Scene text detection with high performance computing

Abstract

Text detection and recognition in natural images find many real time applications for navigation, multimedia, industrial automation, and scene understanding etc., The Text obtained from natural images is difficult to extract because they are of different fonts, sizes, shapes, complex background, sensitivity and Interference. This work proposes a novel Text detection method, which is an important step in End to End Text recognition. There are four main steps in this work, character candidate region extraction, character region classification, Text candidate extraction and Text region classification. The extraction of proper character candidates is vital for any good text detection system. A novel method of MSER and Fast stroke feature transform is introduced to extract character candidate regions. The maximal extremal regions are deducted from the input image. The resulting image is smoothened and segmented into super pixels. The super pixels are then clustered by a Hierarchical clustering algorithm. The background regions are removed from the image by a random forest edge detection algorithm. A stroke edge and a stroke color map are generated based on which character candidate regions are extracted. A Deep convnet Text region classifier is introduced to classify the character candidate regions. Low level features based on color, texture and geometric properties are extracted and fused together by a neural network. High level features are extracted by a deep convolutional neural network. High level features and the fused Low level features are combined together by another neural network. For the classification of character and non character regions, a Random Forest regressor is used. The Resultant Character regions are combined to extract candidate Text regions. The candidate Text regions are classified as Text and non text regions by the Deep convnet region classifier used for the Character newline