Text Understandingform Natural Images with Enhanced Classiftcation using Genetic Algorithm
Ghulam Jillani Ansari, Jamal Hussain Shah, Muhammad Sharif · 2021
In recent times, text retrieval as content from natural images is regarded as challenging and popular in computer vision. This paper illustrates understanding of natural scene text, having unconstrained scene properties in the vein of unstructured background, variant font characteristics, undesirable illumination and much more. These properties make scene text challenging, which encourages us to investigate this research area. To perform natural scene text understanding following steps have been carried out. Firstly, text extraction is done, which include contrast enhancement using L-U-V channel, Maximally Stable Extremal Region (MSER) for text region detection and Stroke Width Transform (SWT) for obtaining connected components and segmentation. After that, novel binary classification approach for classifying text and non text regions using genetic algorithm with integrated Support Vector Machine (SVM). Secondly, character based text recognition and labeling is employed using novel Convolution Neural Network (CNN) model. Finally, the CNN outcome is collected into the text file to examine that it matches to scene text or not. In order to observe the significance of the intended methodology, standard datasets SVT, IIIT5K and ICDAR 2003 have been used. After experimentations, it has been found that proposed methodology performs well while comparing with existing benchmark techniques in terms of accuracy.