Research On The Recognition Method Of Fishing Boat Fishing Behavior Based On Deep Learning

Fishing boats are the essential equipment of the fishery economy and an important part of sea traffic.Fishing boats are small in size and large in number,so fishing boat management is an essential and challenging point in ship management.In addition,the emergence of problems such as illegal fishing and irregular fishing are also spurring the standardization of fishing vessel management.Accurate and efficient fishing behavior identification of fishing vessels can provide technical support for fishing vessel management.Many scholars at home and abroad have used information such as the position,speed,and course of fishing boats to complete the behavior recognition of fishing boats by establishing mathematical models.However,it is difficult to identify the fishing behavior of complex fishing boats based on ship position data.The fishing behavior of fishing boats can only be roughly divided into 2-3 categories.The development of electronic monitoring system for fishing boats provides a new idea for identifying fishing behaviors of fishing boats.This paper aims to accomplish the following two tasks:(1)Based on the electronic monitoring system of fishing vessels,the theory of fishing behavior of fishing vessels is proposed,and the concept,classification,and general identification process of fishing behavior of fishing vessels are summarized.(2)Complete the identification of the fishing behavior of different fishing boats based on this theory.Different fishing boats have different fishing behaviors.According to the scope of application of fishing boats,fishing behaviors can be divided into general behaviors and special behaviors.Among them,common behaviors include stopping,sailing,fishing,etc.Special behaviors are relatively complex,and different ship types have different special behaviors,such as net deployment,harvesting catches,net gear collection,anchoring,etc.,which belong to the special behaviors of stretch net fishing vessels.This paper proposes a general process for identifying fishing behaviors of fishing boats: firstly,the hardware is installed to collect data.Secondly,the classification of fishing behaviors of fishing boats and data processing are completed according to the actual situation.Finally,the network model is designed,and the results of the model are verified.Within the framework of the above theory,this paper researches the identification of fishing behaviors of three fishing vessels: tuna longline fishing,Chinese hairy shrimp stretching net,and Japanese mackerel light netting.The three types of fishing vessels have different characteristics.Among them,the tuna longline fishing vessel has a more prominent target(tuna);The fishing behavior is complex and can provide less computing power.The operation process of the tuna longline fishing vessel is relatively simple,and the target in the operation process is rather large.The mature target detection model can be applied to the fishing behavior recognition of fishing boats in tuna longline fishing.According to the operating characteristics of the tuna longline fishing vessel HNY722,the fishing behavior of the fishing vessel is divided into three categories: "fishing for tuna","fishing for floats" and "others".Or the15578 keyframes of tuna,divide all keyframes and their marked files into 14178 training data and 1400 verification data,and design group training experiments based on four kinds of YOLOV5 deep learning neural network models,including YOLOV5 s,YOLOV5l,YOLOV5 m,and YOLOV5 x Compare training effects.Finally,the average detection accuracy m AP of the three behaviors reached 99.16%.There is a big gap between Acetes chinensis fishing boats and tuna fishing boats: the behavior of Acetes chinensis fishing boats is more complex and difficult to distinguish.First,an electronic monitoring system will be installed on fishing boats with quotas for hairy shrimp to collect work data from June 16,2021,to July 13,2021.Then the working behavior of the fishing boat is divided into five behaviors: stopping,sailing,pulling nets,waiting,and putting nets.Acetes3 DNet,a 3D convolutional neural network,is designed to extract multi-dimensional and multi-level data features and train on the training set.Finally,the training effect is verified in the verification set,and the training results are combined with the AIS ship position data to restore the working process of the fishing boat.The recognition accuracy of the five behaviors in the test set is as high as 97.09%.The multiple behaviors of Scomber japonicus fishing boats also have strong similarities.Still,Scomber japonicus fishing boats have high requirements for real-time recognition of fishing boats,and the model requires a lower computational load.According to the characteristics of Scomber japonicus fishing boats,the operation process is divided into nine kinds of behaviors: "nets pulling","nets putting","fish picking",and "reprinting".In the pre-experiment,four types of networks with different convolutional layers are designed,and each network’s feasibility in recognizing the fishing behavior of fishing boats is observed.The pre-experiments are optimized from the perspective of the data set and network.In terms of data,the size of the optimized data set is significantly reduced while retaining the characteristics of the original data as much as possible.In terms of network,different combinations of pooling layer,long short-term memory network,and attention(including CBAM and SE)are added to the network,and their impact on training time and recognition effect is compared.The combination of LSTM and SE modules has the most apparent optimization effect on the network,and the F1 score of the optimized model on the test set can reach 97.12%.Its F1 score in the test set surpasses classic Alex Net,Vgg Net,and Res Net.Fishing behavior recognition of fishing boats based on the convolutional neural network has strong feasibility and application prospects and can provide technical support for intelligent fishing boat management.In summary,this paper includes the following innovations:(1)Put forward the classification theory of the fishing behavior of fishing vessels.This paper divides the fishing behavior of fishing vessels into general behavior and special behavior.The general behavior has been thoroughly studied,while the research on the special behavior of fishing vessels is less.(2)Propose the general research process on fishing behavior identification of fishing vessels.This paper proposes a universal fishing behavior recognition scheme for fishing boats,which is used to identify the general and special behaviors of fishing boats.Based on surveillance video data,fishing boat recognition has fully verified the behavior recognition scheme.(3)Complete the behavior recognition of the three groups of fishing boats.This paper completes the behavior recognition of three groups of fishing boats,including longline tuna fishing,Acetes chinensis stretching net,and Scomber japonicus lighting net.The recognized behaviors include general behaviors and special behaviors.According to the data characteristics of the three groups of fishing boats,different deep-learning convolutional neural network models were designed for their behavior recognition.