Faster Container Code Detection and Recognition

This paper develops a bi-fusion feature cut method for detecting and recognizing container codes from shipping ports. The container code analysis is basically one application of text analysis, which can be divided into one-stage or twostage categories. The one-stage method is mainly based on heap maps of characters, which is anchor-free but time-consuming and fails to deal with small characters. The two-stage method enlarges each detected text region and thus is accurate for small character recognition, but very inefficient since its time complexity is proportional to the number of text candidates. To alleviate the above issues, we propose a bi-fusion feature cut method, which is efficient since the backbone for text detection and recognition is only applied once, no matter how many text regions are generated. The proposed method also increases the detection accuracy since the bi-fusion module fuses feature maps not only from the top-down direction but also from the bottomup direction to generate richer features for small character recognition. The proposed system is robust under different challenging conditions, such as weathers, light reflections from ground, and container rusting. Compared with SoTA (State of The Art) methods, the system is superior in terms of efficiency, accuracy, and robustness.