type = 'yiliao'
    invoice = {}
    invoice_type_map = {
        #广东增值税电子普通发票
        'putong':{
            'tuple_voice_no':(642,47,703,65),
            'tuple_voice_owner': (137, 122,187,137),
            'tuple_voice_date': (641,70,739,88), # let [top] right [bottom]
            'tuple_voice_amount': (659, 384, 738,404),
            'tuple_voice_amount_cn': (225, 382, 390,405)
        },
        # 广东省医疗门诊收费票据
        'menzhen': {
            'tuple_voice_no': (584, 88, 650, 102),
            'tuple_voice_owner': (92, 120,133,135),
            'tuple_voice_date': (584,121,650,133), # let [top] right [bottom]
            'tuple_voice_amount': (495, 335, 559,351),
            'tuple_voice_amount_cn': (145, 335, 299,351)
        }
    }
    print(invoice_type_map)
    image_voice_no = new_img.crop(invoice_type_map[type]['tuple_voice_no'])
    #image_voice_no.show()
    voice_no = ocr.classification(image_voice_no)
    print('识别出的voice_no为：' + voice_no)
 
    image_voice_owner = new_img.crop(invoice_type_map[type]['tuple_voice_owner'])
    #image_voice_owner.show()
    voice_owner = ocr.classification(image_voice_owner)
    print('识别出的voice_owner为：' + voice_owner)
 
    image_voice_date = new_img.crop(invoice_type_map[type]['tuple_voice_date'])
    #image_voice_date.show()
    voice_date = ocr.classification(image_voice_date)
    print('识别出的voice_date为：' + voice_date)
 
    image_voice_amount = new_img.crop(invoice_type_map[type]['tuple_voice_amount'])
    #image_voice_amount.show()
    voice_amount = ocr.classification(image_voice_amount)
    print('识别出的voice_amount小写为：' + voice_amount)
 
    image_voice_amount_cn = new_img.crop(invoice_type_map[type]['tuple_voice_amount_cn'])
    #image_voice_amount_cn.show()
    voice_amount_cn = ocr.classification(image_voice_amount_cn)
    print('识别出的voice_amount_cn大写为：' + voice_amount_cn)
    print('--------------------------------------------------')
    invoice['voice_no'] = voice_no
    invoice['voice_owner'] = voice_owner
    invoice['voice_date'] = voice_date
    invoice['voice_amount'] = voice_amount
    invoice['voice_no'] = voice_no
    invoice['voice_amount_cn'] = voice_amount_cn
    #print(invoice)
    return invoice
 
def get_output_list():
 
    pass
 
#保存excel
def saveExcel(output_list):
    writer = pd.ExcelWriter('发票记录.xlsx')
    data = pd.DataFrame(output_list)
    data.to_excel(writer, 'sheet_1', float_format='%f', header=True, index=False)
    writer.close()
 
if __name__ == '__main__':
    pdf_path = r"xxx\pdf"
    pic_path = r"xxx\png"
    file_names = []
    files = traverse_directory(pic_path,file_names=file_names)
    print(files)
    output_list = []
    for img_url in files:
        new_img = readPic(img_url=img_url)
        invoice = parse_invoice(new_img);
        invoice['img_url'] = img_url
        output_list.append(invoice)
    saveExcel(output_list)

整体的识别效果只能说能够看得懂，后面会优化为一个发票识别小助手，先埋个坑