SAR Ship Detection Research Based On Lightweight YOLOv5

As an Asian coastal state,China’s coastline is 18,000 kilometers long and its sea area is 3million square kilometers,making up 30 percent of the country’s territory.Maritime safety is of vital importance of our national defence and security.Along with the development of SAR,it is inevitable to use the SAR image to detect the target.However,due to subjective factors and other problems,the traditional method will cause a certain degree of missing alarm rate and unpredictability,which will affect the effect of the final target detection.In addition,with the increase of sample data,artificial feature extraction will cause huge consumption of time cost.However,the existing target detection algorithm is mainly concerned with precision of the model,ignoring the number of parameters and the complexity of the model,which leads to the lower detection speed.In view of the above problems,this paper studies the lightweight YOLOv5 algorithm deployed on Jetson nano.The main research contents of this paper are as follows:(1)The interference from background clutter is easy to affect the image quality of SAR images,and therefore,the ship targets of SAR images are easy to be misjudged.Moreover,since most of the existing models focus on how to increase the model precision,how to realize the model in real time is also an issue.In this paper,the object detection algorithm YOLOv5 is lightened,and its improvement mainly includes the following three aspects: 1)The backbone part of YOLOv5 is redesigned by MNEBlock module;2)Using Sim SPPF pyramid structures to dynamically increase the size of receptive fields,splicing features to preserve the features of different receptive fields;3)Using PEA attention module to retain accurate location information on spatial orientation,while avoiding a lot of computing overhead.Experimental results indicate that the proposed method is superior to YOLOv5,the lightweight YOLOv5 model is more suitable for ship target detection in SAR images.On the basis of certain detection accuracy,it has smaller model volume,fewer parameters and faster detection speed.(2)The lightweight YOLOv5 have good detection performance,but how to apply it to realtime detection is still a problem to be solved.Therefore,in this paper,the trained lightweight YOLOv5 was deployed on the NVIDIA Jetson nano edge device to achieve real-time detection of the target.The practicality of deploying lightweight YOLOv5 on the Jetson nano device was tested by comparing the detection effect on different data sets with the same weight and the detection effect on the same data set with different weights.