利用cv2提取pdf中的表格

依据线条绘制的表格

延长每根线条，使其达到表格边缘

不规范格式的表格

不规范格式的表格提取线条后

延长线条后

实际的表格

在绘制线条中，出现了额外的线条

内容类似OCR的识别结果

非常生动的细节描述

使用三次贝塞尔曲线绘制的文字

对于此类PDF，至今我未能找到解决办法，如果有知道答案的小伙伴，希望能在评论区告诉我解决方案。

结尾

在最后将所有修改后的代码献上

import fitz import numpy as np import cv2 import itertools import copy def to_int(*kwargs): v = [] for k in kwargs: v.append(int(k)) return v def page_to_words_list(page: fitz.fitz.Page) -> list: ''' 将每一页中的textWords信息使用list封装，这样方便后续使用 :param page: :return: ''' # 获取文字及坐标信息 words = page.getTextWords() # 将元素转为list # 因为list[0],list[1]....对于不熟悉代码的人很容易忘记含义，所以用字典封装 # words = [[w[0], w[1], w[2], w[3], w[4]] for w in words] # 此处更适合用实体类，但是调试print的时候不方便，虽然可以重写__str__ word_list = [] for w in words: if '页码' in w[4]: continue # 有些文字旋转过，需要旋转回来 p1 = fitz.Point(w[0], w[1]) * page.rotation_matrix p2 = fitz.Point(w[2], w[3]) * page.rotation_matrix # 旋转后矩形点位置发生改变，需要还原 p3 = min(p1[0], p2[0]), min(p1[1], p2[1]) p4 = max(p1[0], p2[0]), max(p1[1], p2[1]) word_list.append({'rect': [p3[0], p3[1], p4[0], p4[1]], 'text': w[4]}) # 按y坐标排序 word_list = sorted(word_list, key=lambda data: (data['rect'][1], data['rect'][0])) return word_list def draw_pdf_tables(page: fitz.fitz.Page): assert isinstance(page, fitz.fitz.Page), '必须传入fitz.Page对象' # 创建一个白色的画布 pixmap = page.getPixmap(matrix=fitz.Matrix(1, 1)) # 二进制数据，宽，高 img = np.zeros([pixmap.h, pixmap.w], dtype=np.uint8) + 255 draws = page.getDrawings() # 在白色的画布上，画上黑色的线条 for draw in draws: color = draw['color'] fill = draw['fill'] if (color == [1.0, 1.0, 1.0] and fill is None) or (fill == [1.0, 1.0, 1.0] and color is None): continue items_ = draw['items'] for item_ in items_: # print(item) item_ = list(item_) # 线条 if 'l' == item_[0]: p1, p2 = to_int(*item_[1]), to_int(*item_[2]) img = cv2.line(img, (p1[0], p1[1]), (p2[0], p2[1]), (0)) elif 're' == item_[0]: p = to_int(*item_[1]) img = cv2.rectangle(img, (p[0], p[1]), (p[2], p[3]), (0)) # elif 'c' == item_[0]: # print('c', item_) # else: # print(item_) # cv2.imshow('1234', img) # cv2.waitKey(0) # 使用漫水填充算法，将周围变为黑色 # 这样也可以去掉单独的线条 cv2.floodFill(img, None, (1, 1), (0), cv2.FLOODFILL_FIXED_RANGE) kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (3, 3)) img = cv2.morphologyEx(img, cv2.MORPH_OPEN, kernel, iterations=2) return img def get_table_words(page: fitz.fitz.Page, words=None): ''' :param page:一页pdf :param words:从pdf中提取的无序文字 :return: ''' assert isinstance(page, fitz.fitz.Page), '必须传入fitz.Page对象' if words is None: words = page_to_words_list(page) img = draw_pdf_tables(page) # 查找相应的轮廓，得到每个表格cell的矩形框 contours, hierarchy = cv2.findContours(img, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE) table_cell = [] # 判断文字是否在表格cell中 for c in contours: r = cv2.boundingRect(c) r = [r[0], r[1], r[0] + r[2], r[1] + r[3]] ws = [] for word in words[:]: w = word['rect'] center = [(w[0] + w[2]) / 2, (w[1] + w[3]) / 2] if inside_rectangle(center, r): ws.append(word) table_cell.append({'rect': r, 'words': ws}) # 闭运算 kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (3, 3)) morp = cv2.morphologyEx(img, cv2.MORPH_CLOSE, kernel, iterations=3) # 查找相应的轮廓，得到每个表格cell的矩形框 contours, hierarchy = cv2.findContours(morp, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE) tables = [] for c in contours: r = cv2.boundingRect(c) r = [r[0], r[1], r[0] + r[2], r[1] + r[3]] tables.append(r) # 排序 table_cell = sorted(table_cell, key=lambda data: (data['rect'][1], data['rect'][0])) tables = sorted(tables, key=lambda data: (data[1], data[0])) # 将cell合并到表中 tables_words = [] for t in tables: table = {'rect': t, 'cell': []} for cell in table_cell: c = cell['rect'] center = [(c[0] + c[2]) / 2, (c[1] + c[3]) / 2] if inside_rectangle(center, t): table['cell'].append(cell) tables_words.append(table) return tables_words, img def inside_rectangle(point, rect): ''' 判断点是否在框内 :param point: :param rect: :return: ''' x, y = point[0], point[1] x1, y1, x2, y2 = rect if x1 <= x <= x2 and y1 <= y <= y2: return True return False def get_small_cell(table_word, img): ''' 将复杂格式的表格生成最小单元的表格 此方法是根据四周的点来确定最小单位的表格信息，但是不能避免极端情况 :param table_word: :param img: :return: ''' t_r = table_word['rect'] table_img = copy.deepcopy(img) cells = table_word['cell'] # 将每个格子的线条都撑到最大 for cell in cells: r = cell['rect'] table_img[:, r[0] - t_r[0]] = 0 table_img[:, r[2] - t_r[2]] = 0 table_img[r[1] - t_r[1]] = 0 table_img[r[3] - t_r[3]] = 0 # cv2.imshow('img',img) # cv2.imshow('get_small_cell', table_img) # 开运算，避免细小漏洞 kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (3, 3)) table_img = cv2.morphologyEx(table_img, cv2.MORPH_OPEN, kernel, iterations=3) # cv2.imshow('morphologyEx', table_img) # cv2.waitKey(0) # cv2.destroyAllWindows() cells = [] # 查找相应的轮廓，得到每个表格cell的矩形框 contours, hierarchy = cv2.findContours(table_img, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE) for c in contours: r = cv2.boundingRect(c) r = [r[0] + t_r[0], r[1] + t_r[1], r[0] + r[2] + t_r[0], r[1] + r[3] + t_r[1]] cells.append(r) cells = sorted(cells, key=lambda data: (data[1], data[0])) cells_group = itertools.groupby(cells, key=lambda x: (x[1])) return cells_group def equal_rect(r1, r2, border): if abs(r1[0] - r2[0]) < border and abs(r1[1] - r2[1]) < border and abs(r1[2] - r2[2]) < border and abs(r1[3] - r2[3]) < border: return True return False def table_parse(table, img, border=5): ''' 解析表格，形成最终的表格数据 :param table: :param img: :return: ''' table_cell = table['cell'] # 延长表格中的线条，获取到最小的单元格，并按行分组 cells_group = get_small_cell(table, img) # i为行坐标 for i, (k, line_cells) in enumerate(cells_group): line_cells = list(line_cells) # j为列坐标 for j, c in enumerate(line_cells): for cell in table_cell: center = [(c[0] + c[2]) / 2, (c[1] + c[3]) / 2] ''' 如果最小单元格的格子中心，落在表格中，那么他一定是属于这个表格的 因为上文中已经对所有的格子做了x,y轴排序，此处只需对比当前格子和上一个格子的位置关系,就能确定跨行跨列的相关信息 inside是指cell中内部的上一次遇到的表格 ''' if inside_rectangle(center, cell['rect']): # if i == 6: # print('1234') r = cell['rect'] # 起点或者两个框相等 if equal_rect(r, c, border) or (abs(r[0] - c[0]) < border and abs(r[1] - c[1]) < border): cell['col'], cell['row'] = j, i cell['colspan'], cell['rowspan'] = 1, 1 cell['inside'] = c elif 'inside' in cell: # 纵坐标差不多，表示同一行 if abs(cell['inside'][1] - c[1]) < border: cell['colspan'] += 1 cell['inside'] = c # 下面格子顶坐标和上面格子底坐标 elif abs(cell['inside'][3] - c[1]) < border: cell['rowspan'] += 1 cell['inside'] = c else: print(r, c, cell['inside'], i, j) break if __name__ == '__main__': path = r'E:temp603回单1.pdf' # 加载pdf文件 doc = fitz.open(path) # 取第一页数据 page = doc[0] # 获取第一页中所有的表格文字 table_words, img = get_table_words(page) # 获取第一个表格 table = table_words[0] # 将表格的数据 table_parse(table, img) # 测试，显示表格 img = cv2.cvtColor(img, cv2.COLOR_GRAY2BGR) table_cell = table['cell'] for cell in table_cell: p = cell['rect'] print(cell) cv2.rectangle(img, (p[0], p[1]), (p[2], p[3]), (0, 255, 0)) cv2.imshow('123', img) cv2.waitKey(0)