Vessel Target Detection Based On Visible Image

With the development of Artificial Intelligence(AI)related research and devices,AI is beginning to be used on a large scale in different fields,and this is also the case in the marine environment.The marine domain can achieve better automation and intelligence through the involvement of AI,and at the heart of this is the object detection algorithm.However,the marine environment has its unique characteristics,and it has more complex data features than the urban and living environments where current object detection algorithms are commonly used.Thus,current mainstream object detection algorithms can not work better in the marine environment.Current object detection algorithms in the marine environment have the following difficulties:(1)The marine environment is more complex and less,and this caused training the model well to be difficult.(2)Current object detection models are not designed for complex environments,and they cannot be adapted to complex environments well such as the ocean.To address the above problems,this thesis improves the current mainstream object detection algorithms based on the needs of the marine environment,as follows.1.A marine environment dataset is constructed from videos posted on the Internet taken near the shore and harbor.The videos are randomly captured and annotated,and divided into a training set,a validation set,and a test set in the ratio of 6:2:2.2.Based on the analysis of the constructed dataset and the real situation of the marine environment,three characteristics of the marine environment were derived:(1)the data characteristics of the daytime,evening,and nighttime differed greatly.(2)The environment is empty and has a lot of water vapor,compared to urban and other environments that are affected by fog.(3)The marine environment has many rainy days and raindrops tend to form diagonal lines in the image thus changing the image features.Based on these three features that affect the performance of the object detection model,data enhancement algorithms such as darkening enhancement,fog enhancement,and rain enhancement are proposed.3.This thesis analyses the factors that prevent mainstream object detection models from adapting to complex environments,and analyses current research on coping with complex environments.Then proposing a parallel backbone network based on the functions of different parts of the object detection model.The model uses multiple backbone networks to extract features from complex environment data,calculates the weights of different branches,and sums the extracted features as a weight for prediction.Experimental results show that the data enhancement algorithm proposed in this thesis can significantly improve the performance of the model in marine environmental data.The processing capability of complex marine data is further enhanced by using a parallel backbone network.