Script Identification in Natural Scene Text Images by Learning Local and Global Features on Inception Net.

In a multi script environment Script identification is essential prior to text recognition. Compared to document images, Script identification in natural scene images becomes a more challenging task due to complex backgrounds, intricate font styles, poor image quality etc. All this is in addition to the common problem of script recognition, related to similar layouts of characters found for certain scripts. The proposed work involves fine tuning of a pretrained model of Inception net V3 on a combination of word and the constituent character images extracted from the sample images of the scene text. In this model, the relu activation function originally used with Inception net is replaced by an experimentally selected Mish activation function and a softmax layer is added at the output to have a probability distribution for possible scripts of the input word images. To identify the script of an input word image, maximum of the sums of the weighted probability values for the word image and the images of its constituent characters is considered for all the script categories. The proposed method is tested on MLe2e and CVSI-2015 datasets showing encouraging results by slightly crossing the existing benchmarks.