Dear Gautam, 

my answer seems to be simple. but for your application this is the best thing. 

histogram equalization  may solve the problem up to some degree. 

homomorphic filtering is the best solution for the problem. because in Homomorphic filtering, the image is taken as product of illumination and reflectance. f= i*r. 

the simplest of all solutions could be:

the image you have shown is a monochrome image. you can convert in to binary image with some threshold. it could be 0.5.

the illumination is uniform through out the whole image. if you are not satisfied with 255 background. you can change as well with some less intensity.

i hope your problem got solved with these simple solutions.

Try solving this problem using robust method of edge detection, binarization using local threshold

I would suggest to you to use sauvola binarization, niblack binarization.

To cut things short: I agree with Stafford Michahial:

binarizing, using the local average as the threshold should result in a crisp image where the letters are black and 'the rest' is white.

In principle it should be possible to 'improve' the image by a similar technique in the original domain. But this is not really worth the effort and might 'worsed' the situation for the text recognition stage due to the possible introduction of artefacts. Ad-hoc binarization should work a lot better.

BTW: one of the most important stages of image processing is controlling the illumination: positioning the light source not about opposite to the camera would prevent such overexposure as in the upper right corner - which seems to be caused by some reflection of the lamp on the paper. If the light source had only been aside of the camera ...

Regards

Hi Gautam,

I'd like to make you aware of our new software package MIPAR (http://mipar.us) which is well-suited for these kinds of problems. I have made two recipes:

I think both of these approaches were good suggestions by Stafford and U. Dreher. I have attached screenshots of MIPAR's Image Processor with each of these recipes applied, as well as the recipe files themselves if you wish to download a free trial of MIPAR at http://mipar.us/trial to run them yourself.

We will be adding OCR tools shortly as well, but for now, I think MIPAR could at least be a helpful tool for your pre-processing needs.

Cheers,

John

I believe this is an illumination normalization problem. Try to use histogram equalization or any other efficient illumination normalization method such as wavelet_normalization.

 You can try to model the background as accurate as possible at first. Then subtract the modeled background from the original image. The left image will become very easy to be segmented.  

I think it occurs due to image enhancement.so you perform nonlinear transformation like log transformation/Gama correction and then equalize your image.

You could try thresholding the image. This would lead to an image where text is pure black and the background is pure white (hopefully). The Otsu binarization method does not require parameters, and would be a good first try for this kind of image. It is already implemented in opencv, you would only need to load the image and call a function. For example, if you have an image (cv::Mat)  called tmp:

cv::threshold(tmp, tmp, 0, 255, CV_THRESH_BINARY | CV_THRESH_OTSU);

I agree with Khaled Ahmed Nagaty. Histogram equalization is one of the solutions.

A rudimentary way of homogenizing illumination is to apply gaussian blur to the image and then substract the blurred to the original image. Hope it helps!

Júlia

I reccomed you an histogram equalitazion, for example CLAHE.

Best regards.

Dear Gautam,

from my point of view you have to options: enhance the lightning conditions, which will of course solve your problem immediately or, as I assume, enhancing illumination is not challenging enough or even possible. If your illumination conditions are uniform => they don't vary over time, then you might consider use the "Flat Fielding" algorithm to obtain uniformly illuminated image. This technique is often used to enhance images obtained from microscopes. The process is well described in the book "Practical Algorithms for Image Analysis 2nd Edition" written by O'Gorman, Sammon and Seul. Cambridge University Press.

I hope it helps.

Kindest Regards

Dejan

http://www.cambridge.org/us/academic/subjects/computer-science/computer-graphics-image-processing-and-robotics/practical-algorithms-image-analysis-cd-rom-2nd-edition?format=WW

Dear Gautam, 

my answer seems to be simple. but for your application this is the best thing. 

histogram equalization  may solve the problem up to some degree. 

homomorphic filtering is the best solution for the problem. because in Homomorphic filtering, the image is taken as product of illumination and reflectance. f= i*r. 

the simplest of all solutions could be:

the image you have shown is a monochrome image. you can convert in to binary image with some threshold. it could be 0.5.

the illumination is uniform through out the whole image. if you are not satisfied with 255 background. you can change as well with some less intensity.

i hope your problem got solved with these simple solutions.

Hi Gautam,

You might try the method of fitting the intensity of image, as described in Henry's thesis "Neural Network-Based Face Detection", page 19-20.

http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.187.7787&rep=rep1&type=pdf

Hi,Gautam,

Recently, we  used polynomial fitting to fit the illumination. Maybe you can give it a try.

Thanks a lot for all the fruitful answers