23-02-2013, 10:59 AM
A HYBRID APPROACH TO DETECT COLOR TEXTS IN NATURAL SCENE IMAGES
[attachment=51561]
Abstract
Large amounts of information are embedded in natural scenes which are often required to be automatically recognized and processed. This requires automatic detection, segmentation and recognition of visual text entities in natural scene images. In this paper, we present a hybrid approach to detect color texts in natural scene images. The approaches used in this project are region based and connected component based approach. A text region detector is designed to estimate the probabilities of text position and scale, which helps to segment candidate text components with an efficient local binarization algorithm. To combine unary component properties and binary contextual component relationships, a conditional random field (CRF) model with supervised parameter learning is proposed. Finally, text components are grouped into text lines/words with a learning-based energy minimization method. In our proposed system, a selective metric-based clustering is used to extract textual information in real-world images, thus enabling the processing of character segmentation into individual components to increase final recognition rates. This project is evaluated on natural scene image dataset.
[attachment=51561]
Abstract
Large amounts of information are embedded in natural scenes which are often required to be automatically recognized and processed. This requires automatic detection, segmentation and recognition of visual text entities in natural scene images. In this paper, we present a hybrid approach to detect color texts in natural scene images. The approaches used in this project are region based and connected component based approach. A text region detector is designed to estimate the probabilities of text position and scale, which helps to segment candidate text components with an efficient local binarization algorithm. To combine unary component properties and binary contextual component relationships, a conditional random field (CRF) model with supervised parameter learning is proposed. Finally, text components are grouped into text lines/words with a learning-based energy minimization method. In our proposed system, a selective metric-based clustering is used to extract textual information in real-world images, thus enabling the processing of character segmentation into individual components to increase final recognition rates. This project is evaluated on natural scene image dataset.