| The detection and identification of fishing boat license plate is one of the important technical means to improve the scientific and intelligent management of fishing ports,and is of great significance to solve the time-consuming and labor-intensive problems existing in the traditional way of manual supervision of fishing boats.At present,there are still various problems in the application of fishing boat license plate detection and recognition,such as the small amount of data related to the boat license plate,inconsistent location of the boat license plate hanging,inconsistent background color and number of characters,and poor quality of the boat license plate picture,which largely affect the improvement of the accuracy of fishing boat license plate detection and recognition.With the development of artificial intelligence and the continuous progress of deep learning technology,the problems faced by fishing boat license plate detection and recognition are slowly being solved.Deep learning has strong learning ability and generalization ability,which can improve the accuracy and robustness of fishing boat number detection and recognition in surveillance video.In this paper,by studying the text detection network algorithm DBNet and text recognition algorithm CRNN,we have completed the production of fishing boat number detection dataset and fishing boat number text recognition dataset,and the improvement of fishing boat number detection and recognition algorithm.The main research contents of this paper are as follows:(1)A Differentiable Binarization network with Dual Path Networks and convolutional block attention module(DP-CBAM-DBNet)is proposed to fuse Dual Path Networks and convolutional block attention module,which adds the convolutional block attention module(CBAM)to the Dual Path Networks(DPN)and replaces the original feature extraction part of the DBNet network with the improved feature extraction layer for feature detection.The improved feature extraction layer is used to obtain more detailed features of fishing boat license plate,thus improving the overall detection accuracy of fishing boat license plate text.Aiming at the problems of irregular position changes,uneven color backgrounds and unclear boundaries of fishing boat license plate images in natural scenes,this paper researches and improves the DBNet network to achieve the detection of fishing boat license plate images based on the above problem scenarios and improve the accuracy rate of fishing boat license plate text detection.the DP-CBAM-DBNet model and the original network use Res Net18 and Res Net50 networks,there is a 2.54% and 1.49% improvement in the detection accuracy,and the number of model parameters only increases by 1M compared to Res Net18 and decreases by 11.9M compared to Res Net50,thus it can be concluded that the DPCBAM-DBNet model has a better It can be concluded that the DP-CBAM-DBNet model has a better performance in detecting fishing vessel license numbers.(2)A Convolutional Recurrent Neural Network with Multi-head-attention Mechanism(MHA-CRNN)is proposed for fishing boat license number text recognition.The first method is to optimize the convolutional layer,replacing the VGG16 network with high performance Res Net network,which consumes more computational resources,and updating the Re LU activation function in the Res Net network to Mish activation function,so that more data can be added to the training of the network,thus making the model have better accuracy and generalization ability;Secondly,the recurrent layer is optimized,and the Multi-head-attention Mechanism(MHA)is incorporated in the recurrent layer to obtain the dependencies between long text characters more effectively,and the features of different dimensions of the fishing boat number can be obtained.For the problems of varying length of fishing boat license plate and unclear image of fishing boat license plate,this paper researches and improves on the basis of CRNN network to improve the accuracy of fishing boat license plate recognition.the MHA-CRNN network has a comprehensive improvement of 11.53% in accuracy compared with the original CRNN network,so the model has a better recognition ability in fishing boat license plate text recognition.(3)A fishing boat license plate number detection and recognition system is designed and implemented.A two-stage dual-model fishing vessel license plate detection and recognition method is implemented using the improved DP-CBAMDBNet model for fishing vessel license plate text detection and the improved MHACRNN model for fishing vessel license plate text recognition.Based on the high point video monitoring data around the fishing port,it realizes the dynamic visualization and intelligent supervision of fishing boats for the phenomenon of fishing boats illegally fishing at sea during the closed season and fishing boats not docking at the specified home port,while the lag of sea supervision is effectively solved,which can effectively investigate and deal with the illegal fishing boats,assist the relevant management departments in the scientific and intelligent management of the fishing port,and reduce the relevant departments’ various resources. |
