The problem of scene text recognition has gained significant importance because of its numerous applications. A variety of methods has been recently proposed that explore various theoretical and practical aspects to solve this problem. In this work, we focus towards a framework to recognize the text present in outdoor scene images. The text information carries one important property, that is, its colour in comparison to its background. Text information is always placed in such a way that it stands out from its background. In the same way, most of the time the characters in a word possess similar colour that helps us to recognize the letters of a particular word. We exploit this characteristic of text regions to solve the problem of character recognition. The character recognition pipeline is further extended in to a word recognition framework where the estimated word combinations are matched against a lexicon.The existing approaches for scene text recognition can be roughly divided in to two broad categories: Region grouping based methods and object recognition based methods. In this work, we have combined region grouping method with object recognition based strategy to achieve the advantages of both techniques. First, we binarize the image using colour information and perform foreground segmentation to separate characters from background. Next, we extract shape representation features on binary images and perform character classification using a pre-trained classifier. The recognized characters form words that are fed in to a string similarity matching stage where lexicon based search is performed to find the closest matching word.Character Identification: We use the bilateral regression [2] for character identification. However, our approach is different than the original method in that we only use it to estimate the horizontal location of each character in word image. The bilateral regression models the foreground pixels by using a weighted regression that assigns weight to each pixel according to its location with respect to foreground in feature space. The pixels that belong to the foreground get high weights in comparison to the pixels belonging to background. In this case, the regression model in equation 1 represents the quadratic surface that best models the image as a function of pixel locations.We enhance the operation of bilateral regression by a pre-processing step where the foreground colour is estimated a priori. We apply n-level colour quantization to achieve binary image for each quantization level. We use Minimum Variance Quantization (MVQ) originally proposed by Heckbert [3]. We quantize each word image in to three colours and analyse the respective binary maps for three quantization levels to estimate the foreground. The characters are cropped from the actual word images using the estimated horizontal location and width from bilateral regression while the height is kept same as the height of the actual word image. the segmented masks are used to crop the characters from original (coloured) im...