Research On Deep Learning Methods For Deep Sea Garbage Classification And Detectio

Deep-sea debris refers to those wastes that remain in the depths of the ocean all year round.Deep-sea garbage is difficult to clean up,and the pollution caused by it is extremely serious.At present,the deep-sea garbage is mainly cleaned by manual operation of underwater robotic arms,which is not efficient.One way to significantly improve the efficiency of garbage cleaning is to load deep-sea debris classification and detection systems into autonomous underwater vehicles(AUVs)to achieve autonomous cleaning of AUVs.However,there are few datasets and algorithms for deep-sea debris classification and detection.To promote the efficiency of AUV’s autonomous classification,detection,and cleaning of seabed garbage and improve the marine environment,this dissertation is dedicated to creating deep-sea debris classification and detection datasets and algorithms.The research in this thesis is as follows:(1)Build a deep-sea debris classification dataset and detection dataset.Based on the Deep-sea Debris Database provided online by the Japan Agency for Marine-Earth Science and Technology(JAMSTEC),this thesis generates a real deep-sea debris classification dataset(Deep-sea Debris Images dataset,DDI dataset)and detection dataset(Deep-sea Debris Detection dataset,3-D dataset).Each dataset contains a total of 7 types of debris:metal,glass,plastic,rubber,fishing net &rope,natural debris,and cloth.(2)Considering the intra-class diversity and inter-class similarity of deep-sea debris,this thesis constructs a deep-sea debris classification model called Shuffle-Xception.First,Shuffle-Xception uses depthwise separable convolution to extract more information and advanced features of seabed debris.Second,group convolution is taken to improve classification efficiency and model representation ability;at the same time,the channel shuffle strategy is employed to mine channel information between feature groups.Finally,the residual connection is adopted to make the network deeper and easier to learn.Experimental results show that this network is better than the current advanced convolutional neural network methods and has the potential for deep-sea debris classification.(3)To improve the speed and accuracy of deep-sea debris detection,this dissertation proposes a one-stage deep-sea debris detection network ResNet50-YOLOV3.In this detection network,a multi-scale detector called YOLOV3(You Only Look Once)is applied,which can achieve a high detection speed while ensuring the accuracy of detection.ResNet50(Residual Network)is the backbone of this network.It has a strong feature abstraction ability,which can further improve detection accuracy.In this dissertation,the performance of ResNet50-YOLOV3 is verified by experiments.The experimental results show the applicability and effectiveness of ResNet50-YOLOV3 in deep-sea debris detection.