Identifying Binarization method for scanned low quality documents

Telugu Scanned documents and books that are available in net have varying degree of brightness and sometimes very low in quality. Even a human can not identify words in such documents. In pattern recoganization, binarization of a document is an involved process and this article only considers documents that vary in degree of brightness.

It seems in Indian OCRs people are exploited the following concepts from English OCR techniques (with little modification)

1) Histogram based : This is useless for Indian languages
2) Clustering based : K-means and Gaussian density methods
3) Entropy based :
4) Attribute based : based on predefined attributes


Otsu’s algorithm : (complete code and test it)


Let’s see Otsu’s algorithm and how to improve it for better results.


As usual, like a human, when the background and foreground intensities are well separated Otsu’s algorithm gives a better binarized results. In the following picture, at certain parts, we can clearly see the pixels of gray scales of foreground and background are mixed. In such cases otsu’s algorithm gives a wrong result and we certainly need a new method.






A simple approach is divide a document into multiple parts ( as a grid ) and find out threshold value. Use image length and width pixels to make a grid.





Feature 1 : T(otsu) r – T(min) r
Feature 2 : u®
Feature 3 : sigma(t)





Once we extract these 3 features for each grid ( region ) use SVM to determine which binarization action to take. Improve this decision using the training data.