| The navigation safety in the ocean mainly depends on the sailing experience of the crew and the naked eye discrimination of objects on the sea,which bury hidden dangers for the occurrence of marine accidents.Iceberg accidents,ships hitting rocks,and animal collisions are all painful lessons that have occurred in recent years.Therefore,the exclusion of dangerous objects at sea is an important measure to avoid marine accidents.With the development of technologies such as drones,satellites,and cameras,it is relatively easy to obtain pictures.Therefore,the specific types and locations of floating objects at sea can be obtained and analyzed through pictures.However,conventional detection requires a large amount of manpower to participate,which not only expends a lot of manpower resources and material resources,but also has low efficiency.Therefore,the target detection technology is used to quickly detect floating objects at sea,conduct rapid inspection of abnormal floating objects and dangerous fast detection.Analysis plays an important role in water safety and safety rescue after crisis.Although driven by deep learning,object detection technology has developed by leaps and bounds,but due to the complex water conditions,ships,weather and other factors will put forward higher requirements for image detection.Problems such as image occlusion and background interference still need to be solved urgently,which have a huge impact on the detection accuracy.Therefore,this paper will study the marine floating object detection algorithm based on deep learning,which can provide technical assistance for the danger opinion of water rescue,and is of great significance to water operation and accident rescue.In this paper,the following research work is carried out for the detection difficulties of floating objects at sea:(1)Aiming at the problem of the floating objects data set,this paper mixed the actual photographed sea objects with the public data set images,and used Label Img software to annotate the data of all floating objects existing at sea.The detection accuracy of the algorithm is analyzed through the Marine target data set collected in this paper,which can effectively show the application ability of the algorithm in the actual maritime scene.(2)Multi-scale feature fusion is an important step in target detection model components,and it is also an important means to improve model accuracy.It also has a more robust effect on blurred images and complex targets in the marine environment.By improving the feature coalesce method based on Efficientdet detection model,a multi-scale feature coalesce network containing semantic information and spatial feature is obtained.The shallow feature map has higher resolution and rich spatial features,while the higher-level deep feature map contains rich semantic information.It is better adapted to low-resolution and multi-target scene in the marine environment.The detection performance of the model is further improved.In general data set and ocean target data set,the accuracy of the improved algorithm is improved obviously.(3)The size of receiving domain is related to the degree of abstraction of target feature,which affects the performance of detection model.By adding a receptive field enhancement module,the scale changes and occlusion of the target ground are processed.After extracting features from backbone network through class initiation module,multi-scale receptive field and rectangular receptive field are added to the receptive field,and the feature mapping of the channels with dependencies is improved by adding a channel-based attention mechanism.In the general data set and the marine target data set,the improved algorithm has obvious accuracy improvement. |
