Text localization from scene images

Abstract

Text in scene images is a useful source for many applications such as hand-held/wearable devices for visual impaired people to see text around, mobile robot navigation, scene understanding. Locating text in scene images is beyond the capabilities of current optical character recognition packages. Text localization envisages several difficulties including geometric variations, photometric variations, environmental variations and ambiguities. This thesis investigates methods to build an efficient text locating machine in both aspects: computational complexity and accuracy. The proposed system has a cascade framework which includes three main stages: detection, verification and text line generation. In the detection stage, recall is essentially kept high while precision may be lost. The detection stage is deployed with simple and fast operators: edge detection, morphologic operator and connected component analysis. The verification stage is designed to compensate the precision. The verification stage utilizes feature extraction and classification machines to classify potential text areas resulting from detection stage. The verification stage is rather complex and time consuming in comparing with the detection stage. However the cascade structure allows the system save processing time because the detection stage already zoomed into potential text areas and removed many other non-text areas. The proposed system has been tested on the benchmark dataset with different types of features and two classification machines: Support Vector Machines and AdaBoost. The comparative study is drawn to find the suitable features and classifier for scene text locating problem. The results are also compared with recently published results and it shows that the proposed system improves the recall significantly.