A Study On SAR Ship Target Detection Algorithm Based On Lightweight Convolutional Neural Network

Synthetic Aperture Radar(SAR)is widely employed in remote sensing satellites.As an active earth observation Radar,it has a wide monitoring range and can perform all-day and allweather work.Moreover,the Radar wave utilizes side-looking to scan ships without obstacles and the scanning image is relatively complete.Therefore,it has high application value in navigation.In the field of target detection,convolutional neural network has been gradually applied in various computer vision scenes with its characteristics of high precision and low delay.With the continuous in-depth research in recent years,target detection algorithms have been continuously upgraded with higher precision and lower delay.Therefore,convolutional neural network has a great application space in the field of marine ship detection.In view of the phenomena including large model,excessive number of parameters,low real-time performance of target detection and large amount of computer resources in SAR image detection,the main research contents of this thesis are as follows:1.The data set of SAR ship target detection required by the experiment was prepared for annotation and verification.Aiming at speckled noise images,a series and parallel adaptive weight model was established based on morphological method and different structural elements were used to de-noise the noise images.2.Some popular target detection algorithms in the current deep learning field are studied:Faster-RCNN,SSD(Single Shot Multi Box Detector),and YOLOv4 algorithm.The differences of these algorithms are compared through theoretical analysis,and the algorithms are applied in space-borne SAR image ship detection.The results show that YOLOv4 algorithm is better in both simple scenes and complex scenes.3.Mobile Netv2-YOLOv4 model is used to solve the problems of large volume,large number of parameters and slow detection speed of YOLOv4 model.The model uses the backward residuals structure and uses the deep separable convolution to replace the standard convolution,which greatly reduces the number of parameters.The number and volume of the compressed lightweight model are one sixth of the original model.4.Although the linear bottleneck structure can compress the model effectively in the backbone network,the feature information extracted by deep separable convolution of the compressed feature layer will be partially lost.In order to solve the problem of feature information loss often accompanied by model lightweight,Copy Paste-Mobilenetv2-YOLOv4 model was proposed.By using the copy-paste data enhancement method,the image feature information is increased,which improves the detection accuracy while ensuring the lightweight of the model.