Research On Ship License Detection And Recognition Based On Knowledge Transfer

China is rich in inland water transportation resources,and the construction of a complete modern inland water transportation safety supervision system is an inevitable requirement to accelerate the development of inland water transportation in our country.In recent years,with the rapid development of deep learning,the accuracy of face recognition,license plate recognition and other technologies have been greatly improved.More and more industries have begun to widely use digital security and management technologies.Water transportation also needs to be combined with algorithms such as ship license detection and ship license recognition to realize the informatization.Since deep learning is a data-driven algorithm,a large amount of data is required to train a model with acceptable effect,and the ship license image is more difficult to obtain than the license plate image,which brings difficulties and obstacles to the research of related algorithms.At the same time,ship license recognition belongs to a vertical category of text recognition,and its distribution has a large deviation from the distribution of large public datasets or synthetic datasets,which leads to the fact that it is impossible to obtain acceptable recognition effect by directly using datasets from other fields for auxiliary training.In response to the above problems,this study mainly made the following contributions.Firstly,this study collects real ship photos with cameras deployed on the river in Foshan,Guangzhou,and annotates the location,content,and attributes(such as blur degree,shooting angle)of the ship license plates in the images.This study constructs a relatively complete dataset of ship license detection and recognition and is available for the researchers to make up for the lack of data in the field of ship license recognition.Secondly,this study combines the existing knowledge transfer methods with the scene of ship license recognition and optimizes it according to the characteristics of text data.This study proposes a domain adaptation method of coarse and fine-grained feature alignment and character-level pseudo-labels.The transfer of knowledge from data of other fields to the ship license data makes more effective use of data information and improves the accuracy of ship license recognition.Thirdly,this study combines the two knowledge transfer methods of meta-learning and multi-domain adaptation and proposes a new training mode and corresponding multi-domain text recognition dataset,which has achieved better performance in the ship license recognition task.In many cases,it is difficult and costly to obtain and label the data in specific scenarios.Therefore,this study has important significance and application value for such text recognition scenarios with insufficient data.