Although I am new to the computer vision domain, I am tasked with automating handwritten text recognition from bank cheques. I first did some preprocessing(blurring and binarization) and then dilated it so that I could segment the words. But since the words are handwritten (cursive and unconstrained), the segmentation is failing most of the time. I read up on some research papers and it seems you can use neural networks for segmentation.Is that the correct approach? Are there other alternatives? Can someone shed some more light?
As mentioned in comments there is no one right way to recognize handwriting. That said, Here are two techniques you could explore.
Watershed segmentation works by creating a topological map of the image to determine which objects distinct and with or without connection to other objects. I recommend it to you because watershed excels at naturally or man-made round objects. It has been documented for use in handwriting recognition. This is an easy one to get started with too. You can simply call watershed in matlab, OpenCV, or scikit
Machine Learning / Deep Learning (ML) is a daunting term to some but we usually get the chance to stand on the shoulders of giants such that ML on images is about as difficult as editing an large collection of vacation photos. Bag of words is a particular technique which works well on finite object sets. Note that like most ML BoW requires pre-classified training materials so be prepared to provide at least 100 different samples of a cursive uppercase 'Q' along with all the other characters you wish to recognize. Nonetheless, if you've got the time and resources to train it, a ML solution can be incredibly powerful in handwriting recognition.