Methods On Target Recognition Of Sea Surface Based On Ensemble Learning And Broad Learning Neural Network

Unmanned surface vessels hold potential to supplant human involvement in various maritime tasks,enhancing research efficacy and ensuring the safety of maritime personnel.The functionality of these unmanned vessels relies heavily on their autonomous capabilities,particularly in target detection and collision avoidance.As such,camera-based visual obstacle detection,integral to the safe navigation of these vessels,has emerged as a prominent research focus.The camera allows for precise identification of various attributes of obstacles,such as shape,contour,and other specific details.Given the compact size of these unmanned surface vessels,they require supplementary equipment for scientific research and rescue operations.Constrained by space,these vessels possess limited computing power and other resources.Traditional target recognition algorithms utilize deep learning networks,which are resource-intensive in terms of training time,cost,and hardware requirements.To address these constraints,this dissertation aims to enhance the speed and accuracy of maritime surface target recognition,focusing on the following aspects.This thesis proposes a maritime surface target recognition method founded on a Bagging ensemble structure and a broad learning neural network.Initially,a Broad Learning System(BLS)featuring a simple architecture and swift operation is introduced.A dataset of common maritime surface target images is constructed and the concept of Bagging in ensemble learning is employed to design several broad learners for training.Training data is obtained from the training set via Bootstrap sampling.Once training is complete,all broad learners simultaneously predict the test set data,with the Bagging ensemble yielding the final predicted category corresponding to the test image.A K-fold cross-validation method is applied to train the model,mitigating the influence of data volume on model performance.Building upon this,the thesis introduces a target recognition and classification method based on a Stacking ensemble structure and a broad neural network.Inspired by the Stacking integration concept,Resnet networks serve as the foundational model to train the image dataset and generate the predicted values on the test set.Multiple sets of predicted values are produced through K-fold cross-validation and utilized as meta-features.The broad neural network functions as the meta-model,training with the predicted output of the Resnet network as input.Following this,a test set is used to examine the ensemble and generate test results,enabling a verification of algorithm performance.The proposed algorithm is subsequently integrated into an unmanned vessel platform and its performance is assessed in real maritime environments.Experimental results demonstrate that the proposed algorithm,trained and validated on both custom-built and universal datasets,outperforms established algorithms such as the single broad learning algorithm,Resnet network,CNN classification algorithm,and Densenet121.The ensemble algorithms developed in this thesis significantly improve the accuracy of target classification and recognition while reducing training time.Evaluated across metrics including accuracy,recall,F1 value,and the ROC curve,the model proposed in this thesis demonstrates robust performance.In real-world vessel tests,the unmanned vessel equipped with the proposed algorithm effectively identifies existing water surface targets and outputs the classifications and probabilities of these targets.Consequently,this facilitates autonomous collision avoidance decisions,thereby ensuring the safe navigation of unmanned surface vessels.