opencv BGR2GRAY和Pillow转换函数之间的区别

Question

I'm trying to OCR an image which includes both numbers and characters, using Tesseract library with opencv and c++. Before calling tesseract library, i used to gray scale the image with opencv

cvtColor(roiImg,roiImg,CV_BGR2GRAY);

This is the

OCR results for this image wasn't 100% accurate.

Then the same image was tested with pillow library with python. The original image was gray scaled using the following method.

gray = image.convert('L')

This is the

The latter mentioned gray scaled image gave 100% accurate results.

Once i searched through the internet it was mentioned that both opencv BGR2Gray and pillow img.convert methods use the same luma transform algorithm.

What is the reason for two different OCR results?

Thanks in Advance

Answer 1

Pillow can read only 3x8-bit pixels for color image.

Here a quick test to see how both libraries will round the values:

• OpenCV code:

   cv::Mat img(2, 1, CV_8UC3), img_gray; img.at<cv::Vec3b>(0, 0) = cv::Vec3b(248, 249, 249); //BGR img.at<cv::Vec3b>(1, 0) = cv::Vec3b(249, 248, 248); //BGR cv::cvtColor(img, img_gray, cv::COLOR_BGR2GRAY); std::cout << "img:\\n" << img << std::endl; std::cout << "img_gray:\\n" << img_gray << std::endl; float val1 = 249*0.299f + 249*0.587f + 248*0.114f; //RGB float val2 = 248*0.299f + 248*0.587f + 249*0.114f; //RGB std::cout << "val1=" << val1 << std::endl; std::cout << "val2=" << val2 << std::endl; 

img:

[248, 249, 249;

249, 248, 248]

img_gray:

[249;

248]

val1=248.886

val2=248.114

• Python code:

   rgbArray = np.zeros((2,1,3), 'uint8') rgbArray[0,0,0] = 249 #R rgbArray[0,0,1] = 249 #G rgbArray[0,0,2] = 248 #B rgbArray[1,0,0] = 248 #R rgbArray[1,0,1] = 248 #G rgbArray[1,0,2] = 249 #B img = Image.fromarray(rgbArray) imgGray = img.convert('L') print("rgbArray:\\n", rgbArray) print("imgGray:\\n", np.asarray(imgGray)) print("np.asarray(imgGray).dtype: ", np.asarray(imgGray).dtype) 

rgbArray:

[[[249 249 248]]

[[248 248 249]]]

imgGray:

[[248]

[248]]

np.asarray(imgGray).dtype: uint8