Research On Usv Detection Algorithm Based On Improved Deep Learning

In recent years,unmanned surface vessel(USV)have gradually entered people’s sight,and they have received sufficient attention from both military and civilian perspectives.The reconfigurability of USV allow them to perform tasks independently in military activities,which greatly guarantees the safety of personnel.In the civil industry,USV can easily complete highintensity tasks.Although USV have brought convenience to people’s production and life,they have also caused huge threats and troubles to the country’s coastline safety and navigation regulations.Traditional USV detection technology has obvious shortcomings in terms of accuracy,autonomy and monitoring range.The object detection technology based on deep learning can overcome these difficulties to effectively detect the USV.At present,the detection technology of USV based on deep learning at home and abroad is not yet mature.Therefore,based on the in-depth analysis of the object detection technology under the condition of deep learning,this paper conducts the following research work for such fast and small objects as USV:(1)USV is a type of unmanned intelligent equipment combined with cutting-edge technology in the emerging field.When using deep learning to detect these objects,a network model is needed for data training so as to have the ability to detect these features.However,currently common datasets do not mark this type of characteristic.Therefore,this paper produces USV images data sets from two aspects: military USV and civilian USV.In order to complete the task of object detection,Label Img labeling software is used to convert data pictures into VOC data sets.In the task of instance segmentation,the Labelme annotation tool is used to add an additional mask in the USV to distinguish different targets during detection.(2)In the task of object detection,in order to be able to detect objects with fast navigation speed and small size such as USV in real time.This paper proposes a YOLOv3(You Only Look Once version 3)algorithm based on the E-CIo U(Enclosing Center distance Intersection over Union)loss function,which has an advanced advantage in model convergence speed in the imitation of real proof in MATLAB.On the basis of the proposed algorithm,the data set is trained to obtain the ability to detect such features as USV.At the same time,due to the small imaging effect of USV in the long-distance situation,a multi-scale prediction network needs to be introduced.The multi-stage residual network structure of YOLOv3 just meets the detection requirements for small objects.The proposed E-CIo U loss function makes the positioning of the object box more accurate and plays an important role in improving the accuracy of the algorithm.(3)In terms of candidate box selection,an improved Cluster NMS algorithm is proposed.When the USV contains occlusion,it combines the E-CIo U distance loss to ensure that the neighboring object boxes are two prediction boxes of the same type.The smoothing function attenuates the confidence score of the object box to a certain extent,thereby increasing the output probability of the occluded object box and improving the detection accuracy of the model.In the aspect of instance segmentation,the mask prediction framework in the YOLACT segmentation algorithm is adopted.Its one-stage multi-level prediction structure increases the segmentation accuracy of the model for small objects while ensuring real-time performance.The simultaneous use of USV detection and segmentation makes the detection more abundant image information.