DELIGHT-Net: DEep and LIGHTweight network to segment Indian text at word level from wild scenic images

The recognition and detection of multioriented text from textual natural scene images are still challenging in the computer vision community. The segmentation on either word level or character level is a vital step in the entire end-to-end performance of the scene text recognition system. Many academicians and researchers have done work in the prominent field of segmenting the words or characters from complex document images as well as handwritten images for various non-Indian scripts. In this paper, we extensively presented a deep learning-based architecture named DELIGHT-Net which is derived from the general UNet architecture to segment the text at the word level from natural scene images. The method is mainly proposed to segment the Devanagari, Gurumukhi, and English scenic words from complete images collected from day-to-day life. To achieve this, we have introduced a new dataset, i.e., National Institute of Technology Jalandhar-Word Segmentation (NITJ-WS) which has around 2200 text blocks extracted from 1500 natural images containing unilingual, bilingual, and trilingual text. The benchmark comparative assessment of our dataset is performed with the proposed model and two state-of-the-art models, i.e., UNet and ResUNet. Statistical and visual results are evaluated using different evaluation parameters, which depict the efficiency of the proposed model. Some possible future directions are also recommended in the manuscript. We hope that our work is a stepping stone for academicians in the field of natural scene text recognition.