Deep Learning-based Underwater Target Detection On The Seabed Research

At present,underwater target detection is an important technology for exploring the ocean,and China has made important plans to "further care for the ocean,understand the ocean,and manage the ocean",emphasizing the strategy of strengthening the ocean and strengthening deep-sea research.In the marine pastoral industry,target detection technology can assist underwater robots to complete underwater biological fishing,improve fishing efficiency.In terms of national defense security,underwater target detection technology can monitor the seabed in real time to ensure the security of our waters.Existing target detection algorithms can achieve good results for common detection tasks,but the underwater scenes are complex,mainly in terms of serious image degradation,low contrast,target overlap and image blurring,and the current underwater data sets are difficult to collect on a large scale,so the use of existing detection algorithms can not achieve better detection accuracy.To address the above problems,this topic is based on deep convolutional neural network to design and learn a detection model suitable for underwater scenes.The specific work is as follows:(1)Research on underwater target detection based on improved YOLOv3-SPPAn improved YOLOv3-SPP underwater target detection algorithm is proposed for the underwater target detection task with blurred images,complex backgrounds and small targets leading to false and missed detections.Firstly,the URPC underwater dataset is filtered and also re-labelled using Label Img,followed by the recovery of the original underwater images using the UWGAN network and the enhancement of the data using the Mixup method to reduce false label memory;secondly,the YOLOv3-SPP network structure is used as the basis to increase the network prediction scale and improve the small target detection performance;immediately afterwards,the CIo U border is introduced regression loss to improve localization accuracy;and finally,the K-Means++ clustering algorithm is used to filter the best Anchor box.(2)Research on YOLOv5 underwater target detection based on attention mechanismFor the underwater target detection task,the existing algorithms perform inadequately in the feature extraction and fusion stages,while the network parameters are large and difficult to deploy to relevant devices,an attention mechanism-based YOLOv5 detection algorithm is proposed.Firstly,the CA attention mechanism is fused in the feature extraction backbone network to enhance the network feature extraction capability;secondly,the original PANet is replaced by the Bi FPN structure at the Neck end to enhance the feature fusion;then the Ghost convolution block is used to replace the ordinary convolution in the original network,which can reduce the network computation and accelerate the training convergence;finally,to balance the positive and negative sample problems,the EIo U loss is used to replace the CIo U loss.The experimental results show that the network improvement and loss function optimization of this algorithm can significantly improve the network detection accuracy and effectively reduce the false detection problem,while maintaining a lower parameter computation,which takes advantage in practical applications.(3)Underwater target detection system implementation based on attention mechanismDesign of an underwater target detection system following research into the YOLOv5 algorithm based on attention mechanisms.The system enables real-time detection functions for pictures and videos,as well as outputting the detection results,including predicted categories,probabilities and target locations.