Transfer Learning based Kannada Text Detection and Recognition in Natural Scene Images

Abstract

With the ascent in deep learning, computer vision problems have taken center stage among the research community. Text recognition from natural images is an extremely complex task and entails detecting and locating text inside an image by producing the co-ordinates of a bounding box that holds it. You Only Look Once (YOLO) algorithm is proven for object detection by creating a bounding box around a detected object and providing its location. And Transfer learning in Deep Learning has been gaining momentum for complex text and speech detection tasks. Taking cue from this, in this paper we implement the YOLO algorithm on AksharaNet, a convolution neural network (CNN) based classification model and train it on a large Kannada Scene Individual Character (KSIC) dataset. On testing the pre-trained network with the South Indian language dataset, we observe that it classifies Kannada characters with an impressive testing accuracy of 90.17% with an area under the curve (AUC) of 0.932. On comparing the proposed model with existing models, it is observed that pre-trained AksharaNet with YOLO implementation returns the best precision, recall and F1-scores of 88%, 90% and 93.83% respectively, which is greater than 8%, 4% and 5% on an average compared to other models.