TIE- text information extraction from natural scene images using SVM

Abstract

Detection of text and its localisation from natural scenic imagery plays a vital role in in Content Based Imagery Analysis. The major and unavoidable challenges are illumination effects, different orientations in the lines, typical backgrounds, different font styles and sizes. This paper presents a novel methodology Text Information Extraction (TIE) using Support Vector Machine (SVM) which robustly detects and localises the text from natural scene images. Major focus of this paper lies in detection of text object from the whole scene. Initially the image will be pre-processed for noise removal and contrast enhancement. Later, all the objects of scene will be marked and extracted in order to form an object pool. SVM technique is used to locate text object among the object pool. The SVM will be trained with supervised parameter learning. At last the well trained model will perform binary classification to differentiate between text and non-text object from the object pool. Experiments have been performed on ICDAR dataset and achieved results determine that proposed approach leads to utmost Precision and Recall performance when compared to existing state-of-the-art techniques.

Keywords

• Text detection and localisation
• Pre-processing
• SVM