Few Shot Multiple-Font Character Recognition for Ship Monitoring

To effectively manage ships and maintain the safety of port and territorial waters, ship plate recognition is an essential technology. However, there are many different font styles in actual scenes because there is no unified format. Among them, the handwritten font is the most changeable. These complex and changeable font styles will cause difficulties in recognizing ship plates. Furthermore, handwritten ship plates are unique in data collection, which means that it is impossible to collect enough fonts of the same style or all ship plates for training. In this paper, we propose a text recognition model architecture that simulates human learning and literacy to solve the problem of few-shot multifont, which is called learning by analysis (LBA). Humans can recognize multiple types of character through pre-train knowledge. Referring to this concept, LBA is a twin network composed of a benchmark model (BM) and an extended model (EM). BM builds a hypothesis space based on standard fonts, and then EM learns to recognize variable text based on BM through high-dimensional feature mapping and aggregation of embedded spaces. In addition, we also propose a type change block without training, which increases the complexity of the data by making complex type changes to the text. Experiments show that the method achieves 96% accuracy on NIST. The accuracy of ship plate recognition in natural scenes is as high as 91%, which shows that our method has a robust generalizability.