Text Localization in Natural Scene Images Using Extreme Learning Machine

Text localization from natural scenes is a challenging task and has many applications in the modern era of autonomous systems, IOT and so on. It involves many challenges such as variation in shape and sizes of texts, variable illumination, occlusions, variable orientations, and a huge number of non-text objects in nature which has a form similar to textual elements. In our proposed approach we have introduced the use of ELM in this domain based on histogram of oriented gradients and entropy based features. Along with MSER based candidate region selection, filtering false positives with stroke width transform, region merging with morphological operations, a complete system is proposed and tested on the ICDAR 2015 dataset. We achieved with a precision of 94.02%, recall of 95.39% and F-measure of 94.70%. The proposed algorithm surpasses many other approaches on the dataset. The ELM classifier performs better than other common classifiers like the SVM, Bayes or KNN. It is also shown to be more statistically significant than strong classifiers like SVM.