參考：
python-opencv表格識別

思路：
1.提取出橫線
2.提取出縱線
3.得到交叉點，刪除錯誤的交叉點，兩個交叉點距離小于10，取坐標值小的那個交叉點，得到表格行列
4.對每個單元格使用paddle ocr提取文字

在原文代碼基礎上修改了2點
1.pytesseract識別準確率不高，使用paddle ocr代替 pytesseract
2.識別出的表格交叉點有些并非真實交叉點，通過判斷該行像素點個數，丟掉錯誤橫縱坐標

import cv2 import numpy as np import pandas as pd# import pytesseract # import re from paddleocr import PaddleOCRsrc = 'image.png'raw = cv2.imread(src, 1) # 灰度圖片 gray = cv2.cvtColor(raw, cv2.COLOR_BGR2GRAY) # 圖片二值化 使二值化后的圖片是黑底白字 binary = cv2.adaptiveThreshold(~gray, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C, cv2.THRESH_BINARY, 35, -5) # src：灰度化的圖片 # maxValue：滿足條件的像素點需要設置的灰度值 # adaptiveMethod：自適應方法。有2種：ADAPTIVE_THRESH_MEAN_C 或 ADAPTIVE_THRESH_GAUSSIAN_C # thresholdType：二值化方法，可以設置為THRESH_BINARY或者THRESH_BINARY_INV # blockSize：分割計算的區域大小，取奇數 # C：常數，每個區域計算出的閾值的基礎上在減去這個常數作為這個區域的最終閾值，可以為負數 # dst：輸出圖像，可選 # 展示圖片 # cv2.imshow("binary_picture", binary) # cv2.waitKey()def recognize_bgkx(binary):rows, cols = binary.shapescale = 30 # 值越小 橫線越少 40# 自適應獲取核值# 識別橫線:kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (cols // scale, 1))# 矩形eroded = cv2.erode(binary, kernel, iterations=1) # 腐蝕dilated_col = cv2.dilate(eroded, kernel, iterations=1) # 膨脹# cv2.imshow("excel_horizontal_line", dilated_col)# cv2.waitKey()# 識別豎線：scale = 20kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (1, rows // scale))eroded = cv2.erode(binary, kernel, iterations=1)dilated_row = cv2.dilate(eroded, kernel, iterations=1)# cv2.imshow("excel_vertical_line：", dilated_row)# cv2.waitKey()# 將識別出來的橫豎線合起來 對二進制數據進行“與”操作bitwise_and = cv2.bitwise_and(dilated_col, dilated_row)cv2.imshow("excel_bitwise_and", bitwise_and)cv2.waitKey()# 標識表格輪廓# merge = cv2.add(dilated_col, dilated_row) # 進行圖片的加和# cv2.imshow("entire_excel_contour：", merge)# cv2.waitKey()# 兩張圖片進行減法運算，去掉表格框線# merge2 = cv2.subtract(binary, merge)# cv2.imshow("binary_sub_excel_rect", merge2)# cv2.waitKey()# new_kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (2, 2))# erode_image = cv2.morphologyEx(merge2, cv2.MORPH_OPEN, new_kernel)# cv2.imshow('erode_image2', erode_image)# cv2.waitKey()# merge3 = cv2.add(erode_image, bitwise_and)# cv2.imshow('merge3', merge3)# cv2.waitKey()# 將焦點標識取出來ys, xs = np.where(bitwise_and > 0)# 橫縱坐標數組y_point_arr = []x_point_arr = []# 通過排序，排除掉相近的像素點，只取相近值的最后一點# 這個10就是兩個像素點的距離，不是固定的，根據不同的圖片會有調整，基本上為單元格表格的高度（y坐標跳變）和長度（x坐標跳變）i = 0sort_x_point = np.sort(xs)for i in range(len(sort_x_point) - 1):if sort_x_point[i + 1] - sort_x_point[i] > 10:x_point_arr.append(sort_x_point[i])i = i + 1# 要將最后一個點加入x_point_arr.append(sort_x_point[i])i = 0sort_y_point = np.sort(ys)# print(np.sort(ys))for i in range(len(sort_y_point) - 1):if (sort_y_point[i + 1] - sort_y_point[i] > 10):y_point_arr.append(sort_y_point[i])i = i + 1y_point_arr.append(sort_y_point[i])# 橫縱坐標超過3個 代表點對應的行、列超過2data = pd.DataFrame(bitwise_and)drop_y_list = []for i in y_point_arr:# for j in x_point_arr: data[(data.loc[i-1,:]>0)].index.tolist()# y_dot_num = [x for x in data.loc[i-1,:] if x!=0]y_dot_num = 0y_dot_num += len(data.loc[(data.loc[i, :] > 0)].index.tolist())for j in range(1, 5):y_dot_num += len(data.loc[(data.loc[i + j, :] > 0)].index.tolist())y_dot_num += len(data.loc[(data.loc[i - j, :] > 0)].index.tolist())if y_dot_num < 5:print('縱坐標%s并不在框線上刪除，該行只有%s個像素' % (i, y_dot_num))drop_y_list.append(i)for y in drop_y_list:y_point_arr.remove(y)drop_x_list = []for m in x_point_arr:# for j in x_point_arr: data[(data.loc[i-1,:]>0)].index.tolist()# y_dot_num = [x for x in data.loc[i-1,:] if x!=0]x_dot_num = 0x_dot_num += len(data.loc[(data.loc[:, m] > 0)].index.tolist())print('檢測', m)for n in range(1, 5):x_dot_num += len(data.loc[(data.loc[:, m + n] > 0)].index.tolist())x_dot_num += len(data.loc[(data.loc[:, m - n] > 0)].index.tolist())if x_dot_num < 5:print('橫坐標坐標%s并不在框線上刪除 ，該列只有%s個像素' % (m, x_dot_num))drop_x_list.append(m)for x in drop_x_list:x_point_arr.remove(x)print('該表格有%s行%s列 ' % (len(y_point_arr) - 1, len(x_point_arr) - 1))return x_point_arr, y_point_arrx_point_arr, y_point_arr = recognize_bgkx(binary) # 退后ocr = PaddleOCR(use_angle_cls=True, lang="ch")def recognize_text_by_loop():y_point_arr = [x - 3 for x in y_point_arr]# 循環y坐標，x坐標分割表格data = [[] for i in range(len(y_point_arr))]for i in range(len(y_point_arr) - 1):# if i==0:# continuefor j in range(len(x_point_arr) - 1):# 在分割時，第一個參數為y坐標，第二個參數為x坐標cell = gray[y_point_arr[i]:y_point_arr[i + 1],x_point_arr[j]:x_point_arr[j + 1]]# cv2.imshow("sub_pic" + str(i) + str(j), cell)# cv2.waitKey()# cv2.destroyAllWindows()img_path = "cell_image_"+str(i)+'_'+str(j)+".png"cv2.imwrite(img_path,cell)# 輸入待識別圖片路徑# 輸出結果保存路徑result = ocr.ocr(img_path, cls=True)text1 = ''.join([x[1][0] for x in result])print(text1)data[i].append(text1)j = j + 1i = i + 1print(data)table_coordinate = pd.DataFrame(data[1:-1],columns=data[0])table_coordinate.to_excel('saomiaojian.xlsx',index=False)

待優化：

• 真實圖片的單元格并不在一條橫坐標或縱坐標下，使用框線交叉點坐標確定單元格，實際圖像不夠完整，待后續優化，獲取每個單元格坐標

添加優化代碼：
1.找到表格四個角點
2.做平移變換

因為已經平移變換，就刪除了一些代碼，完整優化代碼如下：

import cv2 import pandas as pd import numpy as npfrom paddleocr import PaddleOCRocr = PaddleOCR(use_angle_cls=True, lang="ch")class TableOCR(object):def __init__(self,result_file=''):self.img_file = ''if not result_file:self.result_file = 'result.xlsx'def get_sorted_rect(self, rect):'''獲取排序的四個坐標@param rect:@return:按照左上 右上 右下 左下排列返回'''mid_x = (max([x[1] for x in rect]) - min([x[1] for x in rect])) * 0.5 + min([x[1] for x in rect]) # 中間點坐標left_rect = [x for x in rect if x[1] < mid_x]left_rect.sort(key=lambda x: (x[0], x[1]))right_rect = [x for x in rect if x[1] > mid_x]right_rect.sort(key=lambda x: (x[0], x[1]))sorted_rect = left_rect[0], left_rect[1], right_rect[1], right_rect[0]return sorted_rectdef get_table(self, gray, min_table_area=0):'''從灰度圖獲取表格坐標，[[右下→左下→左上→右上],..]邊緣檢測+膨脹---》找最外圍輪廓點，根據面積篩選---》根據縱坐標排序---》計算輪廓的四個點，再次篩選@param gray:灰度圖 如果是二值圖會報錯@return:'''canny = cv2.Canny(gray, 200, 255) # 第一個閾值和第二個閾值kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (3, 3))canny = cv2.dilate(canny, kernel)_, contours, HIERARCHY = cv2.findContours(canny, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)if not min_table_area:min_table_area = gray.shape[0] * gray.shape[1] * 0.01 # 50000 # 最小的矩形面積閾值candidate_table = [cnt for cnt in contours if cv2.contourArea(cnt) > min_table_area] # 計算該輪廓的面積candidate_table = sorted(candidate_table, key=cv2.contourArea, reverse=True)area_list = [cv2.contourArea(cnt) for cnt in candidate_table]table = []for i in range(len(candidate_table)):# 遍歷所有輪廓# cnt是一個點集cnt = candidate_table[i]# 找到最小的矩形，該矩形可能有方向rect = cv2.minAreaRect(cnt)# box是四個點的坐標box = cv2.boxPoints(rect)box = np.int0(box)sorted_box = self.get_sorted_rect(box)result = [sorted_box[2], sorted_box[3], sorted_box[0], sorted_box[1]] # 右下 左下 左上 右上result = [x.tolist() for x in result]table.append(result)return tabledef perTran(self, image, rect):'''做透視變換image 圖像rect 四個頂點位置:左上 右上 右下 左下'''tl, tr, br, bl = rect # 左下 右下 左上 右上 || topleft topright 左上 右上 右下 左下# 計算寬度widthA = np.sqrt(((br[0] - bl[0]) ** 2) + ((br[1] - bl[1]) ** 2))widthB = np.sqrt(((tr[0] - tl[0]) ** 2) + ((tr[1] - tl[1]) ** 2))maxWidth = max(int(widthA), int(widthB))# 計算高度heightA = np.sqrt(((tr[0] - br[0]) ** 2) + ((tr[1] - br[1]) ** 2))heightB = np.sqrt(((tl[0] - bl[0]) ** 2) + ((tl[1] - bl[1]) ** 2))maxHeight = max(int(heightA), int(heightB))# 定義變換后新圖像的尺寸dst = np.array([[0, 0], [maxWidth - 1, 0], [maxWidth - 1, maxHeight - 1],[0, maxHeight - 1]], dtype='float32')# 變換矩陣rect = np.array(rect, dtype=np.float32)dst = np.array(dst, dtype=np.float32)M = cv2.getPerspectiveTransform(rect, dst)# 透視變換warped = cv2.warpPerspective(image, M, (maxWidth, maxHeight))return warpeddef recognize_bgkx(self,binary):rows, cols = binary.shapescale = 30 # 值越小 橫線越少 40# 自適應獲取核值# 識別橫線:kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (cols // scale, 1))# 矩形eroded = cv2.erode(binary, kernel, iterations=1) # 腐蝕dilated_col = cv2.dilate(eroded, kernel, iterations=1) # 膨脹# cv2.imshow("excel_horizontal_line", dilated_col)# cv2.waitKey()# 識別豎線：scale = 20kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (1, rows // scale))eroded = cv2.erode(binary, kernel, iterations=1)dilated_row = cv2.dilate(eroded, kernel, iterations=1)# cv2.imshow("excel_vertical_line：", dilated_row)# cv2.waitKey()# 將識別出來的橫豎線合起來 對二進制數據進行“與”操作bitwise_and = cv2.bitwise_and(dilated_col, dilated_row)# cv2.imshow("excel_bitwise_and", bitwise_and)# cv2.waitKey()# 標識表格輪廓# merge = cv2.add(dilated_col, dilated_row) # 進行圖片的加和# cv2.imshow("entire_excel_contour：", merge)# cv2.waitKey()# 兩張圖片進行減法運算，去掉表格框線# merge2 = cv2.subtract(binary, merge)# cv2.imshow("binary_sub_excel_rect", merge2)# cv2.waitKey()# new_kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (2, 2))# erode_image = cv2.morphologyEx(merge2, cv2.MORPH_OPEN, new_kernel)# cv2.imshow('erode_image2', erode_image)# cv2.waitKey()# merge3 = cv2.add(erode_image, bitwise_and)# cv2.imshow('merge3', merge3)# cv2.waitKey()# 將焦點標識取出來ys, xs = np.where(bitwise_and > 0)# 橫縱坐標數組y_point_arr = []x_point_arr = []# 通過排序，排除掉相近的像素點，只取相近值的最后一點# 這個10就是兩個像素點的距離，不是固定的，根據不同的圖片會有調整，基本上為單元格表格的高度（y坐標跳變）和長度（x坐標跳變）i = 0sort_x_point = np.sort(list(set(xs)))for i in range(len(sort_x_point) - 1):if sort_x_point[i + 1] - sort_x_point[i] > 10:x_point_arr.append(sort_x_point[i])i = i + 1# 要將最后一個點加入x_point_arr.append(sort_x_point[i])i = 0sort_y_point = np.sort(list(set(ys)))# print(np.sort(ys))for i in range(len(sort_y_point) - 1):if (sort_y_point[i + 1] - sort_y_point[i] > 10):y_point_arr.append(sort_y_point[i])i = i + 1y_point_arr.append(sort_y_point[i])return x_point_arr, y_point_arrdef recognize_text_by_loop(self,gray,x_point_arr,y_point_arr):# y_point_arr = [x - 3 for x in y_point_arr]# 循環y坐標，x坐標分割表格data = [[] for i in range(len(y_point_arr))]for i in range(len(y_point_arr) - 1):# if i==0:# continuefor j in range(len(x_point_arr) - 1):# 在分割時，第一個參數為y坐標，第二個參數為x坐標cell = gray[y_point_arr[i]:y_point_arr[i + 1],x_point_arr[j]:x_point_arr[j + 1]]# cv2.imshow("sub_pic" + str(i) + str(j), cell)# cv2.waitKey()# cv2.destroyAllWindows()# img_path = "cell_image_" + str(i) + '_' + str(j) + ".png"# cv2.imwrite(img_path, cell)# 輸入待識別圖片路徑# 輸出結果保存路徑result = ocr.ocr(cell)text1 = ''.join([x[1][0] for x in result])print(text1)data[i].append(text1)print(data)df = pd.DataFrame(data[1:-1], columns=data[0])return dfdef ocr(self,img_file):self.img_file = img_fileimg = cv2.imread(img_file)gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)rect = self.get_table(gray)sorted_rect = self.get_sorted_rect(rect[0])gray_z = self.perTran(gray, sorted_rect) #如果有多個table 可以循環cv2.imwrite('gray_z.jpg', gray_z)binary_z = cv2.adaptiveThreshold(~gray_z, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C,cv2.THRESH_BINARY, 15, -5)x_point_arr, y_point_arr = self.recognize_bgkx(binary_z)df = self.recognize_text_by_loop(gray_z,x_point_arr, y_point_arr)df.to_excel(self.result_file, index=False)tableOCR = TableOCR() tableOCR.ocr('table.png')

效果：

結果：

總結

以上是生活随笔為你收集整理的opencv+paddle orc 识别图片提取表格信息的全部內容，希望文章能夠幫你解決所遇到的問題。

如果覺得生活随笔網站內容還不錯，歡迎將生活随笔推薦給好友。