Ship License Number Layout Normalization Based on Regional Texts Fine Localization

Ship license numbers (SLNs) are frequently used to identify ships. An automatic SLNs recognition system makes it convenient and fast to recognize SLNs. A significant prerequisite of many characters recognition algorithms is that characters need to be recognized should be segmented as single ones or at least be listed in one horizontal line. However, SLNs from different ships may have different layouts and the characters of SLNs are often written in more than one lines. In order to recognize these multi-layout SLNs, it is momentous to normalize then into one line. In this paper, we present a SLN layout normalization framework based on regional texts fine localization. For an input SLN image, a series of pre-processing operations are first conducted to enhance the quality of the input SLN image. Then, a single shot text detector with regional attention is used to fine detect characters in that SLN. The normalization algorithms are finally proceeded to normalize the input SLN image based on the detected text bounding boxes and the layout priors of SLNs. At last, the proposed framework is tested on 3216 SLN images, the proposed framework is proved to be effective with a normalization rate of 87.38%.