Design And Implementation Of Hull Number Recognition System Based On Deep Learning

As the carrier of maritime activities,the realization of intelligent identification of the ship’s identity has important practical significance.At present,ship target recognition is still at the stage of binary classification of ships and non-ship targets or multi-classification of ship types,and cannot achieve smaller granularity of ship identification.In addition,when the ship’s automatic identification system is not equipped or artificially turned off,it will lead to loopholes in the ship’s identity supervision.The hull number is the ”identity card” of the ship,and the identification of the hull number can realize the identification of the ship at the finest granularity and at the highest level.However,there are still few mature hull number recognition theoretical methods and hull number recognition application systems.Therefore,this paper designs and implements a hull number recognition system based on deep learning for the needs of intelligent identification of ship identities.The main tasks are as follows:(1)Design and implement a hull number recognition method based on deep learning,which is the core method of the hull number recognition system.The hull number recognition method based on deep learning is a two-stage of hull number detection and hull number recognition.In the hull number detection stage,aiming at the problem that the hull number area is relatively small in the image,which makes the hull number detection difficult,a solution is proposed to add a feature layer with a downsampling multiple of 4in the YOLOv5 network for fusion to improve feature extraction capabilities; In the stage of hull number recognition,aiming at the problem that the hull number area is blurred and distorted,which leads to the difficulty of hull number recognition,a super-resolution correction network solution with both super-resolution and correction functions is proposed;In order to solve the problem of low accuracy of hull number recognition by convolutional recurrent neural network,a solution is proposed to introduce a multi-channel feature fusion structure in the feature extraction network to extract more abundant hull number features.The construction of the hull number detection and hull number recognition data set is completed,and the experimental analysis and verification are carried out based on the data set.The experimental results show that the accuracy rate of hull number detection in the hull number detection stage reached 92.4%,and the recall rate reached 89.4%; the hull number recognition accuracy rate in the hull number recognition stage reached 91.3%;the hull number based on deep learning The overall recognition accuracy of the recognition method reaches 89.2%,and the processing speed reaches 12.4FPS,which meets the accuracy and real-time requirements of the ship number recognition system.(2)Desigen a hull number recognition system based on deep learning.Firstly,the demand analysis of the sponson identification is carried out,and the overall design of the sponson identification system is given.Then,the design work of sub-modules is carried out from three aspects: video acquisition function module,sponson intelligent identification function module and data storage function module.Finally,the overall workflow of the sponson identification system is given.(3)Develop a hull number recognition system based on an embedded platform.This system is developed using various technologies,such as Python,TensorRT,PyQt,and Pytorch,and is based on the NVIDIA developer kit.It combines the video capture function module,ship number intelligent recognition function module,and data storage function module,achieving a lightweight application of the system.