Unclear text extraction by Tesseract

Question

I am implementing an OCR system. Tesseract API is used for extraction. Images are preprocessed using OpenCV before extracting text. Under preprocessing grayscaling, sharpening and adaptive thresholding is carried out. After extracting a text in the image the following output is gained.

Expected Output

Let's talk ;-)

Gained output

" yr _ W??? V. ? _
W fag '7? |g§3:? V
é claw?!

Does anybody know the reason for this? I edited the question as took a different path to implement my project.I input an image and used opnCV to sharpen the image. This is the input image.input image

Then I got the following output.Sharpened output image. When I use the sharpened image for Tesseract Api it provides mixture of characters. But if I provide the input image to Tesseract API, it correctly extract the words. How can I remove those shaded areas in the sharpened image?

This is the code I used to sharp the input image

try {
            System.loadLibrary(Core.NATIVE_LIBRARY_NAME);
            Mat source = Imgcodecs.imread("input.jpg",
                    Imgcodecs.CV_LOAD_IMAGE_GRAYSCALE);
            Mat destination = new Mat(source.rows(), source.cols(), source.type());

            Imgproc.equalizeHist(source, destination);
            Imgcodecs.imwrite("sharpen.jpg", destination);

        } catch (Exception e) {
            System.out.println("error: " + e.getMessage());
        } 

Answer 1

Well, you should provide us with input image at least, so we could better see, what is the problem. But as seen from the expected and actual output, that your input image is very bad for scanning, so there could be a few possible (most common) issues with it:

• Your input image could be rotated too much (maybe even rotated by 90 degrees or so (I had this issue, when the photo, taken by android phone was turned by 90 degrees and scan results were terrible). Tesseract can not recognize the skew and rotate your image).
• Characters in the image have too small resolution (try to make them at least 10-12 pix, below that, your scan output will drop significantly, and the higher the resolution, the faster and more accurate your scan will be (processing for bigger resolution images, depending on how you do it, will be slower, so there should be some balance).
• Image could be too blurry, so letters cannon be recognized.
• Letters are mixed with shadows or other objects, too dark or too much influenced by environment to be scanned.

You should provide us with your input image and how you process it, it would be much easier to find your problem, so if it is possible, please, share with us.