Text localization using valid optical flow for recognition of slab numbers

The goal of this study is to develop a new algorithm to determine the placement of text in images using two consecutive frames. The algorithm consists of two steps: detection of top and bottom boundaries (TBB) and determination of left and right boundaries (LRB) of the texts. For accurate detection of the LRB, several algorithms are proposed in this paper. First, morphological and geometrical information of the letters is used to identify TBB. Next, coarse LRB are determined using an optical flow, which is rendered by the combined methods of Lucas/Kanade and Horn/Schunck algorithms in video sequences. Finally, the texts are localized from the image based on optical flow field information. Experimental results of the text localization have 95.7% recall. The algorithm is developed in a module and can hence be embedded in a steel product management system easily.