OpenCV tesserocr 水印检测

Question

所以我的 SQL 表中有大约 12000 个图像链接。 重点是检测哪些图像包含水印文本，哪些不包含水印文本。 所有的文字和边框都是这样的。

我已经尝试过使用 OpenCV 和 tesserocr

img_data = requests.get(url).content
img = Image.open(BytesIO(img_data))
print(url, tesserocr.image_to_text(img).strip())  

但它似乎根本无法识别图像上的文字。

我的第二种方法是使用一些外部开放 API 站点。

for url,id in alldata:
body = {'language':'eng',
        'isOverlayRequired':'false',
        'url':url,
        'iscreatesearchablepdf':'false',
        'issearchablepdfhidetextlayer':'false',
        'filetype':'jpeg'}

headers = {
    'apikey':'xxxxxxx'
}
rsp =  requests.post('https://api.ocr.space/parse/image', body, headers= headers) 

它可以工作，但速度非常慢。对于 11000 张图片，它会持续几天。

Answer 1

tesserocr 未检测到任何文本，因为文本高度较小或文本大小较小。通过裁剪文本区域并使用该图像，pytesseract 可以提取文本。 由于文本尺寸较小，使用轮廓和膨胀来检测文本区域也不起作用。为了检测文本区域，我使用EAST 模型使用this 解决方案提取所有区域并将所有区域组合在一起。将提取的组合区域图像传递给 tesseract 返回文本。要运行此脚本，您需要下载可在here 找到的模型并安装所需的依赖项。
Python 脚本：

import numpy as np
import cv2
from imutils.object_detection import non_max_suppression
import matplotlib.pyplot as plt
import pytesseract

pytesseract.pytesseract.tesseract_cmd = r"C:\Program Files\Tesseract-OCR\tesseract.exe" # I am using windows

image1 = cv2.imread("r9Do4.png",cv2.IMREAD_COLOR) 
ima_org = image1.copy()
(height1, width1) = image1.shape[:2]
size = 640 #size must be multiple of 32. Haven't tested with smaller size which can increase speed but might decrease accuracy.
(height2, width2) = (size, size)  
image2 = cv2.resize(image1, (width2, height2))  


net = cv2.dnn.readNet("frozen_east_text_detection.pb")
blob = cv2.dnn.blobFromImage(image2, 1.0, (width2, height2), (123.68, 116.78, 103.94), swapRB=True, crop=False)
net.setInput(blob)

(scores, geometry) = net.forward(["feature_fusion/Conv_7/Sigmoid", "feature_fusion/concat_3"])
(rows, cols) = scores.shape[2:4]  # grab the rows and columns from score volume
rects = []  # stores the bounding box coordiantes for text regions
confidences = []  # stores the probability associated with each bounding box region in rects

for y in range(rows):
    scoresdata = scores[0, 0, y]
    xdata0 = geometry[0, 0, y]
    xdata1 = geometry[0, 1, y]
    xdata2 = geometry[0, 2, y]
    xdata3 = geometry[0, 3, y]
    angles = geometry[0, 4, y]

    for x in range(cols):

        if scoresdata[x] < 0.5:  # if score is less than min_confidence, ignore
            continue
        # print(scoresdata[x])
        offsetx = x * 4.0
        offsety = y * 4.0
        # EAST detector automatically reduces volume size as it passes through the network
        # extracting the rotation angle for the prediction and computing their sine and cos

        angle = angles[x]
        cos = np.cos(angle)
        sin = np.sin(angle)

        h = xdata0[x] + xdata2[x]
        w = xdata1[x] + xdata3[x]
        #  print(offsetx,offsety,xdata1[x],xdata2[x],cos)
        endx = int(offsetx + (cos * xdata1[x]) + (sin * xdata2[x]))
        endy = int(offsety + (sin * xdata1[x]) + (cos * xdata2[x]))
        startx = int(endx - w)
        starty = int(endy - h)

        # appending the confidence score and probabilities to list
        rects.append((startx, starty, endx, endy))
        confidences.append(scoresdata[x])


# applying non-maxima suppression to supppress weak and overlapping bounding boxes
boxes = non_max_suppression(np.array(rects), probs=confidences)

iti=[]
rW = width1 / float(width2)
rH = height1 / float(height2)


bb = []
for (startx, starty, endx, endy) in boxes:
    startx = int(startx * rW)
    starty = int(starty * rH)
    endx = int(endx * rW)
    endy = int(endy * rH)
    cv2.rectangle(image1, (startx, starty), (endx, endy), (255, 0,0), 2)
    
    bb.append([startx, starty, endx, endy])
    

#combining the bounding boxes to get the text region

csx = 0
cex = 0
csy = 0
cey = 0
for i,box in enumerate(bb[::-1]):
    if i==0:
        csx = box[0]
    else:
        cex = box[2]
        cey = box[3]+5
        esx = box[1]
    
#print(image1)
cv2.imshow("BB img",image1)
cv2.waitKey(0)

rects.append((startx, starty, endx, endy))
confidences.append(scoresdata[x])
it=ima_org[csy:cey, csx:cex]
cv2.imshow("Cropped Text Region",it)
cv2.waitKey(0)

pytesseract.pytesseract.tesseract_cmd = r"C:\Program Files\Tesseract-OCR\tesseract.exe"
thr = cv2.threshold(src=cv2.cvtColor(it,cv2.COLOR_RGB2GRAY), thresh=0, maxval=255, type=cv2.THRESH_OTSU + cv2.THRESH_BINARY_INV)[1]
txt = pytesseract.image_to_string(thr,lang='eng',config='--psm 11')
print(txt.strip())

这是脚本的输出
文本区域的边界框

裁剪的文本区域

提取的文本
\

正如帖子中提到的，您的所有图像都相似，您可以通过重新调整此脚本的用途从中提取文本。脚本相当快，这张图片需要 2.1 秒。