Towards end-to-end container code recognition

Container code recognition can improve the efficiency and economy of the management system in the port. However, the task is different and complex due to the degradation of image quality caused by uneven illumination, background variation, smear, inaccurate character extraction, and so on. Current processing methods on container images usually provide the framework or modules on specific tasks, such as region detection and character classification, which are hard to implement or to be combined into a whole process. In this paper, we propose a fast end-to-end method of automatic recognition of container code that fills the gap by locating the region and detecting characters as well as making the classification. This allows the three tasks to work collaboratively by pipeline, which is critical to identify the container code. For evaluation, we collect around six thousand container images, including all kinds of circumstances from the local port. Compared with a few other methods and two-step approaches consisting of state-of-the-art character detector and character classifier, our system achieves some competitive results. Finally, the proposed system is verified on this dataset and the overall accuracy reaches 97.30%.