使用深度学习阅读和分类扫描文档

[导读]作者|小白来源|小白学视觉收集数据首先，我们要做的第一件事是创建一个简单的数据集，这样我们就可以测试我们工作流程的每一部分。理想情况下，我们的数据集将包含各种易读性和时间段的扫描文档，以及每个文档所属的高级主题。我找不到具有这些精确规格的数据集，所以我开始构建自己的数据集。我决定...

作者|小白

来源|小白学视觉

收集数据

首先，我们要做的第一件事是创建一个简单的数据集，这样我们就可以测试我们工作流程的每一部分。理想情况下，我们的数据集将包含各种易读性和时间段的扫描文档，以及每个文档所属的高级主题。我找不到具有这些精确规格的数据集，所以我开始构建自己的数据集。我决定的高层次话题是政府、信件、吸烟和专利，随机的选择这些主要是因为每个地区都有各种各样的扫描文件。

我从这些来源中的每一个中挑选了 20 个左右的大小合适的文档，并将它们放入由主题定义的单独文件夹中。

经过将近一整天的搜索和编目所有图像后，我们将它们全部调整为 600x800 并将它们转换为 PNG 格式。

简单的调整大小和转换脚本如下：

from PIL import Image
img_folder = r'F:\Data\Imagery\OCR' # Folder containing topic folders (i.e "News", "Letters" ..etc.)
for subfol in os.listdir(img_folder): # For each of the topic folders sfpath = os.path.join(img_folder, subfol)for imgfile in os.listdir(sfpath): # Get all images in the topic imgpath = os.path.join(sfpath, imgfile) img = Image.open(imgpath) # Read in the image with Pillow img = img.resize((600,800)) # Resize the image newip = imgpath[0:-4] ".png" # Convert to PNG img.save(newip) # Save

构建OCR管道

光学字符识别是从图像中提取文字的过程。这通常是通过机器学习模型完成的，最常见的是通过包含卷积神经网络的管道来完成。虽然我们可以为我们的应用程序训练自定义 OCR 模型，但它需要更多的训练数据和计算资源。相反，我们将使用出色的 Microsoft 计算机视觉 API，其中包括专门用于 OCR 的特定模块。API 调用将使用图像（作为 PIL 图像）并输出几位信息，包括图像上文本的位置/方向作为以及文本本身。以下函数将接收一个 PIL 图像列表并输出一个大小相等的提取文本列表：

def image_to_text(imglist, ndocs=10):''' Take in a list of PIL images and return a list of extracted text using OCR ''' headers = {# Request headers'Content-Type': 'application/octet-stream','Ocp-Apim-Subscription-Key': 'YOUR_KEY_HERE', }
params = urllib.parse.urlencode({# Request parameters'language': 'en','detectOrientation ': 'true', })
outtext = [] docnum = 0
for cropped_image in imglist: print("Processing document -- ", str(docnum))# Cropped image must have both height and width > 50 px to run Computer Vision API#if (cropped_image.height or cropped_image.width) < 50:# cropped_images_ocr.append("N/A")# continue ocr_image = cropped_image imgByteArr = io.BytesIO() ocr_image.save(imgByteArr, format='PNG') imgByteArr = imgByteArr.getvalue()
try: conn = http.client.HTTPSConnection('westus.api.cognitive.microsoft.com') conn.request("POST", "/vision/v1.0/ocr?%s" % params, imgByteArr, headers) response = conn.getresponse() data = json.loads(response.read().decode("utf-8"))
curr_text = []for r in data['regions']:for l in r['lines']:for w in l['words']: curr_text.append(str(w['text'])) conn.close() except Exception as e: print("Could not process image
outtext.append(' '.join(curr_text)) docnum = 1
return(outtext)

后期处理

由于在某些情况下我们可能希望在这里结束我们的工作流程，而不是仅仅将提取的文本作为一个巨大的列表保存在内存中，我们还可以将提取的文本写入与原始输入文件同名的单个 txt 文件中。微软的OCR技术虽然不错，但偶尔也会出错。‍‍‍‍我们可以使用 SpellChecker 模块减少其中的一些错误，以下脚本接受输入和输出文件夹，读取输入文件夹中的所有扫描文档，使用我们的 OCR 脚本读取它们，运行拼写检查并纠正拼写错误的单词，最后将原始txt文件导出目录。

'''Read in a list of scanned images (as .png files > 50x50px) and output a set of .txt files containing the text content of these scans'''

from functions import preprocess, image_to_textfrom PIL import Imageimport osfrom spellchecker import SpellCheckerimport matplotlib.pyplot as plt

INPUT_FOLDER = r'F:\Data\Imagery\OCR2\Images'OUTPUT_FOLDER = r'F:\Research\OCR\Outputs\AllDocuments'

## First, read in all the scanned document images into PIL imagesscanned_docs_path = os.listdir(INPUT_FOLDER)scanned_docs_path = [x for x in scanned_docs_path if x.endswith('.png')]scanned_docs = [Image.open(os.path.join(INPUT_FOLDER, path)) for path in scanned_docs_path]

## Second, utilize Microsoft CV API to extract text from these images using OCRscanned_docs_text = image_to_text(scanned_docs)

## Third, remove mis-spellings that might have occured from bad OCR readingsspell = SpellChecker()for i in range(len(scanned_docs_text)): clean = scanned_docs_text[i] misspelled = spell.unknown(clean) clean = clean.split(" ")for word in range(len(clean)):if clean[word] in misspelled: clean[word] = spell.correction(clean[word])# Get the one `most likely` answer clean = ' '.join(clean) scanned_docs_text[i] = clean

## Fourth, write the extracted text to individual .txt files with the same name as input filesfor k in range(len(scanned_docs_text)): # For each scanned document
text = scanned_docs_text[k] path = scanned_docs_path[k] # Get the corresponding input filename
text_file_path = path[:-4] ".txt" # Create the output text file text_file = open(text_file_path, "wt")
n = text_file.write(text) # Write the text to the ouput text file
text_file.close()
print("Done")

为建模准备文本

如果我们的扫描文档集足够大，将它们全部写入一个大文件夹会使它们难以分类，并且我们可能已经在文档中进行了某种隐式分组。如果我们大致了解我们拥有多少种不同的“类型”或文档主题，我们可以使用主题建模来帮助自动识别这些。这将为我们提供基础架构，以根据文档内容将 OCR 中识别的文本拆分为单独的文件夹，我们将使用该主题模型被称为LDA。为了运行这个模型，我们需要对我们的数据进行更多的预处理和组织，因此为了防止我们的脚本变得冗长和拥挤，我们将假设已经使用上述工作流程读取了扫描的文档并将其转换为 txt 文件. 然后主题模型将读入这些 txt 文件，将它们分类到我们指定的任意多个主题中，并将它们放入适当的文件夹中。

我们将从一个简单的函数开始，读取文件夹中所有输出的 txt 文件，并将它们读入包含 (filename, text) 的元组列表。

def read_and_return(foldername, fileext='.txt'):''' Read all text files with fileext from foldername, and place them into a list of tuples as [(filename, text), ... , (filename, text)] ''' allfiles = os.listdir(foldername) allfiles = [os.path.join(foldername, f) for f in allfiles if f.endswith(fileext)] alltext = []for filename in allfiles:with open(filename, 'r') as f: alltext.append((filename, f.read())) f.close()return(alltext) # Returns list of tuples [(filename, text), ... (filename,text)] 接下来，我们需要确保所有无用的词（那些不能帮助我们区分特定文档主题的词）。我们将使用三种不同的方法来做到这一点：

删除停用词
去除标签、标点、数字和多个空格
TF-IDF 过滤

为了实现所有这些（以及我们的主题模型），我们将使用 Gensim 包。下面的脚本将对文本列表（上述函数的输出）运行必要的预处理步骤并训练 LDA 模型。

from gensim import corpora, models, similaritiesfrom gensim.parsing.preprocessing import remove_stopwords, preprocess_string
def preprocess(document): clean = remove_stopwords(document) clean = preprocess_string(document) return(clean)
def run_lda(textlist, num_topics=10, preprocess_docs=True):
''' Train and return an LDA model against a list of documents '''if preprocess_docs: doc_text = [preprocess(d) for d in textlist]
dictionary = corpora.Dictionary(doc_text) corpus = [dictionary.doc2bow(text) for text in doc_text] tfidf = models.tfidfmodel.TfidfModel(corpus) transformed_tfidf = tfidf[corpus]
lda = models.ldamulticore.LdaMulticore(transformed_tfidf, num_topics=num_topics, id2word=dictionary)
return(lda, dictionary)

使用模型对文档进行分类

一旦我们训练了我们的 LDA 模型，我们就可以使用它来将我们的训练文档集（以及可能出现的未来文档）分类为主题，然后将它们放入适当的文件夹中。

对新的文本字符串使用经过训练的 LDA 模型需要一些麻烦，所有的复杂性都包含在下面的函数中：

def find_topic(textlist, dictionary, lda):''' https://stackoverflow.com/questions/16262016/how-to-predict-the-topic-of-a-new-query-using-a-trained-lda-model-using-gensim
For each query ( document in the test file) , tokenize the query, create a feature vector just like how it was done while training and create text_corpus '''
text_corpus = []
for query in textlist: temp_doc = tokenize(query.strip()) current_doc = [] temp_doc = list(temp_doc)for word in range(len(temp_doc)): current_doc.append(temp_doc[word])
text_corpus.append(current_doc)''' For each feature vector text, lda[doc_bow] gives the topic distribution, which can be sorted in descending order to print the very first topic ''' tops = []for text in text_corpus: doc_bow = dictionary.doc2bow(text) topics = sorted(lda[doc_bow],key=lambda x:x[1],reverse=True)[0] tops.append(topics)return(tops) 最后，我们需要另一种方法来根据主题索引获取主题的实际名称：

def topic_label(ldamodel, topicnum): alltopics = ldamodel.show_topics(formatted=False) topic = alltopics[topicnum] topic = dict(topic[1])return(max(topic, key=lambda key: topic[key]))

现在，我们可以将上面编写的所有函数粘贴到一个接受输入文件夹、输出文件夹和主题计数的脚本中。该脚本将读取输入文件夹中所有扫描的文档图像，将它们写入txt 文件，构建LDA 模型以查找文档中的高级主题，并根据文档主题将输出的txt 文件归类到文件夹中。

################################################################## This script takes in an input folder of scanned documents ## and reads these documents, seperates them into topics ## and outputs raw .txt files into the output folder, seperated ## by topic ##################################################################
import osfrom PIL import Imageimport base64import http.client, urllib.request, urllib.parse, urllib.error, base64import ioimport jsonimport requestsimport urllibfrom gensim import corpora, models, similaritiesfrom gensim.utils import tokenizefrom gensim.parsing.preprocessing import remove_stopwords, preprocess_stringimport httpimport shutilimport tqdm

def filter_for_english(text): dict_url = 'https://raw.githubusercontent.com/first20hours/' \'google-10000-english/master/20k.txt'
dict_words = set(requests.get(dict_url).text.splitlines())
english_words = tokenize(text) english_words = [w for w in english_words if w in list(dict_words)] english_words = [w for w in english_words if (len(w)>1 or w.lower()=='i')]return(' '.join(english_words))

def preprocess(document): clean = filter_for_english(document) clean = remove_stopwords(clean) clean = preprocess_string(clean)
# Remove non-english words

return(clean)

def read_and_return(foldername, fileext='.txt', delete_after_read=False): allfiles = os.listdir(foldername) allfiles = [os.path.join(foldername, f) for f in allfiles if f.endswith(fileext)] alltext = []for filename in allfiles:with open(filename, 'r') as f: alltext.append((filename, f.read())) f.close()if delete_after_read: os.remove(filename)return(alltext) # Returns list of tuples [(filename, text), ... (filename,text)]

def image_to_text(imglist, ndocs=10):''' Take in a list of PIL images and return a list of extracted text ''' headers = {# Request headers'Content-Type': 'application/octet-stream','Ocp-Apim-Subscription-Key': '89279deb653049078dd18b1b116777ea', }
params = urllib.parse.urlencode({# Request parameters'language': 'en','detectOrientation ': 'true', })
outtext = [] docnum = 0
for cropped_image in tqdm.tqdm(imglist, total=len(imglist)):# Cropped image must have both height and width > 50 px to run Computer Vision API#if (cropped_image.height or cropped_image.width) < 50:# cropped_images_ocr.append("N/A")# continue ocr_image = cropped_image imgByteArr = io.BytesIO() ocr_image.save(imgByteArr, format='PNG') imgByteArr = imgByteArr.getvalue()
try: conn = http.client.HTTPSConnection('westus.api.cognitive.microsoft.com') conn.request("POST", "/vision/v1.0/ocr?%s" % params, imgByteArr, headers) response = conn.getresponse() data = json.loads(response.read().decode("utf-8"))
curr_text = []for r in data['regions']:for l in r['lines']:for w in l['words']: curr_text.append(str(w['text'])) conn.close()except Exception as e: print("[Errno {0}] {1}".format(e.errno, e.strerror))
outtext.append(' '.join(curr_text)) docnum = 1
return(outtext)

def run_lda(textlist, num_topics=10, return_model=False, preprocess_docs=True):
''' Train and return an LDA model against a list of documents '''if preprocess_docs: doc_text = [preprocess(d) for d in textlist] dictionary = corpora.Dictionary(doc_text)
corpus = [dictionary.doc2bow(text) for text in doc_text] tfidf = models.tfidfmodel.TfidfModel(corpus) transformed_tfidf = tfidf[corpus]
lda = models.ldamulticore.LdaMulticore(transformed_tfidf, num_topics=num_topics, id2word=dictionary)
input_doc_topics = lda.get_document_topics(corpus)
return(lda, dictionary)

def find_topic(text, dictionary, lda):''' https://stackoverflow.com/questions/16262016/how-to-predict-the-topic-of-a-new-query-using-a-trained-lda-model-using-gensim
For each query ( document in the test file) , tokenize the query, create a feature vector just like how it was done while training and create text_corpus '''
text_corpus = []
for query in text: temp_doc = tokenize(query.strip()) current_doc = [] temp_doc = list(temp_doc)for word in range(len(temp_doc)): current_doc.append(temp_doc[word])
text_corpus.append(current_doc)''' For each feature vector text, lda[doc_bow] gives the topic distribution, which can be sorted in descending order to print the very first topic ''' tops = []for text in text_corpus: doc_bow = dictionary.doc2bow(text) topics = sorted(lda[doc_bow],key=lambda x:x[1],reverse=True)[0] tops.append(topics)return(tops)

def topic_label(ldamodel, topicnum): alltopics = ldamodel.show_topics(formatted=False) topic = alltopics[topicnum] topic = dict(topic[1])import operatorreturn(max(topic, key=lambda key: topic[key]))

INPUT_FOLDER = r'F:/Research/OCR/Outputs/AllDocuments'OUTPUT_FOLDER = r'F:/Research/OCR/Outputs/AllDocumentsByTopic'TOPICS = 4
if __name__ == '__main__':

print("Reading scanned documents")## First, read in all the scanned document images into PIL images scanned_docs_fol = r'F:/Research/OCR/Outputs/AllDocuments' scanned_docs_path = os.listdir(scanned_docs_fol) scanned_docs_path = [os.path.join(scanned_docs_fol, p) for p in scanned_docs_path] scanned_docs = [Image.open(x) for x in scanned_docs_path if x.endswith('.png')]
## Second, utilize Microsoft CV API to extract text from these images using OCR scanned_docs_text = image_to_text(scanned_docs)

print("Post-processing extracted text")## Third, remove mis-spellings that might have occured from bad OCR readings spell = SpellChecker()for i in range(len(scanned_docs_text)): clean = scanned_docs_text[i] misspelled = spell.unknown(clean) clean = clean.split(" ")for word in range(len(clean)):if clean[word] in misspelled: clean[word] = spell.correction(clean[word])# Get the one `most likely` answer clean = ' '.join(clean) scanned_docs_text[i] = clean

print("Writing read text into files")## Fourth, write the extracted text to individual .txt files with the same name as input filesfor k in range(len(scanned_docs_text)): # For each scanned document
text = scanned_docs_text[k] text = filter_for_english(text) path = scanned_docs_path[k] # Get the corresponding input filename path = path.split("\\")[-1] text_file_path = OUTPUT_FOLDER "//" path[0:-4] ".txt" # Create the output text file text_file = open(text_file_path, "wt")
n = text_file.write(text) # Write the text to the ouput text file
text_file.close()

# First, read all the output .txt files print("Reading files") texts = read_and_return(OUTPUT_FOLDER)

print("Building LDA topic model")# Second, train the LDA model (pre-processing is internally done) print("Preprocessing Text") textlist = [t[1] for t in texts] ldamodel, dictionary = run_lda(textlist, num_topics=TOPICS)

# Third, extract the top topic for each document print("Extracting Topics") topics = []for t in texts: topics.append((t[0], find_topic([t[1]], dictionary, ldamodel)))

# Convert topics to topic namesfor i in range(len(topics)): topnum = topics[i][1][0][0]#print(topnum) topics[i][1][0] = topic_label(ldamodel, topnum)# [(filename, topic), ..., (filename, topic)]

# Create folders for the topics print("Copying Documents into Topic Folders") foundtopics = []for t in topics: foundtopics = t[1] foundtopics = set(foundtopics) topicfolders = [os.path.join(OUTPUT_FOLDER, f) for f in foundtopics] topicfolders = set(topicfolders) [os.makedirs(m) for m in topicfolders]
# Copy files into appropriate topic foldersfor t in topics: filename, topic = t src = filename filename = filename.split("\\") dest = os.path.join(OUTPUT_FOLDER, topic[0]) dest = dest "/" filename[-1] copystr = "copy " src " " dest shutil.copyfile(src, dest) os.remove(src)
print("Done")

本文代码Github链接：
使用深度学习阅读和分类扫描文档