Research On Ship Target Detection Algorithm For Complex Background And Low Visual Environment

In recent years,China’s inland waterway shipping business has been developing rapidly.While ship transportation is reducing social logistics costs,the safety problem of waterway is becoming more and more prominent.Therefore,the real-time target detection of inland waterway vessels has become an important link in waterway safety monitoring.However,the background of river Bridges in inland waterways is complicated and easy to be integrated with ships,which makes it difficult to extract local features.Low visual environment,such as rain,fog and night,leads to low accuracy of global feature detection,which makes the existing vision-based target detection methods unable to meet the accuracy requirements of ship target detection in inland waterways.Based on Guangxi key research and development project "Research and Application of Ship Assisted Steering System Based on Intelligent Network and three-dimensional environment Perception Technology",this paper conducts research on ship target detection difficulties in complex background and low visual environment in ship visual perception.The main research contents of this paper are as follows:(1)In view of the difficulty in extracting ship local features from complex background images,a ship target detection algorithm combining convolution self-attention and multi-scale features is proposed in this paper.Firstly,A mixed model that enjoys the benefit of both self-Attention and Convolution(ACmix)is added to the feature extraction network layer of YOLOv5 detector.A Self-Attention and Convolution-Spatial Pyramid Pooling Fast(AC-SPPF)module is constructed through residual network to weaken the interference of complex background.And strengthen the ability of feature extraction and information fusion.Secondly,the weighted Bidirectional Feature Pyramid Network(BIFPN)is added to the neck layer for multi-scale feature fusion to improve the local feature expression ability of the model.Finally,a large-scale detection head is added to the detection head layer to improve the target detection ability.In order to verify the above algorithm,this paper carried out verification experiments on the Seaships data set,and compared the method in literature 46,Faster R-CNN and YOLOv5,the m AP was improved by 11.9%,5.9% and 3.8% respectively.The accuracy rate and recall rate were improved by 3.9% and6.8% respectively compared with YOLOv5.(2)Aiming at the problem that image detection algorithm in low visual environment is difficult to effectively extract global features,this paper proposes a ship target detection algorithm based on global context network.By integrating the Global Context Network(GCNet)mechanism and embedding it into the deep and shallow feature extraction networks of YOLOv5,ship detection accuracy in low-visibility environment is improved.By generating the global attention feature map model,focus on the target feature information from the global perspective,so as to reduce the problem of missing and false detection of ship targets in low visibility environment.Finally,experimental verification and analysis are carried out on self-made low visual environment data set.Experimental results show that the proposed method is 9.8%,13.5% and 5.5% higher than the m AP of Faster R-CNN,SSD and YOLOv5,and the accuracy and recall rate are 8.4% and 6%higher than that of YOLOv5.In addition to verifying the above improved method on the public data set and the self-made data set,this paper also collected the actual complex visual background and video footage in low visibility environment as the test set,which verified the effectiveness of the proposed method in improving the accuracy of ship target detection.